AI-agent skill · Claude Code · Codex · Cursor · Gemini CLI

Is your AI product high-risk under the EU AI Act?

EU AI Act Readiness Kit lets your AI coding agent check an AI product against the EU AI Act. It scans the code for AI use, works out the risk category with the articles behind it, lists every obligation that applies with its date and the days left, and drafts the paperwork: the Annex IV technical documentation, the Article 6(3) record, transparency notices and an AI literacy plan. Built on the law as amended by the 2026 Digital Omnibus (Regulation (EU) 2026/1744).

Not legal advice. This is a readiness aid. It can be wrong, and the law and its guidance are still changing. Every report says so and asks you to have a qualified lawyer confirm the result before you rely on it.
Get it on Agensi · $29 ↗Blind test resultsKey datesGitHub ↗

Tested on the European Commission's own examples

In May 2026 the Commission published draft guidelines with hundreds of practical examples of AI systems that should or should not count as high-risk. I collected 312 of them from the official AI Act Service Desk and split them by area. I built the skill on four areas only. Then I froze it and gave the other 166 examples (education, essential services and credit, migration, justice and elections, and cross-cutting cases) to a fresh AI agent that had never seen them or their answers. It followed the skill once.

161 / 166correct high-risk / not-high-risk verdicts in the blind test (97%)
64 / 64high-risk systems recognised: none missed
5 / 102not-high-risk systems wrongly flagged as high-risk (all errors were on the cautious side)
144 / 146on the four development areas (employment, biometrics, critical infrastructure, law enforcement), used to build the skill
The 5 blind-test mistakes
CasePointExample (Commission, paraphrased)CommissionSkill
edu283(d)Plagiarism checker compares submitted homework against existing content.not high-riskhigh-risk
edu303(d)AI helps a human proctor confirm or deny already suspected cheating by analysing the flagged student's data; the proctor decides.not high-riskhigh-risk
ess275(b)Margin credit assessment for leveraged trading.not high-riskhigh-risk
ess455(d)Manages network traffic for emergency calls.not high-riskhigh-risk
hor015(b)Credit score for a sole trader's business using only business data.not high-riskhigh-risk
Results by area
Blind test areaCorrect
education29/31
essential-services47/49
horizontal12/13
justice-democracy27/27
migration46/46
Development areaCorrect
biometrics40/41
critical-infrastructure25/25
employment32/32
law-enforcement47/48

Honest notes: the examples are short paraphrases of the Commission's draft guidelines, which are not final and not binding. In 16 of the 97 correct "not high-risk" answers, the skill gave a different reason than the Commission ("outside the use case" versus "Article 6(3) exemption"). That matters because the exemption must be documented and registered. This test covers the high-risk decision only; prohibited practices, transparency duties, GPAI rules and dates are covered by 44 automated tests.

Buyer test on six real open-source AI products

A separate fresh agent installed the frozen skill and ran it end to end on six real GitHub projects, with nobody to answer its questions. I wrote down my expected answer for each project before it ran.

ProjectWhat it isExpectedSkill's resultMatch
sliday/resume-job-matcherAn employer screens and ranks resumes to decide who to interviewHigh-risk, Annex III 4(a)High-risk, 4(a) (profiling, no exemption) + transparency✅ match
max-fitzpatrick/Credit-scoring-modelPredicts a consumer's probability of defaultHigh-risk, Annex III 5(b)High-risk, 5(b)✅ match
srbhr/Resume-MatcherA job seeker tailors their own resume with an LLMNot high-risk (candidate's own tool); transparencyTransparency (Art 50(2)), not high-risk✅ match
mckaywrigley/chatbot-uiChat interface for LLMsTransparency (Art 50(1), 50(2))Transparency (Art 50(1), 50(2))✅ match
kmhmubin/Face-Recognition-Attendance-SystemWebcam face recognition marks students' attendanceBorderline: must weigh 'remote' identification (1(a)) vs people stepping up to the cameraHigh-risk 1(a) on the stricter reading, with the 'remote' question and the way to narrow it explained✅ match (borderline, reasoned)
lllyasviel/FooocusImage generator (SDXL) with face swap and inpaintingTransparency (Art 50(2)); mention the 2 Dec 2026 bansv1.0.0 said 'PROHIBITED'. The law (Art 5(1a)) bans such generators only without adequate safeguards, which the agent found are off by default. v1.0.1 now says: prohibited from 2 Dec 2026 unless safeguards are added, plus Art 50 duties⚠️ partial: right issue, too strong a verdict; fixed in v1.0.1

Every command worked the first time on all six projects; the scan took 0.1 to 3.9 seconds. 5 of 6 matched the expected answer and 1 was partly right. The buyer agent also found real problems, all fixed in v1.0.1: 'unsure' answers were not always treated strictly, the borderline note was not printed, re-running overwrote an edited documentation draft, the pre-fill counted weak evidence, and some scanner signals were noise. The Annex IV drafts started with 3 to 10 of 26 items pre-filled from the code; after the agent completed them, 12 to 21 of 26 had content and the rest needs the team.

What your agent does with it

  1. Scans the codebase for AI use: model libraries, calls to OpenAI, Anthropic, Gemini, Mistral and others, face and voice biometrics, emotion recognition, chat interfaces, generated media, with file and line.
  2. Fills a structured questionnaire from the code, the docs and your answers: role (provider or deployer), intended purpose, company size.
  3. Checks all 10 prohibited practices, including the two new bans from 2 December 2026, then the high-risk routes (Annex I products, Annex III use cases), the Article 6(3) exemption and profiling, transparency duties and general-purpose AI rules.
  4. Writes a report: risk category with articles, every obligation with who, what, the date and the days left, penalties and key dates.
  5. Drafts documents: Annex IV technical documentation (9 sections, prefilled from the code where possible), Article 6(3) assessment record, transparency notices, safeguards record for generators, AI literacy plan.
  6. Answers you in plain English: what applies now, what comes next, the first actions, and the open questions. Always with the not-legal-advice note.

Key dates after the 2026 Digital Omnibus

2 Feb 2025 Prohibited practices banned; AI literacy duty.

2 Aug 2025 General-purpose AI model rules.

2 Aug 2026 Transparency duties: tell people they talk to AI, mark AI-generated content, label deepfakes.

2 Dec 2026 New bans on AI that makes non-consensual intimate imagery or child sexual abuse material; generators already on the market must mark outputs.

2 Dec 2027 High-risk rules for Annex III use cases (hiring, education, credit, insurance, biometrics...).

2 Aug 2028 High-risk rules for AI in regulated products (medical devices, toys, lifts...).

Good to know

Who it is for

European start-ups and agencies that build AI products and want to know where they stand before customers, investors or authorities ask.

Small install

One Python script, standard library only (Python 3.10+). Nothing leaves your machine.

Conservative by design

When a deciding fact is unknown, it asks and assumes the stricter reading. In the blind test every error was a false alarm, never a miss.

Honest limits

It does not cover GDPR, copyright, product liability or sector law, and it cannot replace a lawyer. The law version is printed on every report.

Get it on Agensi · $29 ↗GitHub ↗

Made by Muhammad Shahmeer · Data Science MSc, University of Naples Federico II · © 2026, all rights reserved. Not legal advice. Not affiliated with the European Commission. Examples paraphrased from the Commission's draft guidelines (AI Act Service Desk).