# Agentique > Agentique is an AI back office for companies that run on documents, built in Romania by ITAILORS - SOFT S.R.L. and offered in Romanian and English. A company uploads contracts, invoices, scans and spreadsheets; the assistant reads them, answers questions with the source document and page, fills the data tables the company defines, remembers what it is told, runs scheduled automations, retrieves e-invoices from ANAF's SPV, watches public tenders on SEAP and EU-funded procedures and drafts the offer dossier in the buyer's forms, and exposes the whole workspace to the company's own tools through an MCP server. Sensitive actions always wait for a person. Data is hosted in the EU and never used to train models. Currently in private beta with a 30-day trial. Every public page exists in Romanian at the neutral URL and in English under /en. Romanian is the primary language and prevails in the legal documents. ## How Agentique is different - Every answer carries its source: the document and the page, on every claim, so checking takes one click. When the documents do not contain the answer, it says so instead of improvising. - Nothing leaves the workspace without a person's confirmation: sending email, deleting, changing a record, submitting an offer. The boundary lives in the code, not in a prompt. - It is a library, not a chat window: documents stay, memory carries across conversations, data tables keep themselves current, automations run on a schedule. - Data stays in the European Union (AWS Frankfurt), each company has an isolated workspace, customer data is never used to train models, and a data processing agreement is published. - It does real company work: e-invoices retrieved from SPV, public tender dossiers drafted from SEAP notices, a web assistant for the company's own website, an MCP server for developers. - Results are measured with the customer during the pilot (precision, approval rate, hours returned) rather than promised. ## Product pages - [Home](https://agentique.eu/): what Agentique does, a live product preview, beta access - [Public tenders](https://agentique.eu/licitatii): the tenders module: a daily radar on SEAP and EU-funded procedures, a GO / MAYBE / NO GO verdict per notice proven from the company's own documents, and the offer dossier drafted in the buyer's forms; market figures from ANAP's 2024 reports (99 bn lei awarded through 30,804 contracts, 91% of contracts won by SMEs, 41% awarded on a single offer, 49% of the value on EU-funded projects); four example companies and one opened verdict - [Frequently asked questions](https://agentique.eu/intrebari-frecvente): the whole product in short answers: what it is, files and OCR, sourced answers, data tables, memory, automations, e-Factura and MCP, public tenders, the web assistant, security and GDPR, pricing and the trial - [Our mission](https://agentique.eu/misiune): enterprise-grade AI for companies of every size, what that means in practice, what we believe, build and offer ## Guides and articles - [How Agentique works](https://agentique.eu/blog/how-agentique-works): the journey from upload to sourced answer, in plain words - [Five jobs you can hand over this week](https://agentique.eu/blog/five-jobs-to-hand-over): concrete use cases with the numbers behind them - [What a back office can automate today, and what it should not](https://agentique.eu/blog/back-office-what-to-automate): the capability map, including what stays human - [What your e-Factura archive knows](https://agentique.eu/blog/arhiva-e-factura): turning Romania's mandatory e-invoice XMLs into answers, totals and due dates - [The MCP server, for developers](https://agentique.eu/blog/mcp-server): endpoint, API keys, scopes, rate limits and configuration examples - [The work between the systems, in short](https://agentique.eu/blog/munca-dintre-sisteme): why an AI layer between existing systems takes over the manual carrying of information while a person keeps approving - [The work between the systems: the thesis](https://agentique.eu/blog/munca-dintre-sisteme-teza): the long-form argument, the read-resolve-propose-confirm-record loop, evaluation metrics, what it means for ERPs - [Chapter 1: the job nobody designed](https://agentique.eu/blog/munca-dintre-sisteme-1): where the hidden between-systems job lives, what it costs, a one-page seam audit - [Chapter 2: handing a seam over safely](https://agentique.eu/blog/munca-dintre-sisteme-2): which seams first, the loop, model tiers per step, the confirmation boundary, staged rollout - [Chapter 3: measuring whether it worked](https://agentique.eu/blog/munca-dintre-sisteme-3): resolution precision, approval rate, escape rate, time to close, queue age, hours returned; a vendor evaluation protocol - [Chapter 4: forms, the human interface](https://agentique.eu/blog/munca-dintre-sisteme-4): the intake loop that replaces most forms, which forms remain, intake metrics - [Chapter 5: how a Romanian company starts](https://agentique.eu/blog/munca-dintre-sisteme-5): the mandates as forcing function, three company shapes and their first seam, a 90-day plan, Romanian-specific pitfalls ## Capabilities - Files: PDF (including scans, read by the AI model itself), Word, Excel, XML e-invoices, plain text, images; zip archives are unpacked and every file inside becomes a searchable document; practical limits 50 MB per file and 100 files per archive; unusable files are listed by name - Answers: every answer cites the source document and, for paged formats, the page; search works without diacritics and with typos; answers only from the company's documents, with an explicit "not found" otherwise - Outputs: PDF reports, Excel tables, Word drafts, translations, emails ready to send, all produced from the sources with the calculations done by a calculator tool - Data tables: structured record types defined in natural language, filled from documents and kept current as new documents arrive, exportable to CSV - Memory: facts, preferences and corrections told once are kept across conversations, visible and editable in the app - Automations: multi-step tasks described in natural language and run on a schedule or on an event, visible, editable and stoppable from the Automations page with a run history - Confirmation: sending email, deleting, changing records and submitting to third parties always require an explicit human confirmation - e-Factura (SPV): one-time ANAF authorisation with the company's qualified certificate; incoming e-invoices are retrieved automatically several times a day and processed like uploads - Integrations: Google Drive and Sheets from the settings; named connectors (ERP, CRM, accounting software) built with pilot customers on request; any system that speaks MCP today - MCP server: a Model Context Protocol endpoint per workspace (JSON-RPC 2.0 over Streamable HTTP) with tools for search, documents, memory, data tables, calculations, Google Sheets and Drive, e-Factura status and sync, report generation and public tenders; per-workspace API keys with read or write scopes and rate limits; destructive actions are not reachable through a key; a stdlib-only Python bridge exists for stdio-only clients - Public tenders: sources are SEAP advertising notices and direct acquisitions, the competitive procedures of EU-funded private buyers and institutions' own websites; rules filter before any model runs; the full documentation is read; verdicts mark each mandatory requirement met, unmet or unknown and name the proving document; decisions are recorded with name and date; the dossier comes out as editable Word files and PDFs; nothing is ever submitted automatically - Web assistant: an embeddable chatbot that answers a website's visitors only from the documents the owner provides, citing document and page, with no visitor account and no personal data collected; built for public institutions and companies with published documentation - Languages: interface and answers in Romanian and English; documents in any language ## Security and data - Hosting on Amazon Web Services in eu-central-1 (Frankfurt): files, database, search index, application servers; nothing leaves the EU in normal operation - AI models (Anthropic Claude, Amazon Titan) called through Amazon Bedrock in EU regions; Bedrock stores no prompts or responses and does not train on them; data does not reach the model provider - One isolated workspace per company: separate search index and storage prefix; role-based access (owner, admin, member, viewer); API keys stored only as hashes - TLS in transit; AES-256 server-side encryption at rest; secrets in AWS KMS; technical logs kept 7 days; no analytics or tracking scripts - Retention: documents until the customer deletes them; account data for the contract plus 30 days; on termination workspace data is deleted within 30 days and backups expire 7 days later, except invoices kept by law; full export on request within 10 working days - GDPR: published data processing agreement, subprocessor list, 72-hour incident notification, data-subject requests answered within 10 working days ## Pricing and access - Private beta by invitation code; every new workspace gets a 30-day free trial - Monthly subscription per workspace; tiers announced from 19 euro a month for one person to 249 euro a month for teams, final price agreed with each pilot customer; connectors and the tender module agreed separately in writing - Onboarding done personally within a day, with direct access to the team during the pilot ## Legal - [Terms of service](https://agentique.eu/legal/termeni) - [Privacy policy](https://agentique.eu/legal/confidentialitate) - [Data processing agreement](https://agentique.eu/legal/dpa) - [Security](https://agentique.eu/legal/securitate): the answers to procurement and security questionnaires, quotable - [Website assistant, version 1](https://agentique.eu/legal/asistent-web): the product sheet and terms of the embeddable web assistant - Operator: ITAILORS - SOFT S.R.L., CUI 48450662, J2/1018/2023, Arad, Romania. Documents version 1.0 of 8 September 2026; English twins under /en/legal/ ## Contact - Beta access and questions: mihnea-rares.gilceava@itailors.eu - Company: https://itailors.eu ## Frequently asked questions (full text) URL: https://agentique.eu/intrebari-frecvente (Romanian) · https://agentique.eu/en/intrebari-frecvente (English) ### What Agentique is **What is Agentique, in short?** An AI back office for companies that run on documents. You give it contracts, invoices, scans, spreadsheets and emails; it reads them, answers questions citing the source and the page, fills the data tables you define, runs scheduled tasks, remembers what you tell it and prepares work for you to approve: reports, emails, tender dossiers. It works in Romanian and English and is available in the app, on your website and inside your own tools, through the MCP server. **Who is it for?** Small and mid-size companies and public institutions where a few people move information between documents and systems every day: accounting and back office, logistics, construction, services, agencies, town halls. If your week includes “look it up in the folder”, “put it in the Excel” and “check whether we qualify”, Agentique is for you. **How is it different from a chatbot?** A chatbot forgets your files when you close the tab. Agentique keeps a permanent library, answers only from your documents and says when it cannot find the answer, cites the document and the page every time, keeps information across conversations, manages structured data tables, runs tasks on a schedule and asks for your confirmation before any sensitive action. It is a colleague who knows your archive, not a text box. **Which languages does it work in?** The interface and the answers are in Romanian and English. Documents can be in any language: a Romanian contract, a German invoice, an English specification. Ask in Romanian and the answer comes in Romanian, with correct diacritics, even if you type without them. **Where is it available: app, website, phone?** Today in the web app, on any device; as a web assistant embedded in your own website; and inside your own tools, through the MCP server (Claude, Cursor or any other MCP client). WhatsApp, Slack and a public API are in development and are prioritised for the pilot customers who request them. ### Documents and answers **Which files can I upload?** PDF, including scans, Word, Excel, XML invoices, plain text and images, including a photo of a paper taken with your phone. A zip archive is unpacked and every file in it becomes a searchable document. Practical limits: 50 MB per file and 100 files per archive. Files it cannot use, for example an installer or a video, are listed by name so you know exactly what was skipped. **Does it read scanned pages with stamps and signatures?** Yes. Scanned pages go through optical character recognition performed directly by the AI model, which handles stamps, signatures, handwritten notes and dense tables. Upload shows progress page by page, and a twelve-page scanned leasing contract with a 72-row payment schedule comes out with every row and every total. **How do I know an answer is right?** Every answer names the source document and, for paged formats, the page number, so any statement can be verified in one click. When the answer is not in your documents, it says so instead of improvising. The official information remains the cited document, not the generated text. **Does search work without diacritics or with typos?** “plati intarziate” finds “plăți întârziate”, and a misspelled supplier name is still matched. Search combines meaning with keywords, then the AI re-ranks the results, so you can phrase the question the way you would ask a colleague. **Can it produce files: reports, tables, emails?** Yes: a PDF report, an Excel table, a Word draft, the translation of a document, an email ready to send. It reads the sources, does the calculations and puts the result in a real file you can forward as it is. Sending the email or changing a record always waits for your confirmation. **What do I do with my existing archive, hundreds of old PDFs?** Send it as it is: a zipped folder, a drive export, the archive a client sent you. During the pilot we migrate it together, classify it and check the first answers with you. Large or unusual archives, such as an export from an old ERP or an email mailbox, are taken over on request as part of the initial setup. ### Tables, memory and automations **What are the data tables?** Structured records you define in natural language: suppliers, contracts, vehicles, employees, certificates, whatever your company tracks, with the columns you want. The AI fills them from the uploaded documents and updates them as new documents arrive, so the Excel sheet someone maintained by hand fills itself. Everything exports to CSV, at any time. **What does “memory” mean here?** Tell it once and it stays remembered in every later conversation: a supplier changed its bank account, a client is always invoiced in euro, a county you no longer serve, a correction to something it got wrong. You see the memory in the app, edit it and delete it whenever you want. **What can an automation do?** A multi-step task described in natural language and run on a schedule or on an event: a summary of the new documents every Monday at 08:00; the list of contracts expiring in the next 30 days; an alert when an invoice passes its due date; a weekly report sent to the accountant. You see, edit and stop any automation from the Automations page, with the history of every run. **Does it act on its own or ask for confirmation?** Reading, searching, filling tables and drafting run automatically. Any sensitive action, sending an email, deleting a document, changing a record, transmitting a document to a third party, stops and asks for your confirmation, and you see exactly what is about to happen. Where you want more autonomy for a routine you trust, we configure it together, with clear limits. **Can you build a workflow specific to my company?** Yes. Describe the process the way you would explain it to a new colleague, with the documents involved and the people who approve each step, and we build it together during the pilot: picking up documents from an email inbox, matching deliveries to orders, month-end closing checks, a reminder chain for unpaid invoices. We start with what you do every day. ### e-Factura, integrations and MCP **How does the e-Factura (SPV) connection work?** Through a one-time ANAF authorisation with the company’s qualified certificate, done together during the initial setup. From then on, incoming e-invoices are retrieved automatically several times a day and go through the same reading, classification and extraction process as manually uploaded documents: totals, due dates, suppliers and invoice lines become answers and rows in your tables. **Does it connect to my ERP, CRM or accounting software?** Yes, in two ways. Through the MCP server, any system that speaks the protocol can read and write your workspace today. For specific systems, SmartBill, Saga, HubSpot, Odoo or the API of your own software, we build the connector together during the pilot, starting with the one you use every day. Google Drive and Google Sheets connect from the settings. **What is the MCP server and who is it useful for?** Every workspace exposes a Model Context Protocol endpoint with a set of tools: search, documents, memory, data tables, calculations, Google Sheets and Drive, e-Factura status, report generation and public tenders. Your developer, or your own Claude or Cursor, connects with the endpoint and an API key and works on the same documents you see in the app. Keys carry read or write scopes and usage limits, and destructive actions are not reachable through an API key. **Can I use only the data extraction, as a service for my own system?** Yes. You send documents and get structured records back, through the MCP server or a connector, in the database structure you need: contract schedules, invoice lines, certificate expiry dates. Reading the scanned leasing contract mentioned above end to end cost under 2 lei, so a price per document is easy to set. **Can it send and read emails?** It drafts and sends emails, with your confirmation every time. Reading a shared mailbox, so that attachments and requests enter the library automatically, is configured on request during the pilot, with the rules you set. ### Public tenders **Which sources does the tender radar watch?** Advertising notices and direct acquisitions on SEAP (SICAP), the competitive procedures of EU-funded private buyers and the notices institutions publish on their own websites, all filtered on your CPV codes, counties and value range. The large procedures with a participation notice come next. **What is an advertising notice on SEAP?** The notice through which an authority requests offers for a direct acquisition, below the tender thresholds. Deadlines are short, sometimes two or three days, and the offer is often sent by email or on paper. For a small company these are the most accessible public contracts, provided you see them in time. **Does Agentique submit the offer for me?** No, never automatically. Agentique finds the notice, reads it, tells you what you can prove and drafts the dossier. You decide whether to bid, then read, sign and submit. The decision is recorded with your name and the date. **What does the offer dossier contain?** The buyer’s forms, reproduced one to one; the cover letter and the table of contents; the technical proposal, with its conformity matrix; the financial offer, with the price table; the required declarations (for EU-funded procedures, including DNSH and beneficial owner); the clarification email; and a checklist of the few items only a person fills in. All as editable Word files and PDFs. **How does it establish that I meet a qualification requirement?** A requirement is marked as met only when a document in your library proves it: the ISO certificate, the balance sheet, a delivered contract with its acceptance report, a CV. The verdict names the document. An expired certificate is not taken into account, and what cannot be proven stays “unknown” until you upload the document. **Do I have to fill in forms with the company’s data?** No. Drop the registration certificate, the balance sheets, the certificates, the delivered contracts and the CVs into the library and the company file is built from them automatically. The page shows what is missing and what expires; when you add a document, the open verdicts are re-read. **What does it cost and how do I start?** While the module is in beta, access is by invitation: send us the CPV codes you cover and the county you work in, we set up the radar together and you see the first digest the next morning. You pay a subscription, not a consultant for every dossier. ### The web assistant for your website **What is the web assistant?** A chatbot you embed in your own website that answers visitors exclusively from the documents you make available to it, citing the document and the page, and that says when it does not have the answer. It replies in the language of the question, Romanian or English. Visitors need no account and are not asked for personal data. It is designed for public institutions and for companies with a lot of published documentation. **How is it installed?** You receive a code snippet and add it to the pages where you want the assistant to appear. The documents it may use live in a dedicated workspace that you control: add a document and it can answer from it, remove it and it no longer uses it. Setup together with your team takes a day. **What does it never do?** It does not browse the internet, does not take decisions or issue approvals in your name, does not file requests and has no access to your internal systems. Visitor conversations are kept for 30 days, for technical support and for your own review, then deleted; earlier on request. **Can it match the look of my site and my tone?** Yes. Colours, position, greeting and the note shown to visitors are configured for you, and the tone of the answers follows the instructions you give it: formal for an institution, plain for a shop. Changes are made together with us during the pilot, without further work on your site. ### Security, data and GDPR **Where is my data stored?** On Amazon Web Services, in the eu-central-1 region (Frankfurt, Germany): files, database, search index and application servers. In normal operation nothing leaves the European Union. The only exception is optional: if you connect Google Drive or Google Sheets yourself, that data goes to Google. **Are my documents used to train AI models?** No. The models, Anthropic Claude and Amazon Titan, are called through Amazon Bedrock, in AWS regions inside the European Union. Bedrock does not store prompts and responses and does not use them for training, and your data does not reach Anthropic. We do not train models on customer data and we do not sell data. **How is my company’s data separated from other customers?** Each workspace has its own search index and its own storage prefix; a question asked in one company can never return a document from another. Access inside the workspace is role-based (owner, admin, member, viewer), with invitations by email. API keys are stored only as hashes. **How is the data protected?** TLS in transit; at rest, files in S3 with server-side AES-256 encryption, an encrypted database, and application secrets protected with AWS KMS. Backups stay in the Frankfurt region. Technical logs are kept for 7 days. The platform uses no Google Analytics, pixels or third-party tracking scripts. **Can I delete my data? What happens when the contract ends?** You can delete documents, memories, tables and conversations from the app at any time. On request you receive a complete export, in JSON (and CSV for tables), within 10 working days. When the contract ends, workspace data is deleted within 30 days and backups expire 7 days after that, except for the documents the law requires us to keep, such as invoices. **Is there a GDPR data processing agreement (DPA)?** Yes, published and applicable to every customer, with the list of subprocessors, security incident notification within 72 hours and the assistance we provide for data subject requests. Requests about personal data are answered within 10 working days. **Can I have a dedicated environment or single sign-on (SSO)?** Yes, on request, for institutions and larger companies: a dedicated environment in the same EU region, sign-in through your identity provider, custom retention periods and a signed contract with the clauses your procurement procedure requires. We define them together in the first conversation. ### Pricing, trial and getting started **What does it cost?** A monthly subscription per workspace. Tiers start at 19 euro a month for one person and go up to 249 euro a month for teams; the final price is agreed with each pilot customer according to volume and the modules used. Custom connectors and the tender module are agreed separately, in writing, before any work starts. **Is there a free trial?** Yes, 30 days for every new workspace, with an invitation code. You upload real documents, ask real questions and decide with the answers in front of you. **How soon can I start?** Within a day. You write to us, we set up the workspace and the first documents together, and the first answers come the same afternoon. Larger migrations and connectors follow in the first weeks of the pilot, with direct access to our team throughout. **Do I need technical staff?** No. Uploading, asking, defining a table and describing an automation are done in natural language. A developer is only useful if you want the MCP server or a connector to your own software, and even then we do most of the work together with them. **Who is behind Agentique?** ITAILORS - SOFT S.R.L., a software company in Arad, Romania. We onboard every pilot customer personally, and our thesis on the work between systems is published on the blog. ## Public tenders page (full text) URL: https://agentique.eu/licitatii (Romanian) · https://agentique.eu/en/licitatii (English) ### Bid on public contracts without a tenders department. Agentique watches the notices on your CPV codes every day, reads the documentation, tells you what you can prove from your own documents and drafts the offer dossier in the buyer’s forms. You decide, sign and submit. At the price of a subscription, not of a whole team. Market figures on the page, published by the institutions named: - 99 bn lei in public procurement awarded in 2024, across 30,804 contracts (10.05% of GDP, before direct acquisitions). Source: ANAP, Analiza indicatorilor de eficiență 2024, https://anap.gov.ro/ro/wp-content/uploads/2025/12/Analiza-indicatori-de-eficienta-2024.pdf - 91% of the 2024 contracts went to small and mid-size companies (78% of the value). Source: ANAP, Analiza indicatorilor de eficiență 2024, https://anap.gov.ro/ro/wp-content/uploads/2025/12/Analiza-indicatori-de-eficienta-2024.pdf - 41% of the 2024 contracts were awarded on a single offer received (most often in construction works and in health). Source: ANAP, Analiza indicatorilor de eficiență 2024, https://anap.gov.ro/ro/wp-content/uploads/2025/12/Analiza-indicatori-de-eficienta-2024.pdf - 2.3 million direct purchases a year through SEAP, below the tender thresholds, worth over 15 bn lei (ADR figures for 2020). Source: ADR, ianuarie 2021, https://www.adr.gov.ro/en/articole/o-noua-functionalitate-in-platforma-seap---peste-21-000-de-autoritati-contractante-pot-sa-publice-noi-tipuri-de-modificari-in-derularea-contractelor Figures published by the institutions, not counted by us: ANAP for 2024, ADR for the direct purchases, with its figures for 2020. Public contracts count only what went through a procedure; the direct purchases sit below. Bid on public contracts without a tenders department. Agentique watches the notices on your CPV codes every day, reads the documentation, tells you what you can prove from your own documents and drafts the offer dossier in the buyer’s forms. You decide, sign and submit. At the price of a subscription, not of a whole team. ### Where the time goes today A tender is rarely lost when the offers are opened. It is lost weeks earlier, in the hours nobody had. - Notices appear when nobody is looking. To see them, someone has to remember to search: on SEAP, on the institutions’ own sites, at the companies spending EU funds. - Most notices that look right are not: another county, a certification you do not hold, a value outside your range. You find out after an hour of reading, every time. - How to submit, by when and with which forms: every authority writes it somewhere else, in its own way. - The same certificates, balance sheets and references come out of the folders for every offer, and someone checks again whether they are still valid. - A dossier written by a consultant starts at a few thousand lei, paid per piece whether you win or not, and the owner still reads every page before signing. For a small company, that price alone decides which notices even get a try. Where the change starts ### The company file builds itself It starts from the documents you already have, not from a form. Drop the registration certificate, the balance sheets, the ISO certificates, the contracts you have delivered and the CVs of your key people into the library. Agentique reads them into a company file: who you are, what you can prove, what expires when. You see what is missing and which requirements stay unverified because of it; add the document, and the open verdicts re-read themselves. From then on, every notice is checked against your file, not against what someone remembers. ### From the notice to the signed dossier - **Watches the sources on your codes** You give it the CPV codes you sell under. Once a day it collects what is new on those codes: advertising notices and direct acquisitions on SEAP, the procedures of EU-funded private buyers, and the notices institutions publish on their own sites. - **Scores them against your company, without a model** Keywords, exclusions, counties, value range and deadline give each notice a score. Only the ones above the line go to the AI, so a notice for printer toner never costs you a cent. - **Reads the documentation** A quick check first. Then the specification is downloaded, put in your library and read in full. You get a verdict, GO, MAYBE or NO GO, with the reasons: what they want, every mandatory requirement marked met, unmet or unknown, the red flags, how and where to submit, the deadline, a suggested price and the next steps. - **Proves it from your own documents** A requirement is marked met only when a document in your library proves it: the ISO certificate, the balance sheet, a delivered contract, a CV. The verdict names the document. An expired certificate does not count. - **Sends the digest, records your decision** One email a morning with what is worth a look. In the app you mark each notice: we bid, we pass, later. The decision is recorded with the name and the date, and the radar learns from it. - **Drafts the dossier when you say GO** The buyer’s forms one to one, the technical proposal requirement by requirement, the financial offer, the declarations and the clarification email, from the tender file and your company file. Word and PDF, for you to read, sign and submit. ### Why it matters - **The same tools, at every size** What a procurement department does for a corporation, reading, matching, drafting, keeping the file current, is now the work of a system. A company of ten gets it at the price of a subscription, not of a department. - **Evenings back** The owner reads a two-page verdict instead of a forty-page file, and signs a dossier instead of writing one. The judgment stays with the person; the carrying does not. - **A better market** In 2024, four public contracts in ten in Romania were awarded on a single offer received (ANAP). Every one of them is a price nobody checked, at a town hall or at a company spending EU funds. When small companies can answer a notice correctly and on time, the buyer gets more offers and public and European money buys more. That is the part of the mission this module carries. ### What you get: the offer dossier, ready to sign This is where the days go, and where a consultant charges per dossier. When you say GO on a notice, the offer package is assembled from the tender file and your company file, in the shape the buyer asked for: - The buyer’s own forms, reproduced one to one, with their original numbering. - The cover letter and the table of contents. - The technical proposal with a conformity matrix: every requirement, what you offer against it, and what goes beyond the minimum. - The financial offer with the price table and the delivery schedule, starting from the suggested price. - The declarations: conflict of interest, exclusion grounds, and for EU-funded procedures DNSH and beneficial owner. - The clarification email to send before the deadline, with the questions that actually change the offer. - A checklist of the few things only a person fills in: the signatory’s ID, signatures, the envelope for paper submissions and the courier date that makes the deadline. Everything comes out as editable Word files and PDFs. A person reads, corrects, signs and submits. During the beta we prepare the dossier together with you; the same drafts are assembled inside the app. ### Know the market before you price In 2024, almost half of the public contracts in Romania, 48% by number, were awarded on the lowest price alone (ANAP). So before you price, the assistant can pull the finalized direct acquisitions from SEAP for any CPV code: how many contracts, at what prices, which authorities buy regularly and who wins them. Ask it in plain words, in the chat or from your own tools over MCP. ### How the work is shared **Agentique** - Watches the sources every day and scores what appears against your company. - Reads the documentation and writes down what is asked, what you can prove, and what is missing. - Recommends GO, MAYBE or NO GO, with the reasons in plain sight. - Assembles the dossier drafts and the clarification email, from your documents. **You** - You choose whether to bid. The decision is recorded with your name and the date. - Read, correct and sign the drafts; a document in your library is what proves a requirement, not a sentence. - Submit the offer, on paper or on the platform. Nothing leaves without you. - Tell it what changed: a new certificate, a county you no longer serve, a price floor. What the radar watches today: advertising notices and direct acquisitions on SEAP, the competitive procedures of EU-funded private buyers, and the notices institutions publish on their own websites. The large open procedures come next. The EU money is not a side channel: in 2024, 49% of the value of the public contracts awarded in Romania was on EU-funded projects (ANAP). ### Why this belongs in your systems A tender is the work between systems in its purest form: someone carries facts from SEAP, from the tender file and from the company folder into one dossier, with judgment at every step. Agentique does the carrying, keeps the judgment visible, and lets you reach the result from wherever you work: the app, the morning email, or your own tools through the MCP server, where the notices, the verdicts, the company file and the market data are plain tools. ### Start with your CPV codes Access is by invitation while the module is in beta. Write to us with the CPV codes you sell under and the county you work in; we set up the radar together and you see the first digest the next morning. ## Mission (full text) URL: https://agentique.eu/misiune (Romanian) · https://agentique.eu/en/misiune (English) Enterprise-grade AI for companies of every size. The best AI systems have so far reached the companies with data teams, budgets, and time. We bring them to companies of ten people and to companies of three hundred, with the same guarantees: sources in view, confirmation before action, measured numbers, data hosted in the European Union. As a product, through Agentique. As services, through the iTailors team that builds it. ### What enterprise-grade means in practice - Every answer carries its source: the document and the page. - Nothing leaves the workspace or changes a record without a person confirming. - Each company’s data isolated; credentials encrypted at rest; API keys with narrow scopes and rate limits. - An audit trail: who did what, when, on what evidence. - Hosted in the European Union. - Measured, not promised: precision, approval rate, what escaped, hours returned. ### What we believe - Small and mid-size companies have the same problems as large corporations, with a tenth of the people. They deserve the same quality of tools. - The most expensive work in a company is the work between its systems: carrying information from one program into another, with judgment at every step. An AI that reads, remembers, and acts can take that judgment, with the person kept at approval. - The measurement is the product. A vendor who cannot show precision, approval rate, and what escaped is asking to be believed rather than measured. ### What we build Agentique: an AI layer that reads the company’s documents, scans, and e-invoices, resolves them against what it already knows, proposes the entry with its source attached, waits for confirmation, and writes back. With memory for the house rules, data tables it keeps current, scheduled workflows, and an MCP endpoint for your own tools. ### Our services For companies that want the result, not the project: an inventory of the crossings between your systems, a pilot on a single crossing, the integrations your company actually needs, and, for technical teams, MCP access to the same workspace. Everything on one platform; nothing built from scratch for a single client and left without maintenance. ### What we set out to do The product is the first step. This is where we want to take it. - Bring the safety and quality of large-company AI systems to every small and mid-size company in Romania, at a cost it can afford. - Make the work between systems disappear from people’s days: documents and e-invoicing first, then the ERP, the bank, transport, human resources, until a company connects its systems in days rather than months. - Build the connectors Romanian companies actually use, in the order the pilots ask for them. - Publish the numbers, precision, approval rate, escapes, hours returned, for every kind of crossing, as we measure them. - Take the same model across Europe as e-invoicing mandates spread country by country. ### How we work In pilots, together with you. The inventory of the crossings between your systems first, then one crossing handed over, in shadow, then assisted, with the numbers on the table every week. The full argument is in the thesis; the method is in the five chapters that follow it. The five practical chapters ### No company should have to spend its people’s time on the work between its systems. The people who did that work are not replaced; they are freed from its mechanical half and kept where their judgment matters, at approval. The iTailors team, the company behind Agentique. ## Legal documents (full text) Version 1.0, in force from 2026-09-08. The Romanian version prevails; this is the English text. ### Terms of Service URL: https://agentique.eu/legal/termeni (Romanian) · https://agentique.eu/en/legal/termeni (English) Version 1.0, effective 8 September 2026. These terms govern the use of the Agentique platform (https://agentique.eu), provided by ITAILORS - SOFT S.R.L.. By creating an account or using the platform, the customer accepts these terms. #### 1. Definitions - "iTailors", "we" or "us": ITAILORS - SOFT S.R.L., a Romanian limited liability company, tax ID (CUI) 48450662, registered office at Calea Aurel Vlaicu, Arad, Arad County, Romania, Trade Register no. J2/1018/2023, website https://itailors.eu. - "Platform" or "Service": Agentique, the application available at https://agentique.eu, together with its modules, programming interfaces and documentation. - "Customer": the company or entity on whose behalf a workspace has been created. - "Workspace": a customer's isolated environment within the platform, with its own search index, its own storage prefix and its own data. - "User": an individual who accesses the customer's workspace through an account, with one of the roles owner, admin, member or viewer. - "Customer Content": the documents, data, questions, memories, entities, recorded decisions and any other material uploaded to or generated in the workspace at the customer's request. - "AI Output": the answers, extractions, summaries, verdicts and documents produced by artificial intelligence models from Customer Content. - "Privacy Policy" and "DPA": the document describing how personal data is processed and, respectively, the data processing agreement under which we process data on the customer's behalf. Both form part of the contract. #### 2. Accounts and invitation codes 2.1. Access to the platform requires an account. Creating a customer's first account requires an invitation code issued by iTailors. The code may not be passed on to other companies. 2.2. The person who creates the workspace becomes its owner. Owners and admins can invite other users by email and assign them the admin, member or viewer role. The customer is responsible for everything done through its users' accounts. 2.3. Account details must be accurate and kept up to date. Passwords are personal. Users keep them confidential and tell us at once if they suspect unauthorised access. 2.4. The Service is intended for companies and professionals only. It is not aimed at anyone under 18 and may not be used in a consumer capacity. #### 3. Trial period and subscription 3.1. Every new workspace comes with a free 30-day trial. During the trial the Service is provided "as is", with no availability guarantees. 3.2. When the trial ends, access continues only under a subscription. If no subscription is activated, access to the workspace is blocked; the data is retained and then deleted as described in section 11. 3.3. The subscription price is the one communicated to the customer at the time of contracting. Payment is made against an invoice issued by iTailors, by bank transfer, within 15 days of the invoice date. The platform has no built-in payment processor and does not collect card details. 3.4. Prices exclude VAT. Late payment entitles iTailors to suspend access after notice, as described in section 11. 3.5. The subscription renews for successive periods equal to the initial one unless either party terminates it before the end of the current period. #### 4. What the Service does and the limits of AI 4.1. Agentique lets the customer upload documents (PDF, DOCX, XLSX, XML, TXT, images) and provides an AI assistant that reads them, answers questions citing the source document and page, extracts data into user-defined tables, runs scheduled workflows, remembers facts across conversations and manages entities (custom data tables). Scanned documents are read through optical character recognition performed by the same AI model. 4.2. AI Output can be wrong or incomplete. The platform shows the source and page precisely so that a person can check. The customer alone is responsible for decisions taken on the basis of AI Output and for verifying it before relying on it towards third parties or authorities, or in documents with legal effect. 4.3. AI Output is not legal, tax, accounting or other professional advice. 4.4. The Tender Radar module reads public notices from e-licitatie.ro (SEAP/SICAP) and, based on the tender documentation, proposes a GO, MAYBE or NO GO verdict. Verdicts are recommendations, not decisions. The decision to take part and the submission of a bid always remain with a human. The platform does not submit bids in SEAP and does not act on the customer's behalf towards contracting authorities. Any document the platform generates (answers, clarifications, forms, declarations) must be reviewed and signed by an authorised person at the customer before it is used. 4.5. Automated actions (scheduled workflows, automatic emails or notifications) are switched on only by the customer, explicitly, and can be stopped at any time from the application. The customer is responsible for the configuration and the effects of the automated actions it enables. 4.6. We do not train AI models on Customer Content. We do not sell Customer Content and do not use it for anything other than providing the Service. The AI models (Anthropic Claude and Amazon Titan) are delivered through Amazon Bedrock in AWS regions within the European Union. Amazon Bedrock does not store prompts and responses and does not use them for training, and the data does not reach Anthropic. #### 5. Customer obligations 5.1. Lawful content. The customer uploads only content it is entitled to use and process in this way. The customer warrants that it holds the necessary rights to its documents and that uploading them to the platform does not infringe intellectual property rights, confidentiality obligations or data protection law. 5.2. Prohibited uses. The platform may not be used for: unlawful content or content that infringes third-party rights; malicious software; attempts to access other customers' data or iTailors' infrastructure; security testing without our written consent; systematic extraction of the Service to build a competing product; deliberately overloading the system. 5.3. Authorised users. The customer ensures that only authorised people have access to the workspace, assigns appropriate roles and removes access for people who leave the company. The customer is responsible for its users' compliance with these terms. 5.4. Personal data in documents. Where documents contain personal data (for example employee CVs, customer or supplier details), the customer is the controller of that data and ensures it has a lawful basis for processing. iTailors processes it on the customer's behalf under the DPA. 5.5. Programmatic access (MCP) and API keys. The customer may issue programmatic access keys to connect the platform to its own tools through the MCP protocol. We store keys only as a hash; they cannot be recovered after issue. The customer keeps them secure, rotates them when appropriate and revokes them immediately if it suspects they have been compromised. Any action performed with a valid key is treated as the customer's action. #### 6. Third-party integrations enabled by the customer 6.1. The platform includes optional modules that connect to third-party services. Each module is enabled only by the customer, explicitly, and can be disabled at any time. 6.2. Google Drive and Google Sheets. The connection is made through OAuth, in the customer's own Google account. Data exchanged with Google is governed by Google's terms and policies and may involve a transfer to Google in the United States under standard contractual clauses or the Data Privacy Framework to which Google adheres. The customer can revoke access at any time, from the application or from its Google account. 6.3. SPV / ANAF e-Invoicing. The module downloads the customer's invoices from the ANAF Virtual Private Space (SPV) using the customer's qualified certificate. The customer is responsible for the validity of the certificate and for its right to use it. iTailors does not submit documents to ANAF on the customer's behalf. 6.4. Tender Radar (SEAP/SICAP). The module reads public notices from e-licitatie.ro, including the contact details of contracting authorities, which are public information. The rules in section 4.4 apply. 6.5. Messaging channels. The WhatsApp, Messenger and Slack connectors are enabled only by the customer, using the customer's own accounts; by default only the in-app chat and the embeddable assistant (iframe) are active. When the customer connects a messaging channel, messages exchanged through that channel are also subject to the channel provider's terms. 6.6. iTailors is not responsible for the operation, availability or changes of third-party services, nor for the consequences of their interruption. #### 7. Data and confidentiality 7.1. The processing of personal data is described in the Privacy Policy. For personal data within Customer Content, where the customer is the controller and iTailors the processor, the DPA applies and forms part of the contract. 7.2. Data is hosted on Amazon Web Services infrastructure in the eu-central-1 region (Frankfurt, Germany). AI models run in AWS regions within the European Union. In normal operation no data is transferred outside the European Union or the European Economic Area. The exception is integrations enabled by the customer, as described in section 6. 7.3. Each customer has an isolated workspace: a separate search index, a separate storage prefix and separate data. Users of one customer have no access to another customer's data. 7.4. Mutual confidentiality. Each party keeps confidential the non-public information it receives from the other in connection with the contract and uses it only to perform the contract. This obligation survives termination. 7.5. Customer control. Users can delete documents, memories, entities and conversations directly in the application. On termination, Customer Content is deleted as described in section 11. 7.6. Security incidents. If an incident affects personal data we process on the customer's behalf, we notify the customer without undue delay and no later than 72 hours after becoming aware of it, and we provide the information the customer needs to meet its own legal obligations. #### 8. Intellectual property 8.1. The platform, including its code, design, the Agentique brand, documentation and all improvements, belongs to iTailors or its licensors. The customer receives a non-exclusive, non-transferable right, limited to the term of the contract, to use the platform for its internal business. 8.2. Customer Content remains the customer's. iTailors receives only the non-exclusive licence needed to host, index, process and display that content for the purpose of providing the Service, for the term of the contract. 8.3. The customer may use AI Output produced from its Customer Content without restriction from iTailors, within the limits of the law and of third-party rights. 8.4. The customer does not copy, decompile, modify or create derivative works of the platform, except where expressly permitted by law. #### 9. Availability and support 9.1. During the trial the Service is provided "as is", with no availability commitment. 9.2. For subscriptions, during the beta period the service is provided on a best-effort basis, with no guaranteed availability level. 9.3. The platform currently runs in a development (beta) environment and may be unavailable overnight, between 23:30 and 09:00 Romanian time, for maintenance. Planned maintenance is announced in advance, in the application or by email, whenever possible. 9.4. Support is provided by email, at the address shown on the website under "Contact support", on business days. We aim to respond within a reasonable time. #### 10. Liability 10.1. iTailors' total liability towards the customer, on any ground connected with the contract, is limited to the subscription fees actually paid by the customer in the 12 months preceding the event giving rise to liability. During the trial, when nothing is paid, iTailors' liability is limited to what the law does not allow to be excluded. 10.2. iTailors is not liable for indirect damage, loss of profit, loss of business, loss of data (other than loss caused by a breach of the DPA) or loss of reputation, nor for decisions the customer takes on the basis of AI Output. 10.3. Nothing in these terms excludes or limits liability for fraud or gross negligence, for personal injury, or in any other case where the law does not allow limitation. 10.4. The customer indemnifies iTailors against third-party claims arising from Customer Content, from a breach of section 5 or from use of the platform in breach of the law. 10.5. Force majeure. Neither party is liable for non-performance caused by events beyond its reasonable control. #### 11. Suspension and termination 11.1. The customer may terminate the contract at any time, effective at the end of the paid subscription period, by notice sent to the support address or from the application. Fees already paid are not refunded for the remaining period. 11.2. iTailors may suspend the customer's access, with prior notice by email or in the application, in case of: late payment; a breach of section 5; a security risk to the platform, to other customers or to third parties; a request from an authority. In urgent security situations suspension may be immediate, with notice following as soon as possible. 11.3. If the reason for suspension is not remedied within a reasonable time after notice, iTailors may terminate the contract. 11.4. Either party may terminate the contract for a serious breach by the other party that is not remedied within 15 days of the notice. 11.5. Effects of termination. On termination, access to the workspace is closed. The customer may export its data before the termination date, and on request we help with the export during the notice period. Customer data is deleted within 30 days of termination, except for data we are legally required to keep (for example invoicing records, 10 years under the Romanian Accounting Law no. 82/1991) and backups, which expire within a further 7 days after deletion. #### 12. Changes to these terms 12.1. We may change these terms. Changes are communicated at least 30 days before they take effect, by email to the owner or by notice in the application. 12.2. If the customer does not agree with the changes, it may terminate the contract before they take effect, without penalty. Continued use after that date constitutes acceptance of the new terms. 12.3. The current version of these terms is published at https://agentique.eu. The version accepted by the customer and the date of acceptance are recorded when the account is created. #### 13. Governing law and courts 13.1. The contract is governed by Romanian law, including the Civil Code, Law no. 365/2002 on electronic commerce and data protection legislation (Regulation (EU) 2016/679 and Law no. 190/2018). 13.2. The parties will first try to settle any dispute amicably, within 30 days of written notice. Failing that, the dispute goes to the competent courts of Arad, Romania. 13.3. These terms, together with the Privacy Policy, the DPA and the commercial offer accepted by the customer, form the entire agreement between the parties regarding the platform. In case of conflict, the following order of precedence applies: the DPA (for personal data), the commercial offer (for price and SLA), these terms. 13.4. In case of any discrepancy between the Romanian and the English version, the Romanian version prevails. #### 14. Contact ITAILORS - SOFT S.R.L., tax ID (CUI) 48450662, registered office at Calea Aurel Vlaicu, Arad, Arad County, Romania, Trade Register no. J2/1018/2023. Support and questions about these terms: the address shown at https://agentique.eu under "Contact support". Operator website: https://itailors.eu. --- ### Privacy Policy URL: https://agentique.eu/legal/confidentialitate (Romanian) · https://agentique.eu/en/legal/confidentialitate (English) Version 1.0, effective 8 September 2026. Agentique is a platform where your company uploads documents and gets an AI assistant that reads them and answers with the source. To do that, the platform processes personal data: yours, as a user, and that of the people who appear in your company's documents. This policy explains what we process, why, where it is kept and what rights you have. We wrote it so a busy person can read it without needing a lawyer. #### 1. Who we are and how to reach us The controller for the agentique.eu website and the Agentique platform is ITAILORS - SOFT S.R.L. ("iTailors"), a Romanian limited liability company, tax ID (CUI) 48450662, with its registered office at Calea Aurel Vlaicu, Arad, Arad County, Romania, registered with the Trade Register under no. J2/1018/2023. Company website: https://itailors.eu. For any question about your data, to exercise your rights or to report a problem, write to the email address shown on the website under "Support contact". The same address handles technical support. We have not appointed a Data Protection Officer (DPO), as our activity does not trigger that obligation. #### 2. Our roles: when we are a controller and when we are a processor Agentique holds two kinds of data, and we play a different role for each. **iTailors is the controller** for the data needed to run our direct relationship with you and your company: your user account, billing details, support correspondence and the technical logs of the website. For this data we decide the purpose and means of processing and we answer to you directly. **iTailors is a processor** for everything your company uploads or generates in its workspace: documents, extracted tables, entities, memories, conversations, the company file, employee CVs. For this content **your company is the controller**: it decides what to upload, who has access and how long to keep it. We process that content only to deliver the service, following the instructions in the contract and in the application, and we use it for nothing else. The details are set out in the Data Processing Agreement (DPA) attached to the Terms of Service. The same rule applies to the Agentique web assistant when an institution or company installs it on its own website: the institution is the controller for its visitors' data and iTailors is the processor. If you appear in a document uploaded by one of our customers (for example you are their employee or supplier) and want to exercise your rights, contact that company first. If you write to us instead, we forward your request to the customer and help them respond. #### 3. What we process, why, on what basis and for how long The categories below are grouped by your situation. Legal bases refer to Article 6(1) of the GDPR. ##### 3.1 Visitors of agentique.eu - **Data:** server logs (IP address, browser type, page visited, date and time), generated by the AWS/CloudFront infrastructure. - **Purpose:** website security, abuse detection, error diagnosis. We do not profile visitors and we do not track them across sites. - **Basis:** our legitimate interest in keeping the site secure and working (Art. 6(1)(f)). - **Retention:** 7 days. ##### 3.2 Access requests and contact messages - **Data:** the name, email address, company and message you send us by email, from your own email client. The website has no form collecting data in the background. - **Purpose:** to reply, send you an invitation code and set up your workspace. - **Basis:** steps taken at your request before entering into a contract (Art. 6(1)(b)). - **Retention:** until your request is handled; if you become a customer, for the term of the contract. Correspondence that does not lead to a contract is deleted after 12 months. ##### 3.3 Platform users - **Data:** name, email address, password (stored only as a hash, never in clear text), workspace role (owner, admin, member, viewer), company, interface language, activity log (what you asked, which documents you uploaded), conversations with the assistant, memories saved across conversations, decisions recorded in the app (for example a go/no-go verdict on a tender, with your name and the date), invitations sent to colleagues. - **Purpose:** to run your account, authentication, access rights and work history; to provide support; to invoice your company. - **Basis:** performance of the contract (Art. 6(1)(b)); legal obligations for invoicing and accounting (Art. 6(1)(c)); our legitimate interest in security, abuse prevention and product improvement, without training AI models (Art. 6(1)(f)). - **Retention:** for the term of the contract, plus 30 days after termination, after which we delete it. Invoicing records are kept for 10 years under the Romanian Accounting Law no. 82/1991. We do not collect card details: there is no payment processor built into the platform. Payment is made against an invoice issued by iTailors, by bank transfer, within 15 days of the invoice date. ##### 3.4 Content of the documents your company uploads - **Data:** any personal data contained in uploaded documents (PDF, DOCX, XLSX, XML, TXT, images): your company's employees, customers, suppliers, partners. From the documents we extract text, search chunks, the tables you define and, at your request, entity records and memories. - **Purpose:** so the assistant can find information, answer with the source and page, and run the workflows you configured. - **Role and basis:** your company is the controller and chooses its own legal basis; we process as a processor on its instructions (Art. 28 GDPR). - **Retention:** as long as you keep it in the app. You can delete documents, memories, entities and conversations from the interface at any time. When the contract ends, workspace data is deleted within 30 days of termination, and backups expire 7 days after that, except what we are legally required to keep (invoices). ##### 3.5 Optional integrations you switch on None of the integrations below is active by default. You enable them in settings and can disable them at any time. - **Google Drive and Google Sheets:** you sign in via OAuth to your company's Google account. The documents you select come into Agentique, and the data you send to Sheets goes to Google under Google's terms. You can revoke access at any time from Agentique settings or from your Google account. - **SPV / ANAF e-Invoicing:** using your company's qualified certificate, the platform downloads your invoices from the Virtual Private Space. Invoices contain data about your suppliers and customers and fall under the rule in 3.4. - **Programmatic access (MCP):** API keys are issued by you from the app and stored only as a hash. What an MCP client does with your data depends on your configuration. - **Messaging channels:** the WhatsApp, Messenger and Slack connectors are enabled only by you, using your company's own accounts; the data received is the messages your users send on that channel. By default only the in-app chat and the embeddable assistant (iframe) are active. ##### 3.6 Tender radar (SEAP/SICAP) This module reads public notices from e-licitatie.ro, scores them against your company profile and, on request, has the AI read the documentation and propose a verdict (GO, MAYBE, NO GO). Notices contain contact details of contracting authorities (name, position, phone, email of the responsible officers), published by those authorities under public procurement law. We process them in your company's legitimate interest of identifying opportunities (Art. 6(1)(f)), without profiling the individuals concerned. The decision and the submission of a bid always rest with a person; the platform does not submit bids in SEAP and does not contact authorities on your behalf. ##### 3.7 Company file and CVs From the documents your company uploads (certificates, financial statements, contracts, CVs) the platform extracts records about your company so you can reuse them in bids. CVs contain personal data of your employees: name, experience, education, certifications, sometimes contact details. This data is processed solely on behalf of your company, which remains the controller and is responsible for informing its employees. We use it for no other purpose and never combine it with other customers' data. #### 4. Where data is hosted and how the AI works All platform data is hosted on Amazon Web Services (AWS) in the eu-central-1 region (Frankfurt, Germany): files (S3), database (RDS MySQL), search index (OpenSearch), application servers (ECS) and processing queues (Redis). Backups: the database has automated daily backups kept for 7 days; files live in AWS S3, durable storage replicated across several availability zones of the Frankfurt region. Each customer company has an isolated workspace: separate search index, separate storage prefix, separate data. Users of one company cannot see another company's data. The AI models (Anthropic Claude for answers and OCR, Amazon Titan for embeddings) are accessed through Amazon Bedrock, only in AWS regions inside the European Union (Frankfurt, Ireland, Paris, Stockholm, Milan, Spain). Amazon Bedrock does not store prompts and responses and does not use them to train models. Data does not reach Anthropic. iTailors does not train models on customer data and does not sell data. In normal operation there is no transfer of data outside the European Union or the European Economic Area. The only exception arises if you connect Google Drive or Google Sheets: the data you select may then be transferred to Google (United States) under standard contractual clauses and Google's certification under the EU-U.S. Data Privacy Framework. #### 5. Recipients and sub-processors We do not sell, rent or hand over personal data to third parties for their own purposes. Data may reach: - **Amazon Web Services EMEA SARL** (Luxembourg): hosting, storage, databases and Amazon Bedrock, in the EU regions listed above. - **Google** (Google Ireland Limited / Google LLC): only if you connect Drive or Sheets. - **Our transactional email provider** for invitations, digests and alerts: messages are sent through the SMTP server configured by iTailors, and the sender is the address shown in the messages. - **Public authorities**, where the law requires it (for example on a court order). - **Your colleagues in the workspace**, according to the roles set by your company's administrator. The full list of sub-processors, together with our obligation to notify you before adding a new one, is part of the Data Processing Agreement. #### 6. Security The main measures we apply: - encryption in transit (TLS) for all traffic between your browser and the platform and between platform components; - encryption at rest through AWS services: files in S3 with server-side encryption (AES-256) on all buckets, an encrypted RDS database, and application secrets protected with AWS KMS; - passwords stored only as a hash; API keys stored only as a hash; - full workspace isolation between customers; - role-based access (owner, admin, member, viewer), controlled by your company's administrator; - technical logs in AWS CloudWatch, kept for 7 days, for diagnostics and incident investigation; - iTailors staff access to production data only for support and maintenance, on a need-to-know basis. If a security incident affects your data, we notify the Romanian supervisory authority (ANSPDCP) within 72 hours of becoming aware of it, as the GDPR requires, and inform you or your company without undue delay. #### 7. Your rights and how to exercise them You have the following rights over your personal data: - **access:** to learn what data we hold about you and receive a copy; - **rectification:** to correct inaccurate data (name, email and language can be changed directly in the app); - **erasure:** to have your data deleted when we no longer have a legitimate reason to keep it; - **restriction:** to limit processing while we review a dispute; - **portability:** to receive the data you gave us in a structured, commonly used, machine-readable format; - **objection:** to object to processing based on legitimate interest; - **withdrawal of consent:** where processing relies on consent, you can withdraw it at any time, without affecting processing carried out before. To exercise your rights, write to the address under "Support contact" on the website. We respond within 10 working days (the law allows up to 30 days); if a request is complex, we will tell you if we need an extension. For your protection we may ask you to confirm your identity before acting on a request. Exercising your rights is free of charge. If your data appears in content uploaded by one of our customers, Section 2 applies: we point you to that company and help it respond. If you believe the processing breaches the law, you have the right to lodge a complaint with the Romanian National Supervisory Authority for Personal Data Processing (ANSPDCP), B-dul G-ral. Gheorghe Magheru 28-30, Bucharest, www.dataprotection.ro, or to bring the matter before the Romanian courts. We would prefer that you write to us first so we can fix the problem directly. #### 8. Cookies and local storage The agentique.eu website uses no analytics or tracking tools. There is no Google Analytics, Meta Pixel, Hotjar or any other third-party script. Fonts are served from our own site, not from Google Fonts, so even loading a page sends no data to third parties. The application does not use cookies for the login session. The data needed to operate is kept in your browser's local storage (localStorage), on your device, and is never sent to any third party: - **auth_token:** the token that keeps you signed in; - **ui_language:** interface language (Romanian or English); - **onboarding_completed:** whether you have completed the getting-started guide; - **sidebar_collapsed:** whether the side menu is collapsed; - **chat_sidebar_width:** the width of the conversation panel; - **subscription_block:** subscription status, so we can show the renewal page correctly. You can clear these at any time from your browser settings; doing so signs you out and resets preferences to their defaults. The only cookies that may appear are the technical ones of the Django framework (csrftoken, sessionid), used only in iTailors' internal administration area, not for website visitors or application users. Because all of these elements are strictly necessary to deliver the service you requested or to remember your preferences, they do not require prior consent under Art. 4(5) of Romanian Law no. 506/2004 (implementing the ePrivacy Directive). That is why we show no cookie consent banner. If we ever add analytics tools, we will update this policy and ask for consent first. #### 9. Children Agentique is a service for companies and is not directed at anyone under 18. We do not knowingly create accounts for minors. If you learn that a minor has been given access through an invitation from your company, write to us and we will delete the account. #### 10. Changes to this policy We may update this policy when the service, our providers or the law change. The version and effective date appear at the top of the document. For material changes we notify you by email or in the app at least 30 days before they take effect. Earlier versions are available on request at the contact address. Applicable law: Regulation (EU) 2016/679 (GDPR), Romanian Law no. 190/2018, Law no. 506/2004, Law no. 365/2002 and the Romanian Civil Code. --- ### Data Processing Agreement URL: https://agentique.eu/legal/dpa (Romanian) · https://agentique.eu/en/legal/dpa (English) Version 1.0, effective 8 September 2026. This agreement (the "Agreement") is entered into between: - the Customer, being the company identified in the subscription contract or, failing that, in the account created on the Agentique platform (the "Controller"), and - ITAILORS - SOFT S.R.L., a Romanian limited liability company, tax ID (CUI) 48450662, Trade Register no. J2/1018/2023, registered office at Calea Aurel Vlaicu, Arad, Arad County, Romania ("iTailors" or the "Processor"). The Agreement supplements the contract for the provision of the Agentique platform (the "Main Contract") and governs the processing of personal data that iTailors carries out on the Controller's behalf, in accordance with Article 28 of Regulation (EU) 2016/679 (the "GDPR") and Romanian Law no. 190/2018. #### 1. Definitions Terms defined in Article 4 GDPR have the same meaning in this Agreement. "Customer Data" means the personal data contained in the documents, conversations, memories, entities and other content that the Controller or its users upload to or generate in their Agentique workspace. "Sub-processor" means any third party engaged by iTailors to process Customer Data. #### 2. Roles of the parties 2.1. For Customer Data, the Controller determines the purposes and means of processing and iTailors processes solely on its behalf. 2.2. For user account data (name, email, role, language, activity logs) and billing data, iTailors is an independent controller under the Privacy Policy published at agentique.eu. Such data falls outside this Agreement. #### 3. Subject matter and duration 3.1. Subject matter: provision of the Agentique platform, which stores and indexes the documents uploaded by the Controller, reads them with the help of AI models and answers users' questions citing the source, extracts data into user-defined tables, runs scheduled workflows and keeps memories across conversations. 3.2. Duration: the term of the Main Contract, plus the period needed to delete or return the data under Section 12. #### 4. Nature and purpose of processing 4.1. Operations: collection (upload), storage, organisation, indexing, optical character recognition (OCR) of scanned documents, generation of vector representations (embeddings), retrieval, extraction, structuring, making available to the Controller's authorised users, and erasure. 4.2. The sole purpose is the provision of the service to the Controller. iTailors does not process Customer Data for its own purposes, does not use it to train AI models, does not sell it and does not make it available to third parties other than the sub-processors listed in Annex 1. 4.3. Optional modules. Google Drive / Google Sheets, SPV / e-Factura ANAF, the Tender Radar (SEAP/SICAP), the Company File, programmatic MCP access and messaging channels (WhatsApp, Messenger and Slack, connected through the Controller's own accounts; the data received consists of the messages the Controller's users send on that channel) are enabled only by the Controller. By default only the in-app chat and the embeddable assistant (iframe) are active. Enabling them constitutes a documented instruction within the meaning of Section 6. #### 5. Types of data and categories of data subjects 5.1. Types of data: any personal data in the Controller's documents (identification, contact, professional, contractual and financial data, CV and invoice data, correspondence) and data derived from it (table extractions, memories, assistant answers, recorded decisions). Special categories of data (Article 9 GDPR) may be uploaded only where the Controller has its own legal basis; iTailors neither requests nor specifically identifies them. 5.2. Data subjects: the Controller's employees, contractors, customers, suppliers and partners; representatives of contracting authorities whose contact details are published on SEAP/SICAP; any other person mentioned in the uploaded documents. #### 6. Controller's instructions 6.1. iTailors processes Customer Data only on the Controller's documented instructions: the Main Contract, this Agreement and the Controller's own workspace configuration (documents uploaded, modules enabled, scheduled workflows, users invited and their roles). 6.2. Further instructions are given in writing to the support address shown on the website. If iTailors considers that an instruction infringes the GDPR or other applicable law, it informs the Controller without delay and may suspend that instruction until the matter is clarified. 6.3. iTailors processes data outside the instructions only where required by Union or Romanian law, in which case it informs the Controller before processing unless the law prohibits such information. #### 7. Confidentiality of personnel Persons authorised by iTailors to process Customer Data are bound by contractual confidentiality obligations or an appropriate statutory duty of confidentiality. Their access is limited to what is strictly necessary to operate, maintain and troubleshoot the platform or to provide support requested by the Controller. #### 8. Security of processing iTailors implements the technical and organisational measures described in Annex 2, taking into account the state of the art, the costs of implementation, the nature, scope, context and purposes of processing and the risks to data subjects (Article 32 GDPR). The measures may be updated provided the overall level of security is not reduced. #### 9. Sub-processors 9.1. The Controller gives general authorisation for the sub-processors listed in Annex 1. 9.2. iTailors gives the Controller at least 30 days' notice, by email or in the application, before adding or replacing a sub-processor. The Controller may object in writing, stating its reasons, within that period. If the parties cannot find a reasonable solution, the Controller may terminate the Main Contract for the affected feature without penalty, with a pro rata refund of any fees paid in advance. 9.3. iTailors imposes on each sub-processor, by contract, data protection obligations at least equivalent to those in this Agreement and remains liable to the Controller for the sub-processor's performance. 9.4. Sub-processors enabled at the Controller's choice (for example Google, for Drive / Sheets) process data only between connection and revocation of access, which can be done at any time. #### 10. Assistance with data subject rights 10.1. Taking into account the nature of the processing, iTailors assists the Controller with appropriate technical and organisational measures in responding to data subject requests (access, rectification, erasure, restriction, portability, objection). 10.2. The Controller can itself search, export and delete documents, conversations, memories and entities from the application. For requests that cannot be handled in the application, iTailors responds to the Controller within 10 business days. 10.3. If a data subject contacts iTailors directly about Customer Data, iTailors refers them to the Controller, informs the Controller without delay and does not respond on the merits except on its instruction. #### 11. Personal data breaches and other assistance 11.1. iTailors notifies the Controller without undue delay and in any event within 72 hours of becoming aware of a personal data breach affecting Customer Data. The notice goes to the workspace owner's email address (owner role) and includes, to the extent known: the nature of the breach, the categories and approximate number of data subjects and records concerned, the likely consequences, and the measures taken or proposed. Information may be provided in phases. 11.2. iTailors cooperates with the Controller in investigating, containing and remedying the breach and in any communications to the Romanian supervisory authority (ANSPDCP) and to data subjects that the Controller deems necessary. 11.3. On request, iTailors provides the information reasonably needed for a data protection impact assessment (Article 35 GDPR) and prior consultation with the authority (Article 36 GDPR), insofar as it relates to the platform. #### 12. Deletion and return of data 12.1. During the term, the Controller may at any time delete documents, conversations, memories and entities from the application. 12.2. On termination of the Main Contract, iTailors deletes Customer Data from production systems within 30 days of termination, except for data it is legally required to retain (billing documents, 10 years under Romanian Accounting Law no. 82/1991). Before deletion, the Controller may request an export of its documents and structured data in the formats available in the application. 12.3. Backup copies are removed according to the backup rotation cycle, within a further 7 days after deletion from production; until then they remain protected by the measures in Annex 2 and are not used for any other purpose. 12.4. On request, iTailors confirms deletion in writing. #### 13. Audit and information 13.1. iTailors makes available the information necessary to demonstrate compliance with Article 28 GDPR: the measures in Annex 2, the sub-processor list, AWS compliance reports (through AWS Artifact) and answers to the Controller's security questionnaires. 13.2. The Controller, or an independent auditor bound by confidentiality, may carry out an audit at most once a year, and additionally when a supervisory authority requires it or after a personal data breach, on 30 days' prior notice, during business hours and without compromising the security of other customers. Audits are conducted through document review and interviews; physical access to AWS infrastructure is replaced by AWS compliance reports. 13.3. Audit costs are borne by the Controller unless the audit reveals a material non-compliance on the part of iTailors. #### 14. International transfers 14.1. In normal operation, Customer Data is stored and processed exclusively within the European Union: infrastructure in the AWS eu-central-1 region (Frankfurt, Germany) and AI models through Amazon Bedrock in EU AWS regions. No transfer to a third country takes place. 14.2. If the Controller connects Google Drive or Google Sheets, data accessed through that integration may be transferred to Google in the United States under Google's standard contractual clauses and its certification under the EU-US Data Privacy Framework. The Controller alone decides whether to enable the integration and is responsible for assessing the transfer. 14.3. Any other transfer takes place only on the Controller's instruction and in compliance with Chapter V GDPR. #### 15. Liability 15.1. Each party is liable for damage caused by its own breach of the GDPR and of this Agreement, in accordance with Article 82 GDPR. 15.2. The limitation of liability in the Main Contract (the amount paid by the Controller in the 12 months preceding the event giving rise to the claim) also applies to this Agreement, except for liability for wilful misconduct or gross negligence and where the law does not permit limitation. 15.3. The Controller warrants that it has a legal basis for the data it uploads and that it has informed data subjects in accordance with Articles 13 and 14 GDPR. #### 16. Final provisions 16.1. The Agreement is governed by Romanian law; disputes fall under the jurisdiction of the Romanian courts. 16.2. In case of conflict between the Agreement and the Main Contract, the Agreement prevails on data protection matters. 16.3. The Agreement is accepted together with the Main Contract when the workspace is created; the version and date of acceptance are recorded. Changes are notified 30 days in advance by email or in the application. 16.4. Data protection contact: the address shown at agentique.eu under "Support contact". iTailors has not appointed a data protection officer, as its activities do not require one. #### Annex 1. Authorised sub-processors | Sub-processor | Service | Processing location | Condition | |---|---|---|---| | Amazon Web Services EMEA SARL | Hosting (S3, RDS MySQL, OpenSearch, ECS, Redis, CloudWatch) and AI models through Amazon Bedrock | EU: eu-central-1 region (Frankfurt); Bedrock inference in EU AWS regions | Permanent | | Google (Google Ireland Limited / Google LLC) | Google Drive and Google Sheets | EU and USA (standard contractual clauses, EU-US Data Privacy Framework) | Only if the Controller connects the integration | | SMTP server configured by iTailors | Transactional email (invitations, digests, alerts); the sender is the address shown in the messages | As per iTailors' SMTP configuration | Permanent | #### Annex 2. Technical and organisational measures 1. Tenant isolation. Each Controller has its own workspace with a separate search index, a separate prefix in file storage and separate database records. Users of one customer cannot access another customer's data. 2. Infrastructure. The platform runs in AWS, eu-central-1 region (Frankfurt), on the services listed in Annex 1. AWS compliance reports (ISO 27001, SOC 2) are available through AWS Artifact. 3. Encryption. All traffic is encrypted with TLS. Data at rest is encrypted: files in S3 with server-side encryption (AES-256) on all buckets, the RDS database is encrypted, and application secrets are protected with AWS KMS. 4. User access control. Access by account and invitation code; passwords are stored only as hashes. Workspace roles: owner, admin, member, viewer, with differentiated permissions. The Controller invites users and revokes access from the application. 5. API keys (MCP access). Issued by the Controller and stored only as hashes; they cannot be recovered from the system afterwards and can be revoked at any time. 6. AI processing. Anthropic Claude and Amazon Titan models are called through Amazon Bedrock in EU regions. Bedrock does not store prompts or responses and does not use them for training; data does not reach Anthropic. iTailors does not train models on Customer Data. 7. Logging. Technical logs are kept in AWS CloudWatch with a retention of 7 days. The application keeps user activity logs (questions, documents uploaded, decisions recorded with user name and date), visible to the Controller. 8. Backups. The database has automated daily backups, retained for 7 days; files are kept in AWS S3, durable storage replicated across several availability zones of the Frankfurt region. Backups remain in the eu-central-1 region. 9. iTailors staff access. Restricted to the people who operate the platform, on a need-to-know basis, under a duty of confidentiality. Document content is not accessed except for troubleshooting or at the customer's request. 10. Integrations and automation. Google Drive / Sheets use OAuth in the Controller's own Google account; SPV / e-Factura uses the Controller's qualified certificate. Automated actions are enabled explicitly by the Controller and can be stopped at any time. --- ### Security and data URL: https://agentique.eu/legal/securitate (Romanian) · https://agentique.eu/en/legal/securitate (English) Version 1.0, effective 8 September 2026. This page answers the questions that usually come up in security and public procurement questionnaires, and can be quoted directly. Provider: ITAILORS - SOFT S.R.L., tax ID (CUI) 48450662, Romania. Product: Agentique (agentique.eu). #### 1. Where the data lives - All data is hosted in the European Union on Amazon Web Services, eu-central-1 region (Frankfurt, Germany): files, database, search index, application. - AI models (Anthropic Claude, Amazon Titan for embeddings) are called through Amazon Bedrock, exclusively in EU AWS regions. - In normal operation there is no transfer of data outside the EU/EEA. - The only exception, enabled by the customer alone: if it connects Google Drive or Google Sheets from its own Google account, the data accessed may reach Google (USA) under standard contractual clauses and the Data Privacy Framework. The integration can be revoked at any time. #### 2. Who can see the data - Each customer has an isolated workspace: separate search index, separate storage prefix, separate database records. Users of one customer cannot see another customer's data. - Access requires an account and an invitation code and is role-based: owner, admin, member, viewer. The customer invites and removes users itself. Passwords and API keys (MCP access) are stored only as hashes. - iTailors staff access is restricted on a need-to-know basis and subject to confidentiality. Document content is not viewed except for troubleshooting or at the customer's request. - All traffic is encrypted with TLS. Data at rest is encrypted: files in S3 with server-side encryption (AES-256) on all buckets, the RDS database is encrypted, and application secrets are protected with AWS KMS. #### 3. What the AI does with the data - The assistant reads the customer's documents, answers with the source document and page, extracts data into user-defined tables and runs scheduled workflows the customer has explicitly enabled. Scanned documents are read by OCR with the same model (Claude Vision), also within the EU. - Amazon Bedrock does not store prompts or responses and does not use them for training. Data does not reach Anthropic. iTailors does not train models on customer data. - AI answers can be wrong; that is why every answer shows its source, so a person can check it. In the tender module, verdicts are recommendations only: the decision stays with a human, and the platform never submits anything to SEAP. #### 4. What is retained, and for how long - Documents, conversations, memories, entities and activity logs (questions, uploads, decisions recorded with user name and date): for the term of the contract, or until the customer deletes them. - User account data (name, email, role, language): for the term of the contract and a further 30 days after termination, then deleted. - Technical logs (AWS CloudWatch): 7 days. Backups: the database has automated daily backups, retained for 7 days; files are kept in AWS S3, durable storage replicated across several availability zones; they stay in the Frankfurt region. - Billing documents: 10 years, under Romanian Accounting Law no. 82/1991. - Web Assistant embedded on the customer's website, visitor conversations: retained for 30 days, for technical support and so that the institution can review them, then deleted; the institution may request earlier deletion. #### 5. How data is exported and deleted - The customer deletes documents, memories, entities and conversations itself, from the application, at any time. - Exports use the formats available in the application; on request, iTailors provides a full export before termination. - On termination, data is deleted from production within 30 days of termination, except for what must be kept by law (invoices). Backup copies are removed according to the rotation cycle. Written confirmation of deletion on request. - Data subject requests (access, rectification, erasure, portability) are handled by the customer from the application; iTailors assists where needed. #### 6. How to report an incident - Contact: the address shown at agentique.eu under "Support contact". No data protection officer has been appointed, as iTailors' activities do not require one. - A security incident affecting customer data is notified to the workspace owner (owner role) without delay and within 72 hours of discovery: nature of the incident, data affected, likely consequences, measures taken. Information may come in phases. - iTailors cooperates with the customer on the investigation and, where required, on notifying the Romanian supervisory authority (ANSPDCP) and the data subjects. #### 7. What we do NOT do - No visitor tracking: agentique.eu uses no Google Analytics, Meta Pixel, Hotjar or other third-party scripts; fonts are self-hosted. - No session cookies; the authentication token and preferences live in the browser's localStorage and are never sent to third parties. No consent banner is required. - No model training on customer data. No selling or renting of data. - No card data collected: payment is by invoice issued by iTailors, settled by bank transfer within 15 days of issue. - No bids submitted to SEAP and no automated actions the customer has not explicitly enabled; any automation can be switched off at any time. #### 8. Legal framework and documents available - Law: GDPR, Romanian Law no. 190/2018, Law no. 506/2004. Supervisory authority: ANSPDCP (www.dataprotection.ro). For document content, the customer is the controller and iTailors the processor; for account and billing data, iTailors is the controller. - Available on request: the Data Processing Agreement (Article 28 GDPR), the sub-processor list (AWS; Google, only if connected; the SMTP server configured by iTailors, for transactional email), AWS compliance reports (ISO 27001, SOC 2), and answers to the institution's own questionnaire. SLA: during the beta period the service is provided without an availability guarantee (best effort); the platform may be unavailable at night, between 23:30 and 09:00 Romanian time, for maintenance. --- ### Agentique Web Assistant v1 URL: https://agentique.eu/legal/asistent-web (Romanian) · https://agentique.eu/en/legal/asistent-web (English) Version 1.0, effective 8 September 2026. For public institutions and companies that embed the Agentique Web Assistant on their own website. Three parts: product sheet, specific terms for the embedding institution, and a ready-to-display notice for visitors. The service is provided by ITAILORS - SOFT S.R.L. („iTailors”), a Romanian limited liability company, tax ID (CUI) 48450662, Trade Register no. J2/1018/2023, registered office at Calea Aurel Vlaicu, Arad, Arad County, Romania. Website: https://agentique.eu. Support and data protection contact: the address shown on the website under „Support contact”. #### A. Product sheet ##### 1. What it is The Agentique Web Assistant v1 is a chatbot that the institution displays on its own website. Visitors ask questions in plain language and the assistant answers solely from the documents the institution has published in its Agentique workspace: regulations, forms, decisions, guides, fee schedules. Every answer names the document and page it comes from, so visitors can check the source. The official information is the cited document, never the generated text. ##### 2. What it does - Answers using only the documents the institution has made available (PDF, DOCX, XLSX, XML, TXT, scanned images). - Cites the source: document name and page number. - Says when it cannot find the answer in the documents, instead of improvising. - Replies in the language of the question, in Romanian or English. ##### 3. What it does not do - It does not browse the internet and uses no source but the institution's documents. - It makes no decisions on the institution's behalf and issues no documents, opinions or approvals. - It does not file applications, complaints or petitions and has no access to the institution's internal systems. - It does not ask visitors for an account, name or other identifying details and stores no such data; visitors do not need to log in. - It is not used to train AI models. - Answers may be wrong or incomplete, which is why the source is always shown. ##### 4. How it is installed iTailors provides a code snippet that the institution adds to the pages where the assistant should appear: an HTML script configured with the identifier of the workspace dedicated to the assistant, with options for language, colour, position and welcome text. Designated staff upload documents to the institution's Agentique workspace. The institution decides which documents the assistant may use by placing them in the workspace dedicated to the assistant (the assistant answers from the whole workspace, not from a per-document selection) and can update or withdraw them at any time. ##### 5. Where the data is - Hosting: Amazon Web Services, region eu-central-1 (Frankfurt, Germany): files (S3), database (RDS MySQL), search (OpenSearch), application servers (ECS), queues (Redis). - AI models: Anthropic Claude and Amazon Titan through Amazon Bedrock, in EU AWS regions (Frankfurt, Ireland, Paris, Stockholm, Milan, Spain). Bedrock neither stores prompts and responses nor uses them for training; the data does not reach Anthropic. - Each institution has an isolated workspace: separate search index and storage prefix, kept apart from other customers. - Encryption: TLS in transit; at rest, files in S3 with server-side encryption (AES-256) on all buckets, the RDS database encrypted, and application secrets protected with AWS KMS. - Technical logs: AWS CloudWatch, retained for 7 days. - No transfer outside the EU/EEA in normal operation. ##### 6. Common requirements of public institutions | Requirement | How it is met | |---|---| | Data hosted in the EU | Storage and processing in AWS Frankfurt; AI models run in EU AWS regions. | | No transfer outside the EU | None in normal operation. Google (USA) is involved only if the institution connects Drive/Sheets, which the assistant does not need. | | Conversation log on request | Visitor conversations are retained for 30 days, for technical support and for the institution to consult, then deleted; the institution may request earlier deletion. During that period they are provided to the institution on request. | | Export and deletion | The institution deletes documents and conversations in the application. Export on request, in JSON format (and CSV for tables), within 10 working days. At contract end, data is deleted within 30 days; backups expire within a further 7 days. | | Role-based access | Workspace roles owner, admin, member, viewer; invitations by email; API keys stored only as hashes. | | Security incident notified within 72 hours | See Part B, point 7. | | No training, no sale of data | iTailors neither trains models on the institution's data nor sells data. | | No tracking tools | The platform uses no Google Analytics, pixels or other third-party scripts. The snippet uses localStorage in the visitor's browser only to keep the current conversation; it sets no cookies and does not track the visitor. | #### B. Specific terms for the institution that embeds the assistant ##### 1. Roles of the parties For visitors' data and the content of the published documents, the institution is the controller and iTailors is the processor. iTailors processes this data only on the institution's instructions, under the Data Processing Agreement (DPA). For the institution's user accounts (name, email, role, language, activity logs) and billing data, iTailors is the controller, as described in the Privacy Policy. ##### 2. Notice for visitors The institution displays the Part C notice next to the chatbot, visibly, before the first question, with its name and privacy policy link filled in. The wording may be adapted as long as it still says that visitors are talking to an automated assistant, that answers come from published documents and may be wrong, what data is processed and where the policy is. ##### 3. Published content The institution is responsible for the documents it makes available: accuracy, currency, the lawfulness of publishing them and, where they contain personal data, the legal basis for processing it. iTailors does not review or edit the content. ##### 4. Logs Visitor conversations are retained for 30 days, for technical support and for the institution to consult, then deleted; the institution may request earlier deletion. Access to logs is role-based, limited to people the institution designates. ##### 5. Export At the institution's request, iTailors provides the uploaded documents and, if retained, the conversations, in JSON format (and CSV for tables), within 10 working days of the request. ##### 6. Deletion The institution can delete documents, conversations, memories and entities in the application at any time. At contract end, iTailors deletes the institution's data within 30 days of termination (backups expire within a further 7 days), except invoicing records, kept for 10 years under the Romanian Accounting Act no. 82/1991. ##### 7. Security incident If iTailors becomes aware of a security breach affecting the institution's data, it notifies the institution without undue delay and within 72 hours at the latest. The notification describes the incident, the categories of data and people affected, and the measures taken and recommended. iTailors supports the institution in notifying the supervisory authority (ANSPDCP) and, where required, the people concerned. ##### 8. Rights of data subjects Visitors' requests about their data (access, erasure, objection and the other GDPR rights) are handled by the institution. iTailors assists in time for the institution to reply within 30 days. ##### 9. Sub-processors Amazon Web Services EMEA SARL (hosting, Amazon Bedrock). Google, only if the institution connects Drive/Sheets. The full list and the change notification procedure are in the DPA. ##### 10. Availability During the trial the service is provided „as is”. Availability for subscriptions: during the beta the service is provided without an availability guarantee (best effort). The current environment is a beta with scheduled night-time downtime, between 23:30 and 09:00 Romanian time, for maintenance; the institution should bear this in mind before showing the assistant to the public. ##### 11. Applicable documents This document is supplemented by the Agentique Terms and Conditions, the Privacy Policy and the Data Processing Agreement (DPA), available at https://agentique.eu/legal/termeni, https://agentique.eu/legal/confidentialitate and https://agentique.eu/legal/dpa. For personal data processing, the DPA prevails. Romanian law applies and disputes are settled by the courts of Romania. Changes are notified 30 days in advance, by email or in the application. #### C. Notice for visitors For the institution to display next to the chatbot, replacing the text in square brackets. You are talking to an automated assistant, not a member of staff. Its answers are generated from the documents published by [institution name] and may contain errors or omissions. Each answer names the document and page it comes from; for official information, read that document or contact the institution. The questions you type and connection details (IP address, browser type) are processed on the institution's behalf by the service provider, iTailors, on servers in the European Union, and are not used to train AI models. Please do not enter personal data such as ID numbers, bank details or medical information. Privacy policy and your rights: [address of the institution's privacy policy]. --- ## Full blog content The complete text of every post, in English. Each post also exists in Romanian at the URL listed with it. ### The Interface Layer, part 1: the job nobody designed (2026-09-02) URL: https://agentique.eu/blog/munca-dintre-sisteme-1 (Romanian) · https://agentique.eu/en/blog/munca-dintre-sisteme-1 (English) This is the first chapter of *The Interface Layer, in practice*, the working companion to [the thesis](https://agentique.eu/blog/munca-dintre-sisteme-teza). The thesis says why; these chapters say how, and they do not retell the argument. Written by the team at iTailors, the company behind Agentique, for the people who run companies and the people who run their back offices. This chapter is about finding the job. The ones that follow are about handing it over safely, measuring whether it worked, replacing forms with intake, and how a Romanian company starts. #### Three offices **A distributor.** Forty suppliers, a few hundred incoming invoices a month, an ERP, an accounting program, a bank portal, the state's e-invoicing system, and a shared inbox. One person in the office knows how they all fit. Each morning she downloads what arrived in SPV, opens each invoice, finds the purchase order it belongs to, checks the delivery note against it, fixes the supplier name that came in spelled a new way, types the totals into accounting, files the PDF in the right folder, and emails the two suppliers whose quantities did not match. She was hired as an office manager. Nobody wrote this list into her contract; the list wrote itself. **A transport company.** Trucks, clients, subcontractors. A dispatcher matches delivery confirmations to orders, invoices the client from one screen, enters the transport declaration in another, answers "where is my shipment" from a third, and keeps the real state of the world in a spreadsheet that only he understands. When he is on holiday, the spreadsheet is on holiday too. **An accounting firm.** One accountant, thirty-five client companies. Every month, thirty-five downloads from SPV, thirty-five reconciliations against what each client sent by email or by photo, a hundred small questions asked on WhatsApp because an invoice has no purchase order or a scan is half legible, and thirty-five sets of entries. The accountant's expertise is tax. Most of her hours are logistics. The three have nothing in common on paper. They share one job: moving information between systems that never learned to talk to each other, with judgment applied at every step. Nobody designed that job. It accreted, one system at a time, and it landed on whoever was closest. #### The five shapes Look closely at any of the three offices and the job decomposes into five recurring motions. **Transcription.** A value read in one place and typed into another. Invoice totals from a PDF into the ledger. Order lines into a supplier's portal. Delivery quantities from a photo into the ERP. **Reconciliation.** Deciding that two records describe the same thing. The supplier known under three spellings. The payment that settles this invoice and not the other one with the same amount. The delivery that closes that order. **Translation.** Converting meaning between systems that model the world differently. Their "customer" is our "account". Their payment term is our code 4. Their product code is our internal SKU, except for the ones that were renamed last spring. **Routing.** Deciding where something goes and who must see it. Which cost center. Which approver, given the amount. Which folder, which client, which month. **Exceptions.** Everything that does not fit. The credit note without an invoice. The duplicate that is not quite a duplicate. The scan cut off at the bottom. The email that says "same as last time" about a last time nobody remembers. The first three shapes generate volume. The last two generate judgment. A useful thing to notice already: every previous attempt to automate this job attacked the first three and stopped at the last two. That is chapter two's subject. For now, the point is that the five shapes are how you will recognize the job when you go looking for it. #### Why it grows faster than the company Systems do not exchange information with each other; they exchange it through pairs. With two systems there is one seam between them. With four there are six. With six there are fifteen. The formula is n multiplied by n minus one, divided by two, and it is merciless: every system you add creates a seam with every system you already have. Now count honestly what a system is. The ERP and the accounting program, obviously. But also the bank portal, the e-invoicing system, the transport declaration portal, the supplier's portal, the shared inbox, the WhatsApp group with the drivers, and the spreadsheet that holds the real state of the world. Each of those is a place information lives and must be moved out of by hand. A company that thinks it has three systems usually has nine, and thirty-six seams. Here is the part that stings. Most of those seams were added in the name of saving time. A new portal, a new tool, a new channel, each one useful on its own, each one adding seams to everything that already existed. The company gets more capable and the person in the middle gets more seams. The seams never appear on any invoice. They appear as the office manager staying late. #### What it really costs The visible cost is the hour: the time a knowledgeable person spends transcribing, matching, and re-keying. That hour is real and is the smallest part. **Switching.** Every seam is a change of screen and a reload of context. A person crossing fifteen seams in a morning is not doing fifteen small tasks; they are doing one large task with fifteen interruptions built into it. The work that requires concentration, the actual judgment, gets the leftover attention. **The error loop.** A number mis-keyed in March is found in April by someone else. Finding it costs more than typing it did; fixing it costs more than finding it; explaining it to the supplier, the accountant, and sometimes the tax authority costs more than all of the above. The interface layer is where errors enter, and the layer's hourly cost never includes the loop. **Latency.** Things wait for the person. The invoice is approved when she gets to it. The client learns where the shipment is tomorrow. The month closes when the last reconciliation is done. None of this shows up as cost; it shows up as a company that is slower than its systems. **Key-person risk.** The day the person is sick, on holiday, or gone, the company discovers it was running an undocumented system whose only copy lived in one head. The handover, if there is one, is a list of screens and a shrug at the judgment calls. **The talent cost.** You hired a capable person because the judgment calls need one. Then you filled a large part of their day with the copying that surrounds the judgment. Capable people notice. The ones who stay learn to work around the seams instead of through them, which is how the spreadsheet with the real state of the world is born. #### Why nobody sees it There is no line in the budget called "interface layer". The work is distributed across roles that all have other names. The people doing it are competent and quiet, which is exactly why they were given it. Management sees the outputs, the closed month, the paid invoices, and not the seams that produced them. Software vendors sell modules, not seams; a demo shows one system doing its job beautifully, never the moment its output has to enter a different system by hand. Consultants sell new systems, and a new system adds seams to every old one. So the job stays invisible, and invisible costs do not get fixed. The first move, before any technology, is to make it visible. #### The one-page audit This takes an hour, and it is the most useful hour a company can spend before touching any AI. Sit with the two or three people who actually do the work; they know the seams better than anyone. **List every system.** Everything that information lives in. Include the inbox, the messaging groups, the portals, and the spreadsheets. If a person moves information out of it by hand, it is a system. **Map the seams.** For every pair that exchanges information through a person, write one line: what moves, from where to where, who does it. **Time each seam.** How often it happens and how many minutes each time. Be honest about the interruptions; a five-minute task done twelve times a day is an hour and a broken morning. **Name the judgment.** For each seam, what decision does the person make while crossing it? Matching two records, classifying a document, choosing a cost center, judging whether a number is plausible, deciding what to do with an exception. Write it in plain words. This column is the one that matters later. **Rank.** Multiply hours by month, weigh by what an error at that seam costs, and sort. Expect a handful of seams to carry most of the hours and most of the risk. The inventory will show you which. What you hold at the end is the seam inventory: one page that describes the job nobody designed, in the company's own words, with the judgment inside each seam named. Most companies have never seen this page. Reading it is usually a quiet moment. #### What to do with the page Two things not to do. Do not buy a new system to fix it; a new system is a new set of seams. And do not automate the easy seams first because they are easy; the easy seams are the ones where a machine already does most of the work, and the hours are in the hard ones. Look at the judgment column instead. For fifty years, that column was the reason the job stayed human. Chapter two is about why that stopped being true, and where it is still true. The seam inventory is also the document we ask for in a first conversation. If you make one, we would like to read it. --- ### The Interface Layer, part 2: handing a seam over safely (2026-09-02) URL: https://agentique.eu/blog/munca-dintre-sisteme-2 (Romanian) · https://agentique.eu/en/blog/munca-dintre-sisteme-2 (English) [Part 1](https://agentique.eu/blog/munca-dintre-sisteme-1) ended with a page: the seam inventory, every place a person carries information between systems, with the judgment inside each seam named. This chapter is about what to do with it. It applies to any implementation of [the loop](https://agentique.eu/blog/munca-dintre-sisteme-teza), ours included, and it is written so that a company could follow it with its own team. #### Which seams first The inventory tempts you toward the biggest number. Resist it for a week. The first seams you hand over teach the company how this works, and a bad first seam teaches the wrong lesson. Pick by these five criteria, in this order. **The judgment repeats.** Look at the judgment column. A seam where the decision has the same shape every time ("which purchase order does this invoice belong to") is ready. A seam where every case is a different kind of exception is not ready; it is where you go third, once the layer has memory of the house rules. **The input is at least half structured.** E-invoices, bank exports, order confirmations, delivery notes with a fixed layout. Photos of handwritten notes are possible and come later. **A wrong decision can be undone.** A mis-matched invoice caught at approval costs a minute. A wrong tax declaration costs a quarter. Start where mistakes are cheap, because there will be mistakes, and the point of the first seam is to see them. **The hours are real.** Among the seams that pass the first three tests, take the one with the most monthly hours. Nobody should have to squint to see the effect. **The owner is willing.** The person who crosses that seam today will design the handover with you, and will become its approver. If they dread it, pick another seam and come back. Chapter 1 named them; this is where their judgment gets written down. Two seams that usually pass all five: supplier invoices into the ledger, and client status questions answered from operational systems. Two that usually fail the third test and should wait: anything filed with the tax authority, and anything that sends money. #### One seam, decomposed Take the supplier-invoice seam at the distributor from chapter 1 and lay it on the loop. **Read.** The invoice itself, structured if it came through e-invoicing, scanned if it came by email. The purchase orders open for that supplier. The delivery notes of the last weeks. The rule the office manager stated once: "this supplier's freight line goes to logistics, not to goods." **Resolve.** Which supplier record, given that the name arrived with a new spelling. Which purchase order, given that the amounts match two of them and only one is for the delivered quantity. Which cost centers, per line, given the rule above. **Propose.** The ledger entry, line by line, with each value pointing at its source: this total from page one of the invoice, this quantity from the delivery note, this cost center from the stated rule. And a plain-language note where something is off: "quantity on the invoice exceeds the delivery note by twelve units." **Confirm.** The office manager sees the proposal and the evidence, corrects the one line that needs it, approves. Thirty seconds instead of eleven minutes, and the mistake she catches is the system's, not hers. **Record.** The entry is written to the accounting program, the invoice is filed and linked, the supplier record gains the new spelling as an alias, and the correction she made becomes a memory: next time, the system proposes it her way. Every seam in the inventory decomposes like this. Doing it on paper for the first two seams, with the owner, is an afternoon, and it is the design document. #### Which step needs which model Not every step of the loop needs the same machinery, and most of the cost of a bad implementation is using the expensive tier for work a cheaper one does better. Assign each step a tier. **Tier zero, no model.** Parsing the structured e-invoice fields. Looking up the supplier by tax id. Exact matching of an invoice number. Validation rules the accounting program already has. Scheduling. This tier is exact, free, and should carry everything it can. **Tier one, small and fast.** Classifying a document ("this is a credit note, not an invoice"). Extracting fields from a semi-structured PDF. Producing the shortlist of candidate purchase orders. Deciding which of three fixed actions a message asks for. **Tier two, the strongest model.** Reading a half-legible scan. Choosing between two candidates when the amounts match both. Writing the note that explains the discrepancy. Handling the case nobody wrote a rule for. The discipline is a single rule: push every step down to the cheapest tier that meets the precision bar, and measure the bar, which is chapter 3. In the invoice seam above, roughly everything is tier zero and one except the ambiguous match and the explanation. That is typical, and it is why the layer is affordable. #### Where the confirmation boundary goes The default is simple and should be written down before anything runs: anything that leaves the company or changes a record of consequence waits for a person. Sending an email to a supplier, posting an entry, changing a payment field, deleting a document. Reading, resolving, and drafting run freely; acting waits. The default is not the end state. Autonomy is earned per class of decision, never granted per system. A class of decision is "matching an e-invoice to a single open purchase order with an exact amount", not "invoices". For each class, the approver's own behavior tells you when it is safe: after a run of approvals with no corrections, long enough to have seen the month's variety, that class can move to automatic with a sample reviewed. One in ten at first; the sample is where drift will show up. Classes that fail the diagram's question, the irreversible and the outbound, stay at propose-and-confirm permanently. Their value is in the proposal, not in the autonomy. Two rules that hold at every level of autonomy. **A proposal with no evidence is not a proposal.** If the system cannot point at the document and the field a value came from, it must say so and stop, and the design should make silent guessing impossible rather than discouraged. And **the "I could not determine this" path must exist and must be visible**: a queue of unresolved items that a person works through daily. The size of that queue is a health signal, and each item resolved is a candidate for a new memory or a new rule. #### The mechanics that prevent the classic failures **Idempotency.** Every write is keyed by an external identifier: the invoice number and supplier, the message id from the e-invoicing system, the order number. Running the same seam twice, because a schedule fired twice or a person clicked twice, must produce nothing new. Duplicates are the fastest way to lose trust, and they are entirely preventable. **Least privilege, assuming a leak.** The layer gets read access broadly and write access narrowly. Credentials are scoped to the job, rate-limited, and stored as if they will one day be exposed, because some credential somewhere always is. A key that can read invoices should not be able to email clients. **Memory with an owner.** Rules stated by the approver ("treat these two suppliers as one") become memory the system applies. Memory needs an owner and a list; unowned memory is how a wrong rule from March quietly shapes every decision in June. **An audit trail from day one.** Who approved what, on what evidence, when. Less for the auditor than for the Tuesday when a number is wrong and the person who has to find out why has ten minutes. #### The rollout **Shadow.** For two to four weeks the system produces proposals and the person keeps doing the seam the old way. At the end of each day, compare. This is where you discover the spelling variants, the rule nobody stated, the one supplier whose invoices always look wrong. Nothing has been risked, and the design document gets its corrections. **Assisted.** The person now works from the proposals: reviews, corrects, approves. This is the state most seams live in for a long time, and it is already most of the value: the hours drop, the errors are caught at approval, the judgment is captured as memory every time a correction is made. **Sampled.** For the classes that pass the decision above and have earned it through the approval record, automatic with a reviewed sample. This is the state that returns the last of the hours, and it is entered one class at a time, never by flipping a switch on the whole seam. Two to six weeks per seam is a realistic pace with a willing owner. Faster usually means the shadow stage was skipped, and the shadow stage is where the cheap mistakes are found. #### What goes wrong **Automating the exceptions first.** They are the most annoying part of the person's day, so they attract attention. They are also the part with the least pattern. Do them third. **Silent drift.** A supplier changes its invoice layout; a portal renames a field. The parser at tier zero starts producing wrong values with full confidence. The sampled review and the unresolved queue exist to catch this; if nobody looks at them, nothing catches it. **Proposal fatigue.** If the approver sees forty low-stakes confirmations a day, they stop reading them, and the boundary becomes theatre. Move the safe classes to sampled autonomy and keep the human attention for the decisions that need it. **Tool sprawl.** Each new source connected is a new seam with the layer itself. Connect what the inventory ranked, in that order, and stop when the hours are back. #### What the handover produces At the end of one seam handed over you have: a design document that names the judgment; a memory of the house rules in the owner's words; an approval record that shows, class by class, what the system gets right; an unresolved queue that shows what it does not; and an approver whose day is shorter. You also have the raw material for the only question that matters next: did it work, and how would you know? That is chapter 3. --- ### The Interface Layer, part 3: measuring whether it worked (2026-09-02) URL: https://agentique.eu/blog/munca-dintre-sisteme-3 (Romanian) · https://agentique.eu/en/blog/munca-dintre-sisteme-3 (English) [Part 2](https://agentique.eu/blog/munca-dintre-sisteme-2) ended with two artifacts every handover produces on its own: an approval record, decision by decision, and a queue of what the system could not resolve. Those are the instruments. This chapter is about reading them. It is deliberately dry, because the alternative to measuring is believing, and believing is how companies end up paying for a feeling. #### The six numbers Each of these has a plain definition, a source you already have, and a reason it exists. Four are about the system; two are about the company. **Resolution precision.** Of the times the system said "this invoice matches that order" or "this is the same supplier", how often was it actually right, judged by the final outcome? A proposal corrected at approval counts as wrong, and so does one that was accepted and found wrong later. Computed per class of decision from the approval record plus the escapes traced back to it. Measured per class, never as one blended number, because a blended number hides the class that is failing. **Approval rate.** Of the proposals a person saw, how many did they accept without touching, at the moment they saw them? This looks like precision but measures behavior rather than truth: a person can wave through a wrong proposal, which shows up later as an escape, or correct a right one out of habit. What it tells you is whether the judgment is actually being taken and trusted. A system can be precise on what it proposes and still propose so little, or so cautiously, that the person is doing the job with extra steps. A rising approval rate on a class is the evidence that earns it sampled autonomy; a falling one is the earliest warning you get. **Escape rate.** How many wrong values reached the ledger, the client, or the authority? This is the number that matters most and is the hardest to see, because an escape is by definition something nobody caught at the time. It surfaces later: a supplier's complaint, a correction at month close, a query from the accountant. Count each one and trace it back to the decision that produced it. A system that never reports escapes is a system whose escapes nobody is counting. **Time to close.** Per item, from arrival to recorded, with the human step included. The system's part is seconds; the human's part is where the number lives. If time to close does not fall after handover, the proposals are not being trusted or the boundary is in the wrong place. **Queue age.** The unresolved queue has a size and an age. Size tells you how much judgment the system is declining to take; age tells you whether anyone is working the queue. A queue that grows and ages is a handover that stopped halfway. **Hours returned.** The business number, and the only one an owner will ask for. The seam's monthly hours from the [inventory](https://agentique.eu/blog/munca-dintre-sisteme-1), minus the hours the approver now spends on it. Measured honestly it includes the queue work and the sample reviews. It should be reported next to escape rate every time, because hours returned with escapes rising is a loss dressed as a saving. One more, derived: **cost per correct decision**. Everything spent on the seam in a month, model calls and people included, divided by the number of decisions that were correct. This is the unit economics of the interface layer, and it is what "cheaper model" or "more automation" must be judged against. A cheaper model that lowers precision often raises this number. #### The baseline you get for free The question "did it work" needs a "compared to what". The shadow stage from part 2 provides it without any extra effort: while the person keeps doing the seam the old way and the system proposes in parallel, you are measuring both at once. The person's time per item, their own error rate found at month close, and the system's precision against the person's decisions all come out of the same weeks. Record the baseline before the assisted stage begins and keep it. Six months later, when the seam feels normal and nobody remembers the eleven minutes per invoice, the baseline is the only thing that shows what changed. #### What a demo cannot show Every vendor demo, ours included, shows a document read correctly and a fluent answer. That demonstrates capability, which is real and is not the question. The question is what happens on your documents, with your spelling variants, over a month that contains a format change, a new supplier, and a week when the approver was on holiday. Precision on your variants, drift, the queue, and escapes are all things only a month shows. There is no way to shorten this, and a vendor who claims there is has not measured their own product. #### Reading the patterns The numbers are more useful together than apart. A few combinations recur. **Precision high, approval rate falling.** The system is right but the person has stopped trusting it, usually because of one visible mistake or because proposals are arriving without their evidence. Look at the last correction they made and what the proposal showed them. **Approval rate high, escape rate rising.** The dangerous one. Approvals have become a reflex, the boundary has turned into theatre, and errors are passing through with a signature on them. Reduce the volume the person sees by moving safe classes to sampled autonomy, so their attention returns to the decisions that need it. **Queue growing, precision stable.** The system is declining more cases than before. Usually a source changed: a new invoice layout, a renamed field, a partner switching formats. This is drift, and the queue caught it as designed. **Hours returned flat, everything else fine.** The approver is re-doing the work before approving, either from habit or because the proposal is not laid out in the order they work. Watch them for an hour. **Precision falling on one class only.** Something specific: a supplier whose invoices break the pattern, a rule in memory that was right in March and wrong now. Per-class numbers exist for exactly this moment. #### The cadence Weekly during shadow and assisted stages, on one page, reviewed with the approver present. Monthly after that. The review is a fifteen-minute conversation about the numbers, the queue, and the one thing to change; it is not a report to be filed. #### The evaluation protocol These are the questions to put to any vendor of an interface layer, including us. They are questions a serious one will welcome. - **Define your metrics.** Ask for their definitions of precision, approval, escape, and time to close. If the definitions are not per class of decision, the numbers will hide the failing class. - **Run in shadow on our data.** Two weeks minimum, on the seam we chose, against the person who does it today. The output is a baseline and a per-class precision number, or it is a sales call. - **Show us the queue.** Where does "I could not determine this" go, who sees it, and how old is it allowed to get? - **Show us a correction becoming memory.** The approver fixes one line; what happens next time? - **Show us an escape.** What is the path from "a wrong value reached the ledger" back to the decision and the evidence that produced it? How long does it take? - **Where does the boundary live?** In code that cannot be talked past, or in instructions to the model? - **What can this credential do?** Read what, write what, and what happens if it leaks? A vendor that answers all seven with specifics is one you can measure. A vendor that answers with a demo is one you can only believe. #### The limits of measuring Small numbers lie. A class with fourteen decisions a month cannot show a precision difference between two setups; wait for the numbers or merge classes carefully. Seasonality matters: the month with the year-end invoices is not the month of August. And the oldest trap of all, the measured number becoming the goal: an approval rate rewarded on its own will rise for the wrong reason. Pair it with escape rate, always, and let the two argue. Measurement will not make the interface layer work. It will tell you, week by week, whether it is working and where it is not, which is the difference between a system you run and a system that runs you. Part 4 turns to the other side of the same layer: the forms your people and your customers fill, and what replaces them. --- ### The Interface Layer, part 4: forms, the human interface (2026-09-02) URL: https://agentique.eu/blog/munca-dintre-sisteme-4 (Romanian) · https://agentique.eu/en/blog/munca-dintre-sisteme-4 (English) The first three chapters were about seams between systems. This one is about the seams between people and systems, and every company has a name for those: forms. The supplier registration form, the expense claim, the onboarding packet, the service request, the order entry screen. They are the interface layer's other side, and they carry the same hidden cost with an extra ingredient. People hate them, and the hatred is deserved. #### What a form actually asks of a person A form is the user interface of the deterministic era. It is a system declining to do any judgment and asking the human to do all of it in advance: read the document you hold, find the twelve values the system wants, translate each into the system's vocabulary, type them in the system's order, and satisfy validation rules whose logic you cannot see. The form asks for the tax id that is printed on the invoice in your hand. It asks for the address the company already has from last year. It has no memory of the last time you filled it. When you get something wrong, it answers in red. The hidden costs follow directly. **Abandonment**: the customer who closes the tab, the supplier who "will send it later". **Fake data typed to satisfy validation**: the date that is always the first of January, the field that says "n/a" because the form would not proceed without it, now permanent in the record. **The phone call after the form**: because the form could not ask a follow-up question, a person has to. And the same key-person problem as chapter 1: the one employee who knows how to fill the form correctly, because the form's real rules were never written anywhere. #### Inventory the forms The audit from [part 1](https://agentique.eu/blog/munca-dintre-sisteme-1) extends naturally. List every form, internal and external. For each, write down: who fills it, how many times a month, how long it takes, how often it is abandoned or comes back wrong, and, the column that matters, **which of its fields already exist in a document the person holds**. For a supplier registration, nearly all of them exist on the company's registration certificate and its first invoice. For an expense claim, they exist on the receipt. For a service request, they exist in the email that prompted it. That column tells you the shape of the replacement. A form whose fields all exist in a document is not a form; it is a transcription seam wearing a user interface. #### The intake loop **Drop what you have.** A document, a photo of one, a forwarded email, or a sentence. The person gives the system what they actually possess, in the form they possess it. Nothing is pre-structured by a human. **Extract and resolve.** The fields come out of the document by parser where it is structured and by reading where it is not. Then the same resolution as chapter 2: is this supplier already known, does this employee exist, which project is this claim against. The system fills what it can and knows what it could not. **Ask only the gaps.** Two or three questions, in the person's words, about the things no document could have said: which department, which of two matching projects, whether the delivery address is the same as last time. A rule with no exceptions: never ask what the document already said. Every question that breaks that rule teaches the person the system did not read what they gave it. **Review.** The form does not disappear. Its job changes, from input surface to review surface. The person sees it prefilled, each value carrying a small marker of where it came from, corrects the one thing that needs it, and confirms. This is the confirmation boundary from chapter 2 applied to the human side, and it answers the power users who genuinely prefer a form: they get one, already filled, with a keyboard. **Record.** Validation runs exactly as before; the form's rules encoded real constraints and they stay. What goes is the typing, the translating, and the follow-up call. The schema stays; the labor goes. That sentence is the whole design. #### Which tier does the work Intake is one of the cheapest interface-layer jobs per correct record, because most of it is the bottom tiers. A parser reads the structured document. A small model classifies which intake this is and extracts fields from the semi-structured ones. Validation is deterministic and already written. Resolution against existing records is a lookup plus a small model for the ambiguous cases. The strongest model appears in two places only: the gap conversation, which must sound like a person and handle whatever the person answers, and the genuinely ambiguous document. The push-down rule from chapter 2 applies unchanged. #### Designing the gap conversation This is where intake succeeds or becomes a worse form. - **Fewest questions.** If the system needs five answers, the extraction is not finished; go back and read harder before asking. - **The person's vocabulary.** "Which team is this for?" and not "select cost center code". - **Accept any shape of answer.** "The Cluj warehouse, like last month" is a complete answer; the system resolves it. - **Show what was understood** before asking the next thing, so a wrong extraction is caught at question two rather than at review. - **Correct inline.** If the person says "no, the other Popescu", the system fixes it and moves on, without restarting. - **End with the whole picture.** The last step is always the prefilled form, visible, before anything is recorded. #### External intake The same loop works for people outside the company: a link sent to a new supplier, a new hire, a customer with a request. "Drop your documents here and answer two questions" replaces the forty-field page, and the abandonment rate is the first number to move. Three things matter more outside than inside. **Provenance is shown to the submitter too**, so they see what the system read from their document and can fix it themselves. **Purpose and retention are stated plainly**: what the documents are for, how long they are kept. And **a plain form remains available as a fallback**, because some people, devices, and situations want one, and forcing the conversation on them is the old mistake in new clothes. #### Which forms to keep Some forms should stay forms, and knowing which is part of the discipline. **Tax and regulatory declarations.** The structured view is the safety. The fields have exact legal semantics, the person must see exactly what is being declared, and the form is the audit artifact. Intake can prefill it; the form remains the surface. **Attestations and signatures.** "I declare that the above is true" is not data entry. It is an act, and the form is how the act is recorded. **High-volume entry by experts.** A person who keys two hundred order lines a day from memory is faster on a keyboard than in any conversation. Give them prefill and shortcuts, not a chat. The rule that follows: replace intake, keep attestation. If the form's purpose is to collect information a document already contains, replace it. If its purpose is for a person to commit to something, keep it and make it shorter. #### Measuring intake The numbers from [part 3](https://agentique.eu/blog/munca-dintre-sisteme-3) apply with small renames. **Completion rate**, against the old form's abandonment. **Time to complete**, submitter side and reviewer side. **Fields corrected at review**, per field, which is the extraction precision in disguise and tells you exactly which field to fix. **Follow-up contacts per submission**, the calls and emails that used to happen after the form. And **escapes**: wrong values that reached the record, counted and traced, as always. #### Rollout Start inside, with one intake where the documents already exist and mistakes are cheap. Supplier onboarding is close to ideal: the registration certificate and the first invoice contain almost everything, the person doing it today is easy to find, and a wrong field is caught at the first payment. Run it in shadow: the old form stays, the prefilled version sits next to it, and you compare for two weeks. Then assisted, then the external link. Expense claims and service requests follow the same path. Two to four weeks per intake with a willing owner, the same pace as a seam. The interface layer has now been described from both sides: between systems, in parts 1 to 3, and between people and systems, here. What remains is the practical question every Romanian company asks at this point: with the state's e-invoicing, reporting, and transport mandates already in place, where does one begin? That is part 5. --- ### The Interface Layer, part 5: how a Romanian company starts (2026-09-02) URL: https://agentique.eu/blog/munca-dintre-sisteme-5 (Romanian) · https://agentique.eu/en/blog/munca-dintre-sisteme-5 (English) Four chapters described the interface layer: [finding the job](https://agentique.eu/blog/munca-dintre-sisteme-1), [handing a seam over](https://agentique.eu/blog/munca-dintre-sisteme-2), [measuring it](https://agentique.eu/blog/munca-dintre-sisteme-3), and [replacing forms](https://agentique.eu/blog/munca-dintre-sisteme-4). This last chapter is about the practical question a Romanian company asks at the end of all that: with everything the state has required of us in the last few years, where do we begin? It turns out the state answered part of the question without meaning to. #### The Romanian starting position In a short span, Romanian companies were required to send and receive invoices through the national e-invoicing system, to prepare a standardized ledger export for the tax authority, and to declare certain transport movements before they happen. Each mandate landed as cost: downloads and uploads in SPV, declarations typed into portals, a monthly dossier assembled for the accountant, and the seams that all of that added to already-stretched back offices. Look at the same facts from the other side. Every invoice a company receives now exists as structured XML with a schema. The ledger has a standard machine-readable shape. Transport data is captured in a defined format. For the first time, the raw material of the interface layer, the documents that used to be paper, photos, and PDFs with a hundred layouts, is uniform and machine-readable by law. Companies elsewhere in Europe are heading toward the same position as e-invoicing mandates spread; Romania simply arrived early and abruptly. The pain was real. So is the position. #### The seam most companies forget to count In a large share of Romanian companies the accountant is external, and the accountant is a system. Information crosses that seam every month: invoices photographed and sent on a messaging app, the dossier of originals, questions about what a payment was for, the return trip of entries and declarations. Chapter 1's inventory usually shows this as the widest seam in the company, and it rarely appears on anyone's list because it looks like a relationship rather than a data flow. Count it. The accountant will thank you: they are crossing the same seam from the other side. #### Three shapes, three first seams **The small company with an external accountant.** The first seam is the one to the accountant. Intake replaces the photos: incoming e-invoices arrive from SPV on their own, receipts and contracts are dropped in as they come, and the monthly dossier assembles itself, complete, categorized, with sources. The accountant receives structure instead of a folder of images, and the questions in both directions drop. Nothing is filed with the authority by the layer; the accountant's declarations remain the accountant's acts. **The mid-size company with finance in-house.** The first seam is supplier invoices into the ledger, exactly as chapter 2 laid out, because the input is now structured and the judgment repeats. The second is client status questions answered from operational systems. Transport declarations come third, once the layer has memory of routes, partners, and the rules that govern them. **The accounting firm.** The seam is SPV multiplied by the number of clients, and then the client questions. A firm that pulls every client's e-invoices automatically, categorizes them, and prepares each dossier has removed a day of logistics per month per accountant and gained something more valuable: the ability to answer "what did this client buy from that supplier this year" in a sentence. The accountant's expertise, tax, gets the hours back. #### The ninety-day plan **Days 1 to 10: the two inventories.** The seam inventory from chapter 1 and the forms inventory from chapter 4, made with the people who do the work. Deliverable: two pages, ranked, with the judgment column filled in. **Days 10 to 30: first seam, design, shadow.** Pick by chapter 2's five criteria. Write the design document: the seam on the loop, a tier per step, the confirmation rule. Start shadow mode: the system proposes, the person keeps working as before, you compare daily. Deliverable: the baseline and a first per-class precision number. **Days 30 to 60: assisted, memory grows, first numbers.** The person works from proposals and corrects. Every correction becomes memory. The weekly one-page review from chapter 3 starts. Deliverable: approval rate per class, the unresolved queue and its age, hours returned against the baseline. **Days 60 to 90: earned autonomy, second seam.** Classes with a clean approval record and an undoable failure mode move to sampled autonomy. The second seam enters shadow. Deliverable: a report an owner can read in five minutes: hours returned, escape rate, cost per correct decision, and what comes next. Ninety days is realistic with one willing owner and a company that keeps the review cadence. It stretches when the review is skipped, not when the technology is slow. #### What to have in place before day one - **An owner per seam.** The person who does it today, willing, with the time to review daily during shadow. - **Access to the documents.** SPV authorization for the company's certificate holder, a folder or inbox where the rest already lands. - **The accountant in the loop.** Shown the audit trail early, because their trust in what the layer produces decides whether the monthly dossier is accepted or re-done. - **The confirmation rule, written.** What waits for a person: anything leaving the company, anything filed, anything moving money. One paragraph, agreed before anything runs. - **A decision on where data lives.** For most companies, hosting inside the European Union with a processing agreement in place; for some, stricter, and that decision is easier made before the first document is uploaded than after. #### Data, law, trust The interface layer processes personal data: names on invoices, employees in onboarding, clients in contracts. The basics apply and are not optional: a data-processing agreement with whoever operates the layer, stated retention periods, EU hosting or a documented reason for anything else, and access limited to the people who need it. Nothing here is new; the layer must simply meet the standard the company already owes. Two things the layer must never do. It must never file anything with an authority on its own; declarations are human acts, as chapter 4 argued, and the layer's job ends at a prefilled, sourced draft. And it must never replace the originals: legal retention rules for accounting documents continue to apply, and the layer keeps originals and links to them, it does not stand in for them. #### Pitfalls that are specifically Romanian **The tax id in several spellings.** With and without the country prefix, with and without spaces, occasionally with a typo carried for years in someone's address book. Normalize at tier zero and match on the normalized form, never on the string as it arrived. **Diacritics.** The same company name with and without them, from different sources. A resolution step that treats them as different names will create duplicates for months. **The messaging group as a system.** If invoices, delivery confirmations, or client instructions arrive on a messaging app, that app is a system in the inventory, with its own seam. Pretending it is not is how the layer misses a third of the traffic. **Closed software with no API.** Some local accounting and ERP programs expose nothing. The seam then runs through their human-readable exports and imports, which is slower and still far better than a person retyping. The layer's job is to prepare exactly the import the program accepts. **The certificate and the token.** Access to the national e-invoicing system is authorized with a qualified certificate and lives on tokens that expire. Whoever operates the layer must refresh them ahead of time and tell you when it cannot. Ask about this before day one; it is the single most common reason an automatic pull stops quietly. #### Where we fit, and where we do not We build one implementation of the layer, and the series was written so that a company could apply it with any implementation. Ours today reads documents, scans, and e-invoices pulled from SPV once the company's account is connected, a step we do together at onboarding. It fills and updates data tables from what it reads, reads and writes Google Sheets and Drive, calls systems that expose an API from scheduled workflows, answers questions with the source document and page attached, and stops at the confirmation boundary before anything leaves. What we do not have: ready-made connectors for specific Romanian ERP and accounting programs, and the broader write-back that comes with them. We are choosing those by what pilot companies actually run. If you make the two inventories from chapter 1 and chapter 4, they are the first thing we would like to read. #### The person who did this work The series began with one person: the office manager, the dispatcher, the accountant, whoever ended up carrying information between systems that never learned to talk. Five chapters later the argument is simple to state. The judgment they applied at every seam can now be taken by a layer that reads, resolves, proposes, and writes back, provided the company measures it and keeps a person at the boundary. Nobody is replaced: the person is freed from the mechanical half of the work and kept exactly where their judgment matters, at approval, at choosing, at owning the outcome. The hours that return are the hours nobody hired them for. The inventory is the first page. Everything else follows from it. --- ### The work between the systems (2026-09-02) URL: https://agentique.eu/blog/munca-dintre-sisteme (Romanian) · https://agentique.eu/en/blog/munca-dintre-sisteme (English) Every company has one. The person who copies order lines from the ERP into a supplier's portal. Who types invoice totals from a PDF into the accounting system, exports them to Excel for the Monday report, then pastes the report into an email. Nobody designed that job. It grew in the gaps between systems that never learned to talk to each other, and it is some of the most expensive work in the building, because it is done by someone who knows the business. #### Why integrations never finished the job Point-to-point integrations have existed for decades, and they never removed this person, for a simple reason: most of the glue is not data transfer, it is judgment. Is this the same supplier as last month, spelled differently? Which cost center does this line belong to? Is this invoice the one we were waiting for? Code could not answer those, so the integration took the easy half and the person kept the hard half. Every new system added another seam, and another slice of somebody's week. #### What changed An AI that reads a document the way a person does, remembers what it was told, and can act on other systems is exactly the missing piece: it takes the judgment half. We think this is the shift that matters most for the next years of AI in business software. A layer that sits between the systems you already have and does the walking between them, with a human approving what matters, rather than yet another system to migrate to. If you want the full argument, with the history and the failure modes, [the long version is here](https://agentique.eu/blog/munca-dintre-sisteme-teza). #### What that looks like in Agentique today **Invoices arrive on their own.** Connect SPV once (a step we activate together at pilot onboarding) and incoming e-Factura invoices are pulled in automatically, read, filed, and ready to answer questions. The person who used to download and file them gets that hour back. [How it works](https://agentique.eu/blog/arhiva-e-factura). **Documents become records.** Data tables the assistant fills and updates from what it reads: suppliers, contracts, deadlines. The re-keying step disappears, and the table stays queryable in plain language. **Sheets and Drive, both directions.** It reads from and writes to the Google Sheets your team already lives in, so a scheduled report can land where people actually look, instead of in one more inbox. **Systems with an API.** When one of your systems exposes an API, a scheduled workflow can call it, bring the data back, and cross-check it against your documents. Nothing to buy, just a description of what to fetch and when. **Your own tools.** For developers, the same workspace is reachable from any MCP client with one endpoint and an API key. [The developer guide](https://agentique.eu/blog/mcp-server). And the rule that makes any of this safe to hand over: anything that leaves the workspace or changes a record stops and asks first. Nobody is replaced: the person who did this work is freed from its mechanical half and kept exactly where their judgment matters, at approval. #### What we do not have yet Ready-made connectors for specific ERPs. Today a system needs an API and a workflow that describes the exchange; named one-click connectors are on the roadmap, and we are choosing them by what pilot customers actually run. If you want yours first, say so. --- ### The interface layer: a thesis on AI between systems (2026-09-02) URL: https://agentique.eu/blog/munca-dintre-sisteme-teza (Romanian) · https://agentique.eu/en/blog/munca-dintre-sisteme-teza (English) This is the long version of [an argument we made in short form](https://agentique.eu/blog/munca-dintre-sisteme). Rather than a product tour, it is an attempt to state precisely what changed, why the previous fifty years of integration technology did not change it, what can go wrong, and what discipline makes it work. If you run a company with more than two systems, or you build software for such companies, it is written for you. The argument is deliberately short; the practical companion, chapter by chapter, [starts here](https://agentique.eu/blog/munca-dintre-sisteme-1). #### 1. The job nobody designed Take any company with an ERP, an inbox, a spreadsheet, a bank portal, and the state's e-invoicing system, and you will find work whose only purpose is moving information from one of those to another. Call it the interface layer. It has five recurring shapes: - **Transcription.** Reading a value in one place and typing it into another: invoice totals from a PDF into the ledger, order lines into a supplier portal. - **Reconciliation.** Deciding that two records are the same thing: the supplier spelled three ways, the payment that matches this invoice, the delivery that closes that order. - **Translation.** Converting meaning between systems that model the world differently: their "customer" is our "account", their "net 30" is our payment-term code 4. - **Routing.** Deciding where something goes and who must see it: which cost center, which approver, which folder. - **Exception handling.** Everything that does not fit: the credit note without an invoice, the scan that is half legible, the duplicate that is not quite a duplicate. Three properties make this layer expensive. It is performed by people who know the business, because the judgment calls require that knowledge. It is invisible: no org chart has a box called "interface layer", so its cost is never summed. And it grows faster than the company: with n systems there are up to n(n-1)/2 seams, and every seam is a place where someone reads from one side and writes to the other. #### 2. Fifty years of trying Every generation of enterprise technology has attacked this layer. The pattern of what each achieved, and where each stopped, is the whole argument. **Electronic data interchange** gave large trading partners structured messages for orders and invoices. It worked where both sides invested in rigid formats and stable relationships. It did nothing for the long tail of smaller partners, unusual documents, or anything not anticipated by the standard. **Middleware and service buses** addressed the n-squared problem with a hub: every system talks to the bus, the bus talks to every system. Mappings became a specialist's job, expensive to build and brittle when either side changed. The hub moved the seams; it did not remove the judgment inside them. **APIs and integration platforms** made the plumbing cheaper and the connectors more numerous. Deterministic mappings became cheap. But a mapping is still a rule, and rules only cover the cases their author imagined. The exceptions kept flowing to a person. **Robotic process automation** took the most honest approach: automate the human's clicks exactly as they happen. It also revealed the problem most clearly. A robot that replays a workaround encodes the workaround. When the screen changes, it breaks; when the case is unusual, it stops and raises a ticket, which lands on the same person. Look at what these waves have in common. Each automated the deterministic half of the work, the part expressible as a rule, and stopped at exactly the same wall: the point where the next step depends on reading something ambiguous, remembering a past decision, or judging plausibility. Half a century of tooling, one wall. #### 3. Why judgment was the wall Four things made the judgment half resistant to code. **Open-ended input.** The inputs to the interface layer are documents, emails, scans, and screens, produced by other organizations for humans. Their variety is unbounded. A rule that parses this month's invoice layout fails on next month's. **Drift.** Even structured sources change: a field is renamed, a partner switches formats, a new tax code appears. Rules decay silently, and the decay surfaces as exceptions. **Context.** "Is this the same supplier?" cannot be answered from the record alone. It requires knowing that the entity was renamed last year, that this is the branch office, that the finance team decided in March to treat the two as one. That knowledge lived in a person's head. **The long tail.** The exceptional cases are individually rare and collectively dominant. Automating the common path removes the pleasant part of the job and concentrates the person's day on the hard part. The headcount does not move; the job gets worse. #### 4. What changed, precisely Four capabilities arrived together in large language models, and it is the combination that matters. **Reading at human level.** A model can read a scanned invoice, an email thread, or a portal page and extract what a person would extract, without a layout-specific rule. This covers transcription and translation, the two shapes that generated the most volume. **Resolution in context.** Given a candidate and the records it might match, a model can weigh spelling, address, tax id, and history and decide, or say it cannot decide. This is reconciliation, the shape that generated the most judgment. **Acting.** Models can call tools: query a system, write a record, send a message, run a calculation. Reading without acting would only have produced better reports. Acting closes the loop. **Memory.** A decision stated once ("treat these two entities as one", "invoices from this supplier go to that cost center") can be stored and applied consistently afterward. This is what turns a clever reader into a colleague who learns the house rules. Now the limits, stated with the same precision. These systems are probabilistic: they can be confidently wrong, and they do not know when. They have no authority of their own: a model that can write records is a model that can write wrong records at machine speed. They cost money per step, so wasteful loops are real. And they are only as good as what they can see: an assistant that cannot read your documents is a chat window. None of these limits is fatal. All of them dictate the architecture. #### 5. The loop that makes it safe The interface layer, done right, is a loop with a specific shape. Each stage has a discipline attached, and the disciplines are what separate a working system from a demo. **Provenance on every fact.** Everything extracted carries its source: which document, which page, which field. The cost of verifying a claim drops from "find the original" to "one click". Far from a nicety, this is what makes the human approval step fast enough to be real rather than a rubber stamp. **A hard confirmation boundary.** Reading, resolving, and drafting are free to run. Anything that leaves the workspace or mutates a record of consequence waits for a person: sending, deleting, changing money-bearing fields. The boundary has to live in code, not in a prompt, because prompts are suggestions and boundaries must be guarantees. **Idempotency.** Every synchronization is keyed by external identifiers, so running it twice produces nothing new. Systems that pull from other systems will be re-run, by schedules and by impatient people, and duplicates are the classic way an automation destroys trust. **Least privilege, assuming compromise.** Credentials are scoped to what the job needs, rate-limited, and stored as if they will leak one day, because some of them will. A key that can only read invoices cannot email your clients. **Fail toward the human.** When the evidence is insufficient, the correct output is "I could not determine this, here is what I found", never a plausible guess. A system that improvises on the hard cases is worse than one that never existed, because its errors arrive with the confidence of a system. **An audit trail.** Every action recorded with who, what, when, and on what evidence. Less for compliance than for the Tuesday when something is wrong and someone must find out why in ten minutes. #### 6. The economics, honestly Where does the time actually go in the interface layer? Rarely in the happy path. It goes into exceptions, lookups, and the switching cost of holding three systems in one head. The value of automating the layer is therefore the fraction of judgment calls the system resolves correctly, multiplied by the cost of the human minute it saves, minus the cost of the ones it gets wrong, and only marginally "the robot does the typing". That formula has a consequence for how to evaluate any vendor, including us. The numbers that matter are measurable: - **Resolution precision.** When the system says two records match, how often is it actually right, judged by the final outcome? - **Approval rate.** How often does the human accept the proposal unchanged? A high rate means the judgment is being taken; a low rate means the human is still doing the job with extra steps. - **Escape rate.** How many errors reach the ledger, the client, or the tax authority? - **Time to close.** Per item, from arrival to recorded, with the human in the loop. A vendor who cannot report these on your data is selling a feeling. The measurement is the product. ##### The cost objection The first objection to all of this is cost: models are paid per token, and the interface layer runs all day. Three things make the objection weaker than it sounds. The comparison is against a human minute, and the gap is enormous. A resolution call that reads an invoice and a handful of candidate records costs a fraction of a cent; the minute of a person who knows the business costs more by orders of magnitude. The price of tokens has also fallen steeply year over year and shows no sign of stopping, while the price of the human minute rises. Most of the loop does not need the strongest model. Classifying a document, routing a trigger, extracting fields from clean XML, checking a plausibility rule: these are small-model tasks, fast and cheap, and they are the bulk of the volume. The strongest models are for the hard tail: ambiguous reconciliation, multi-document reasoning, a half-legible scan. A tiered design, small model first and strong model on the exceptions, spends the expensive tokens exactly where the human minutes used to go. Some of it needs no model at all. A structured e-invoice is an XML with a schema: the supplier's tax id, the total, the due date, the VAT lines are read by a parser, exactly and for free, and a model that re-reads them is wasted spend. Keyword search is tokenization plus scoring and finds an invoice number or a tax id better than any embedding. A fixed set of commands ("pull the invoices now", "what is due this week") is a small classifier and a dispatch table. Scheduled triggers are a clock. The discipline is to push every task down to the cheapest tier that meets the precision bar: parser and rules where the input is structured, small model where it is semi-structured or the choice is among a closed set, strong model only where reading is open-ended, the match is ambiguous, or the next step must be planned. Most of the volume lives in the bottom two tiers. Most of the value, and nearly all of the risk, lives in the top one. The real cost risk is waste, not price. An agent that loops, re-reads whole documents it could have retrieved from, or re-sends the same instructions on every call burns money without producing decisions. The disciplines are known: retrieval instead of whole documents, caching of the parts of a prompt that repeat, a budget per run, cooldowns on triggers, batching of scheduled work. Token economy is an engineering practice, and it is where most of the cost difference between implementations lives. On running models on your own hardware: a mid-sized open model on a decent server can carry the small-model tier well, and for organizations with strict data-residency requirements that is a legitimate design. Two honest caveats. The hard tail is exactly where smaller models are weakest and where errors cost the most, so a local-only design needs the confirmation boundary more, not less. And a model server is an operations commitment: updates, evaluation, capacity. For most companies, a tiered design hosted inside the EU delivers the same economics without a closet full of GPUs. The right metric for any of these choices is cost per correct decision, never cost per token. #### 7. What this means for ERPs The ERP is the system of record: the ledger, the stock, the contracts, the truth. That role does not go away, and any argument that AI replaces the ERP misunderstands what an ERP is for. What changes is the system of work: the surface people touch during the day. That surface is moving out of the ERP's forms and into a layer that reads what arrives, resolves it against the records, proposes the entry, and writes it back once approved. Three predictions follow. **Connectors become a commodity.** As more systems expose APIs and standard protocols emerge for tools to talk to models, the question "do you have a connector for X" loses weight. The plumbing was never the hard part; the previous four waves proved it. **Judgment becomes the product.** The differentiator becomes which decisions the layer can take reliably, with what evidence, under what boundary. This is where the disciplines of section 5 stop being engineering hygiene and become the competitive substance. **Bolted-on chat is not an interface layer.** Expect every ERP to grow a chat window. A chat window without provenance, without memory, without tools, and without a confirmation boundary is a search box with better grammar. The test is simple: ask it to do the reconciliation, then ask it how it knows. #### 8. What stays human Three things, and they are not consolation prizes. **Choosing.** Between two suppliers, two offers, two ways to structure a deal. The system can lay out the comparison with every number; the choice carries responsibility, and responsibility does not transfer to software. **Committing.** A promise to a client, a signature, a payment released. These are acts of the company, made by people the company has authorized. **Owning the error.** When the interface layer gets something wrong, a person decides what to do about it, and that person needs the audit trail from section 5 to do it in ten minutes rather than a day. Nobody is replaced. The person who did this work sees their day change from doing the transfers to approving them, and approval is precisely the part that always needed a human. The hours saved are the hours that never needed one. #### 9. Where we are We build one implementation of this thesis, and we would rather be judged on the specifics than on the argument. Today the layer reads documents, scans, and Romanian e-invoices pulled from the state's SPV system once a company connects its account; it fills and updates data tables from what it reads; it reads and writes Google Sheets and Drive; it calls systems that expose an API from scheduled workflows; it exposes the workspace to any tool that speaks the Model Context Protocol; and every consequential action stops at the confirmation boundary. What we do not have yet: named one-click connectors for specific ERPs, and the broader write-back that comes with them. Pilots are how we learn which judgments are safe to take next, and in what order. If the argument above describes a job in your company, we would like to see the job. --- ### What your e-Factura archive knows (and how to ask it) (2026-08-20) URL: https://agentique.eu/blog/arhiva-e-factura (Romanian) · https://agentique.eu/en/blog/arhiva-e-factura (English) Every month the invoice folder from SPV grows a little more. Files with cryptic names that nobody opens, because XML was never meant for human eyes. And yet your company's entire purchasing life is in there: who invoiced you, how much, when, with what due date and what VAT. e-Factura (Romania's mandatory e-invoicing system) had a side effect few companies actually use. For the first time, every incoming invoice exists in a format a machine can read perfectly. Clean, structured data, not a crooked scan of a paper. The only problem: you don't read XML. An AI assistant does. There are two roads from SPV to answers: the automatic one and the manual one. #### The automatic way: connect SPV once In Settings, a workspace admin connects the company's SPV account: a one-time authorization in the browser, with the company's qualified certificate. In the pilot we do this step together, at onboarding. From then on, incoming e-Factura invoices are pulled in automatically several times a day: downloaded, read, recognized as invoices, filed, key data extracted. Nothing to download, nothing to drag. The assistant knows about the connection too. "When did SPV last sync?" gets a real answer, and "pull the invoices now" starts a sync on the spot instead of waiting for the next scheduled one. #### The manual way: the archive, by hand No connection yet, or an older archive to bring in? Download it from SPV and drag it into Agentique: a whole zip works, up to 100 files at a time, so a month usually loads in one move. From there on the two roads merge: every XML is read, recognized as an invoice, filed, and its key data extracted. Then you ask. In your own words, not in accounting-speak: - "How much did we buy from supplier X in July, in total?" - "Which invoices are due next week?" - "How much VAT was on last month's incoming invoices?" Every answer comes with a link to the source invoice, so any number can be checked in one click. And totals are computed, not guessed: when the assistant adds up amounts, it uses its calculation tool, the same way you would reach for a calculator. #### More than questions If you want the picture always current instead of on demand: a supplier table the assistant fills and updates as you work, a scheduled weekly report with new invoices and upcoming due dates, and memory for your rules, stated once ("invoices from X go to cost center Y") and kept. #### The limits, stated plainly The SPV connection is activated together with us, during the pilot: the one-time ANAF authorization needs the company's qualified certificate, and we walk through it with you at onboarding. Manual zips take up to 100 files each; split a bigger month in two. And, as everywhere in Agentique, nothing leaves your workspace without your confirmation. --- ### The MCP server: your workspace, from your own tools (2026-08-06) URL: https://agentique.eu/blog/mcp-server (Romanian) · https://agentique.eu/en/blog/mcp-server (English) Your company's documents, memory and data tables live in the Agentique workspace. You live in a terminal, an editor, or an agent you are building yourself. The MCP server connects the two: one HTTP endpoint that exposes the workspace's tools to any client speaking the [Model Context Protocol](https://modelcontextprotocol.io). MCP is an open standard, and Agentique implements it as a plain JSON-RPC 2.0 endpoint over the Streamable HTTP transport. There is no SDK to install and no client we favor: a desktop AI app, an IDE plugin, a cron script, an internal service, anything that can send an HTTP POST with one header can use the workspace. The server speaks protocol versions 2024-11-05 through 2025-06-18, so both older and current clients negotiate cleanly. #### What the endpoint exposes Twenty-two tools, curated on purpose. Search and retrieval (`intelligent_search`, `keyword_search`, `list_documents`, `get_document`), workspace memory (`search_memory`, `save_memory`), data tables read-only (`list_entity_types`, `search_entity_records`), web access read-only (`web_search`, `fetch_webpage`), pure calculation (`calculate`, `convert_units`), Google Sheets and Drive when your workspace has Google connected, SPV e-Factura status and sync (`spv_status`, `spv_trigger_sync`), and findings-report generation (`create_findings_report`). Just as deliberate is what is missing. Nothing destructive is reachable with an API key: no deleting, no sending email, no arbitrary outbound requests (the web tools are search and read-only fetch). Actions like those stay in the workspace UI, where a human confirms each one. A leaked key must never be able to email your clients or empty a folder, so the surface is built as if every key will leak one day. That is the default posture, not a hard ceiling. If your integration genuinely needs one of the excluded actions, say so during the pilot: we would rather extend the surface deliberately, with the same guardrails, than ship dangerous defaults to everyone. #### Connect in two steps Everything you need sits in Settings, in the MCP server panel: the exact endpoint URL of your workspace, and your API keys. A workspace admin creates a key with a name and scopes, and the plaintext (an `ak_` token) is shown exactly once, at creation. Only a hash is stored, so copy it then. Keys can be revoked at any moment, can carry an expiry date, and show when they were last used. First, prove the connection with curl: ```bash curl -s https://your-workspace-endpoint/api/v1/mcp/ \ -H "Content-Type: application/json" \ -H "X-API-Key: ak_your_key" \ -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}' ``` You get back all twenty-two tools with their JSON schemas. A plain GET on the same URL returns server info, which makes health checks trivial. Then point your client at it. Clients that speak Streamable HTTP take a config in this shape. Claude Code reads it from `.mcp.json` at the project root, and Cursor takes the same `url` plus `headers` pair in `.cursor/mcp.json`: ```json { "mcpServers": { "agentique": { "type": "http", "url": "https://your-workspace-endpoint/api/v1/mcp/", "headers": { "X-API-Key": "ak_your_key" } } } } ``` Some clients only speak stdio. For those there is a bridge: a single-file Python script, standard library only, Python 3.10+, downloadable from the same Settings panel. It reads JSON-RPC from stdin and forwards it to your endpoint over HTTPS. Claude Desktop is the usual example: ```json { "mcpServers": { "agentique": { "command": "python3", "args": ["/absolute/path/to/agentique-mcp-bridge.py"], "env": { "AGENTIQUE_URL": "https://your-workspace-endpoint/api/v1/mcp/", "AGENTIQUE_API_KEY": "ak_your_key" } } } } ``` #### Scopes and limits, in the open Every key belongs to one workspace and can only ever see that workspace's data. Scopes keep a key as narrow as its job: `read` covers search, documents, memory reads and table queries; `write` adds memory saves, Sheets/Drive writes and the SPV sync trigger; the calculation tools need no scope at all; `admin` unlocks everything and is rarely what you want. Rate limits are 60 requests per minute per key, and 15 per minute for the expensive tools (semantic search, the web tools, Drive import, report generation). Responses are synchronous JSON, no SSE: a request returns when the tool returns. #### Three things worth building **Ask the workspace without leaving the editor.** "What payment terms did we agree with this supplier?" answered inside the IDE, source document and page in the response, while you write the code that depends on the answer. The same answers the workspace gives, in the tools where you already are; [how the answering works is documented here](https://agentique.eu/blog/how-agentique-works). **A reconciliation script.** A nightly job pulls records from a data table, compares them with your internal system, and writes the differences into a Google Sheet the team already watches. Read scope for the query, write scope for the sheet, twenty lines of code. **Your own agent, with real context.** Building an agent in any framework that supports MCP? One config block hands it your company's documents and memory as tools, with every guardrail above already enforced server-side, not promised client-side. --- ### What a back office can automate today, and what it should not (2026-08-04) URL: https://agentique.eu/blog/back-office-what-to-automate (Romanian) · https://agentique.eu/en/blog/back-office-what-to-automate (English) A back office runs on a few repeated motions. Find the document. Check the number. Copy it somewhere else. Write the summary. Chase the deadline. Some of that work is now genuinely automatable. Some of it should never be, no matter what any AI company tells you. The difference matters more than any feature list, so here is the map we actually believe in. #### The work you can hand over The test is simple: if describing the task sounds like "read, find, count, summarize, repeat", a machine can do it, and probably better than a tired human at 17:40 on a Thursday. **Answering questions from your own documents.** "What payment terms did we agree with this supplier?" A person answers by remembering which folder to dig through. An AI answers by having read everything, and a good one shows you the source document and the page, so trust never depends on memory. **Sorting what comes in.** The email attachments, the scans, the folder from an old laptop. Reading each file, figuring out what it is, filing it, writing a two-line summary: this is exactly the work nobody has time for, and exactly the work a machine never gets bored of. **The recurring report.** If a report has the same shape every week, a schedule can produce it and a human can read it. Producing it by hand was never the valuable part. **Arithmetic across documents.** Totals that live in four contracts, rates scattered across invoices. Collecting and adding numbers is mechanical work. Checking the result, one glance, stays pleasant. **Remembering what changed.** A supplier changes their bank account, a discount is renegotiated. Facts like these should be stated once and then simply hold, in every future answer. #### Why you can actually let go Handing work over only makes sense if you can trust what comes back, and here the trust is not asked for on faith. What it takes from your documents, it shows: a link to the source, with the page for anything that has pages, so checking takes one click. The way you would trust a careful new colleague: on receipts, not on charm. When the documents do not contain the answer, it says so instead of improvising. You keep full control of the result. You just stop paying the afternoon it used to cost. #### The work that stays yours The other test: if the task sounds like "decide, approve, promise, sign", it stays human. Not because the technology fails there, but because responsibility does not transfer. **Decisions.** An AI can put two supplier offers side by side with every number that matters. Choosing one is your job, and should be. **Anything that leaves the building.** An email to a client, a deleted document, a changed record. Agentique prepares these but stops and asks before any of them happens. Every time, by design. An assistant that acts on your behalf without asking is not an assistant, it is a liability. **Judgment under ambiguity.** Whether to enforce a penalty clause on a good client, whether a deadline is worth a discount. The AI brings the clause, the history and the numbers. The wisdom is yours. **Questions your documents cannot answer.** The most underrated feature in this whole category is a system that says "this is not in your documents" and asks how you want to proceed, instead of producing something confident and wrong. If a tool cannot say "I do not know", it is not ready to be near your business. #### The rule of thumb Read, find, count, summarize, repeat: hand it over, today. Decide, approve, promise, sign: keep it, and let the machine bring you everything you need to do it in minutes instead of afternoons. Automation done this way does not replace your people. It gives your most capable person back the hours they currently spend being a search engine. --- ### Five jobs you can hand to Agentique this week (2026-07-16) URL: https://agentique.eu/blog/five-jobs-to-hand-over (Romanian) · https://agentique.eu/en/blog/five-jobs-to-hand-over (English) Every company that runs on documents has the same five time sinks. Someone digs through folders for one clause. Someone retypes numbers from invoices into a spreadsheet. Someone writes the same Monday summary every single week. Here is what happens when you hand those jobs to Agentique instead. #### 1. Finding the one thing you need "What is the late payment penalty in the supplier contract?" Ask exactly like that, in Romanian or English, and you get the answer with a link to the document and the page it came from, so checking takes one click instead of ten minutes of scrolling a scan. Two details matter here. You can type without diacritics, "plati intarziate" finds documents that say "plăți întârziate". And the answer only uses what your documents actually say. If the information is not there, it tells you so and asks how you want to proceed, instead of inventing something plausible. How well does the finding part work? We test it on a fixed set of 91 questions written the way real users ask, in both languages, with typos and missing diacritics included. 90 of the 91 land on the right document. ![A question answered with a structured reply and links to the source documents](https://agentique.eu/landing/01-chat-preview.png) #### 2. The pile nobody has time to sort That folder of 60 unsorted PDFs from an old laptop, or the archive a client sends you as a single zip. Drop the whole thing in. Agentique accepts zip archives with up to 100 files inside, reads each supported file, writes a short summary of every one, and files them into categories on its own. Files it cannot use, an installer, a video, are listed by name so you know exactly what was skipped. Scans and phone photos count as documents too. A photographed delivery note with a stamp on it is read page by page, like a person would read it. The practical limits, so there are no surprises: files up to 50MB each, archives up to 100 files. ![The Knowledge Base after a bulk import: documents categorized, summarized and indexed](https://agentique.eu/landing/02-knowledge-base.png) #### 3. The report you write every week Tell it once, in chat: "every weekday at 8:00, send me a summary of the documents that came in the day before". That becomes a scheduled automation you can see, edit or pause in the Automations page. Hourly, daily, weekdays, weekly or monthly. The Monday summary stops being someone's job. ![The Automations page with scheduled recurring tasks and their run history](https://agentique.eu/landing/05-automations.png) #### 4. The numbers scattered across documents "What is the total value of the leasing installments we pay this year?" That answer lives in four different contracts. Agentique reads them, does the arithmetic, and can put the result in a real file: a PDF report or an Excel table you can forward as is. If you keep structured records, it fills your data tables directly, and you check the result instead of typing it. #### 5. The things everyone forgets Tell it "the supplier changed their bank account in March" and every future answer about payments takes that into account. Corrections, agreed discounts, internal rules, they persist across conversations, and the Memory page shows everything it has remembered so you can edit or delete any of it. #### One more thing about trust Anything sensitive stops and asks first. Sending an email, deleting a document, changing a record: you see exactly what is about to happen and confirm it, every time. Your documents stay in your company's own space, are never used to train AI models, and never mix with anyone else's. #### Try it on your own documents The beta is open, with invite codes and a 30-day trial. Bring a folder of real documents, the messier the better, and ask the first question five minutes later. [Write to us](mailto:mihnea-rares.gilceava@itailors.eu?subject=Agentique%20beta%20invite) and we will set your team up. --- ### How Agentique works, explained simply (2026-07-10) URL: https://agentique.eu/blog/how-agentique-works (Romanian) · https://agentique.eu/en/blog/how-agentique-works (English) You do not need to understand how Agentique works to use it. But if you are curious what happens between upload and answer, here is the whole journey, in plain words. #### The journey of a document **You upload anything.** A contract as PDF, an Excel file, a Word document, an XML invoice, a scanned page full of stamps and signatures, or simply a photo of a paper taken with your phone at the warehouse. Drag it into the app or attach it in a conversation. **The AI reads every page.** Documents with normal text are read directly. Scans and photos are different: there is no text inside them, only a picture. For those, the AI looks at each page the way a person would and writes down what it sees, including tables, stamps and handwritten notes. You can watch the progress live, page by page. **It figures out what the document is.** An invoice, a contract, a delivery note, a report. Based on that, it files the document in the right folder and pulls out the information that matters for that kind of document. For a delivery note, for example: the note number, the date, the vehicle plate, the supplier and the quantities. **From that moment, the document answers questions.** Everything it contains can be found again, in your own words, even months later. ![The Knowledge Base with categorized documents, AI summaries and indexed status](https://agentique.eu/landing/02-knowledge-base.png) #### Ask in your own words There is no special way to ask. Write the way you would ask a colleague: "how much material did we receive from our suppliers in June?" or "which contracts expire this year?". Romanian or English, both work. The answer uses only what your documents actually say, and every claim comes with a link to the exact source, so you can always check with one click. ![A conversation with a sourced, structured answer and document links](https://agentique.eu/landing/01-chat-preview.png) #### From answers to deliverables Answering questions is the beginning. In the same conversation you can ask for real deliverables: It fills in your data tables, drafts and sends emails, translates documents, sets reminders and runs scheduled tasks like a weekly summary of new documents. For anything sensitive, deleting a document or sending an email to someone, it always stops and asks you to confirm first. And if an answer is going in the wrong direction, there is a stop button. #### It remembers what you tell it Tell it once that a supplier changed their bank account, and every future answer takes that into account. Corrections, preferences, rules of your business, they all persist across conversations. You can see and edit everything it has remembered, in the Memory page. ![The Memory page with saved facts, preferences and corrections](https://agentique.eu/landing/04-memory.png) If you work with several clients or projects, each project keeps its own space: conversations and documents from one project never mix with another. #### Your documents stay yours Every company on Agentique is completely separate. Your documents, your data tables and your AI's memory live in your own space. Your data is never used to train AI models and is never shared with anyone. Your documents are exactly that: yours. Downloads work through protected links that expire in minutes. When you delete a document, it disappears from answers too, immediately. #### Try it The beta is open, with invite codes and a 30-day trial. [Write to us](mailto:mihnea-rares.gilceava@itailors.eu?subject=Agentique%20beta%20invite) and we will set your team up. --- ### Agentique is in beta (2026-07-10) URL: https://agentique.eu/blog/agentique-beta (Romanian) · https://agentique.eu/en/blog/agentique-beta (English) Someone at your company spends hours in documents every week. Checking a contract clause before noon. Matching an invoice total against the delivery note. Retyping a document photographed at the gate. That work is real, and it swallows the hours of your best people. Agentique does the reading, so their hours go where judgment is needed. **Starting today, the beta is open.** Upload contracts, invoices, reports, scans, spreadsheets. Ask in Romanian or English, the way you would ask a colleague. You get answers with links to the exact source, tables filled in, reports generated and emails drafted, all in one conversation. ![A conversation with a sourced, structured answer and document links](https://agentique.eu/landing/01-chat-preview.png) #### What you will find inside - **Watch your documents being read.** Uploads show live progress, page by page on scanned files. - **Scans behave like real documents.** Dozens of pages, stamps, signatures: read and searchable. - **Change your mind mid-answer.** A stop button ends a response instantly, and you simply rephrase. - **Tidy up by asking.** Tell the assistant to create folders and file documents into them. It does it. - **One space per client.** Projects keep tasks and documents separated by engagement, so nothing mixes. #### How the beta works Access is by invite. Each workspace gets a 30-day trial with the full feature set, in English and Romanian. During the beta, the platform pauses nightly for maintenance between 23:30 and 09:30 (Bucharest time). Want an invite? [Write to us](mailto:mihnea-rares.gilceava@itailors.eu?subject=Agentique%20beta%20invite) and tell us a little about your team.