EU AI Act compliance where every sentence of the technical file is cited, every question is closed on evidence, and a named person signs
Every high-risk AI system documented, cited and signed before 2 December 2027.
See one case, screen by screen ↓EU AI Act Compliance builds the technical documentation, risk file and conformity evidence for high-risk AI systems. Seven agents classify each system against Article 5, Article 6 and Annex III, write each Annex IV section only from cited passages in your own documents, keep the Article 9 risk file, question the file as an authority would and prepare the declaration and registration record. Owners answer with documents and accept residual risk; the AI compliance lead signs.
How to choose AI solutions for back-office and enterprise work →
Updated
An AI system used on people — a model that shortlists job applicants, scores credit or allocates shifts — needs a technical file covering the nine points of Annex IV, an Article 9 risk management file, tests of accuracy for the groups it is used on, a human oversight design, a signed EU declaration of conformity and a registration in the EU database, all before it is used. The evidence exists, but it is scattered across the model registry, the code repository, the design wiki, the test library and the document library, and someone has to find it, check it is dated and signed, and write it up.
Today that writing is done by hand from model cards, slide decks and design notes, and claims creep in that no document supports. Then the model is retrained, and sections 2(d), 3 and 6 are wrong the week after signature because nobody connected the retrain to the file. An authority asks for accuracy by age band, or how an override is logged, and the answer is in a notebook nobody filed.
Estimated split for the first technical file of one in-house high-risk system, by hand and with the solution.
Seven agents classify, find evidence, draft each section from sources and question the file; owners answer, the compliance lead signs.
Kestrel & Vane Group builds its own candidate shortlisting model, AIS-0417 v3.2, used by recruiters in 30 countries — about 212,400 applications a year. As the provider, it needs a complete technical file before 2 December 2027. Here is what happens, screen by screen, in the working solution.
Of the 186 systems in the register, 23 need a decision or carry duties. Eleven are high-risk — 6 the group provides, 5 it deploys — with 414 days to 2 Dec 2027, files 44% filled on average and 14 assessor questions open. One prohibited system, emotion analytics at work, is blocked. Six items need her today, ranked by risk to the date.
Needs you today: “AIS-0417: 3 assessor questions stand between the file and sign-off.”
The agent runs seven tests, each with the rule it applies: an AI system under Art. 3(1); no prohibited practice under Art. 5; not a safety component under Annex I; recruitment under Annex III point 4(a). The Art. 6(3) derogation cannot apply because the model profiles applicants. Kestrel is provider and deployer, the duties apply from 2 Dec 2027, and the route is internal control under Annex VI. Confidence 0.97.
What would change this: “If it only turned CVs into application fields, with no ranking — not high-risk: Art. 6(3)(a), a narrow procedural task.”
The evidence finder searches six sources — the design wiki, document library, test library, model registry, code repository and data catalogue — and matches 23 documents to the nine Annex IV points. It flags one not reviewed in 12 months. The drafters write sections 1 to 7: 318 sentences, each cited to a passage, none without a source. Then the assessor reads the file and raises 9 questions, closing 6 on evidence the agents already hold.
“Nothing invented. Each sentence carries its document, version and passage; a person can check any of them in one click.”
The file opens as nine Annex IV tabs, each with its fill, its documents and anything missing in red. Helena clicks the first sentence of section 1 and the source lights up: the intended purpose statement IPS-0417-3, §1.1, from the design wiki, version 3.2, signed. Beside it sit the Article 9 risk file — 14 risks, 3 open — and the 17-point conformity checklist, 8 met.
“The system ranks each application against the vacancy's published requirements and presents a shortlist of no more than 25 applicants to the recruiter.”
Section 3 states accuracy of 0.91 overall but no figure for the groups of persons the system is used on. The drafters will not write it without a source, so it shows as missing. The assessor, reading as a market surveillance authority would, raises Q-0417-03 under Art. 15(1) and Annex IV 3 and sends it to Lucía Ferreira-Hahn, due Wed 21 Oct.
“Missing: Accuracy for the specific groups of persons the system is used on (age bands) … the drafters will not write this without a source.”
Lucía uploads TR-0417-19, the subgroup accuracy test report, 14 pages, signed on 13 Oct 2026. The agents read the accuracy tables for 4 age bands, 2 sexes and 30 countries, check the date and signature, and compare the bands with the 0.05 tolerance: 55 and over at 0.88 against 0.92 for 18–29, a largest gap of 0.04.
The question closes only on a document or a signed answer.
The drafters add six cited sentences: 0.92, 0.92, 0.90 and 0.88 by age band, 0.91 for women and men, and 0.89 in Poland to 0.93 in Ireland by country. Section 3 reaches 100%, the file 76%, and risk R-07, lower accuracy for applicants aged 55 and over, moves to mitigated. Sign-off stays locked: two questions are still open.
“Section 3 complete · Q-0417-03 closed · R-07 mitigated”
Tomasz answers Q-0417-08 with the override logging specification OLS-0417-1: every override records the recruiter, one of 6 reason codes and the time. For Q-0417-07 the drafter builds a post-market monitoring plan only from processes Kestrel already runs — a monthly sample of 2,000 shortlists, weekly drift, complaints — and he approves it. Then he accepts residual risk for R-07, R-09 and R-12 with reason code A1.
“Signing as Tomasz Weller … prepared by the risk file keeper, not by you.”
With every question closed, the declaration DoC-0417-01 unlocks. Helena signs with reason code D1, “Conforms: file complete, every question closed”, and the declaration preparer writes section 8 from the signed document. The EU database record — Annex VIII Section A, 11 fields — is prepared for her; she submits it herself and enters the number, EU-HR-2026-004127.
The declaration preparer never signs and never submits to the EU database.
On 6 Oct the model was retrained as v3.3. The change monitor checks it against the pre-determined change plan PCP-0417-02: architecture, features and data window unchanged, overall accuracy 0.91, but applicants aged 55 and over at 0.85, under the 0.87 floor. Promotion is held and v3.2 stays in service. If promoted, sections 2(d), 3 and 6 would re-open.
“Whether this is a substantial modification (Art. 43(4)) is for the system owner to decide.”
The same technical file, seen by the five people who carry it — what the work looked like, and what it looks like now.
Seven agents classify, find evidence, draft each section from sources and question the file; owners answer, the compliance lead signs.
Tests each system against Art. 3(1), Art. 5, Annex I, Annex III and Art. 6(3), finds the company's role and the date the duties apply, and explains every step with the rule and what would change it.
Searches the model registry, code repository, design wiki, document library, data catalogue and test library, matches each document to the Annex IV points it supports, and flags old or unsigned documents.
Write each Annex IV section only from cited passages, and list what is missing in red with the article and the person asked.
Keeps the Article 9 register: risks from tests, impact assessments and workshops, mitigations linked to evidence, residual risk proposed for the owner.
Reads the finished file as a market surveillance authority would, raises numbered questions with the article, routes each to the person who holds the evidence, and closes it only on a document.
Watches new model versions, drift and incidents, checks each retrain against the pre-determined change plan, re-opens only the sections affected and runs the 15-day serious-incident clock.
Assembles the Annex V declaration of conformity and the Annex VIII registration record from the file.
Signs the declaration. The agents propose; a named person decides.
Ask about any AI system, section, question or deadline. Every answer cites its source — and decisions go to the person who owns them.
What stops AIS-0417 from being signed?
3 questions and one acceptance. Q-0417-03, accuracy by age band (Lucía Ferreira-Hahn, due Wed 21 Oct); Q-0417-08, override logging (Tomasz Weller, due Fri 23 Oct); Q-0417-07, no post-market monitoring plan (Tomasz Weller, due Wed 28 Oct). Then Tomasz accepts residual risk for R-07, R-09 and R-12. The file is 73% filled, with 414 days to 2 Dec 2027.
Why is the shortlisting model high-risk?
It ranks job applications for recruiters, which is Annex III point 4(a): “to analyse and filter job applications, and to evaluate candidates”. The Art. 6(3) exemption cannot apply because it profiles applicants; a system that profiles people is always high-risk. Olivier Bassett confirmed it on 2 Jul 2026 (reason C1).
Which systems could miss their date?
One: AIS-0444, the promotion readiness recommender. Its file is 22% filled; at the pace of the last 8 weeks it finishes in Mar 2028, after 2 Dec 2027. AIS-0431 is close: 38%, forecast Aug 2027, with Q-0431-01 one day past due.
Is the v3.3 retrain a substantial modification?
That is Tomasz Weller's call, not mine. v3.3 passed 4 of 5 gates of change plan PCP-0417-02; it failed one: accuracy for applicants aged 55 and over is 0.85, under the 0.87 floor. Inside the plan it would not be substantial (Art. 43(4)); outside it, a new assessment is needed. Promotion is held and v3.2 stays in service.
13 screens from the working solution, on its sample data. Pick one to see it large.
How the agents work for your group — when they escalate, the rules they follow, and how fast questions must be answered. Changes are logged.
| Setting | Default | Choose from |
|---|---|---|
| Escalate a classification below | 0.80 confidence | A value between 0 and 1; a person always confirms |
| Largest accuracy gap between groups | 0.05 | Set per system |
| Accuracy floor for any group (AIS-0417) | 0.87 | Taken from the system's change plan, PCP-0417-02 |
| A document needs review after | 12 months | Months without review; the owner is asked to confirm |
| Drift review above | 0.20 | Population stability index, checked weekly |
| Answer an assessor question within | 5 working days | Reminder after 3, escalation to the AI Compliance Lead after 5 |
| No source, no sentence · no sign-off with an open question · no one signs what they prepared | On | Locked — cannot be switched off |
| Every retrain is checked against its change plan · prohibited practices block procurement | On | On or off |
Every figure is labelled: a target the solution is built to, an estimate, a typical published result, or a proven one.
“demo” = seen in the working solution, on its sample data · “target” = the design goal, measured in the live solution · “estimated” = our estimate · “typical” = published figures (Regulation (EU) 2024/1689, Art. 99(4) — up to €15M or 3% of worldwide turnover for breaching high-risk duties; consultancy estimates of €30–60k per system to write the first file, directional). Dates are as amended by Regulation (EU) 2026/1744 and named as context. People, companies, systems and documents named on this page are fictional — characters and sample data in the working solution.
EU AI Act Compliance builds the technical documentation, risk file and conformity evidence for high-risk AI systems. Seven agents classify each system against Article 5, Article 6 and Annex III, write each Annex IV section only from cited passages in your own documents, keep the Article 9 risk file, question the file as an authority would and prepare the declaration and registration record. Owners answer with documents and accept residual risk; the AI compliance lead signs.
The classification agent tests the system against Art. 3(1), Art. 5, Annex I, Annex III and the Art. 6(3) derogation, decides whether you are provider or deployer and finds the date the duties apply. Each step shows the rule, the confidence and what fact would change the outcome. Legal counsel confirms every classification, and anything below 0.80 confidence is escalated.
The section drafters write the prose, but only from passages in your own documents, and every sentence links to the passage it came from. Anything no document supports is listed in red with the article and the person asked. On the sample file, 7 of 9 sections were drafted from 23 documents as 318 cited sentences.
A separate assessor agent reads the finished file as an authority would and raises numbered questions with the article — for example, accuracy by age band under Art. 15(1). Each goes to the person who holds the evidence with a due date, and closes only on a document or a signed answer. Sign-off stays locked while any question is open.
The change monitor checks every new version against the pre-determined change plan and re-opens only the sections a change affects. When a gate fails — on the sample, accuracy for applicants aged 55 and over fell to 0.85, under a 0.87 floor — promotion is held. The system owner decides whether it is a substantial modification under Art. 43(4); the agent never does.
The declaration preparer fills the Annex VIII registration record from the file and the signed declaration. A person submits it in the EU database and enters the registration number; the agents never submit for you.
Yes. Counsel confirms classifications, system owners accept residual risk and decide on retrains, and the AI compliance lead signs the declaration — each with a reason code, and never on work they prepared. Every action by a person or an agent is on a chained audit trail kept 10 years.
The Agentic Solution Engine builds and deploys it from your requirements and documents — your AI system register, a sample of model cards, test reports and design notes, and your thresholds — and it goes live once every quality gate has passed. We will walk you through it on one of your own systems first.
Every request answered inside its legal clock, with no other person's data released.
1-min demoPrivacy Impact AssessmentEvery new project assessed before launch, with each answer traced to its source.
1-min demoData Processing Agreement ReviewEvery supplier DPA checked on all eight Article 28 terms, redlined from your playbook.
1-min demoWe'll run EU AI Act Compliance on one of your own high-risk systems and the documents you already hold for it.