Practical guide

EU AI Act checklist for product teams.

Who does what at each stage of delivery, and what evidence to keep. Use it for any AI feature that talks to people, creates content, or helps decide about them.

1 · Before you build

  • Check it isn’t prohibited (Article 5), including the ban that applies from 2 December 2026 on AI that generates sexual deepfakes of real people or child sexual abuse material.
    Owner: Product with Legal · Keep: prohibited-practice check
  • Write the intended purpose, and what the AI is not for.
    Owner: Product · Keep: intended-purpose statement
  • Check whether it’s high-risk. Is the use listed in Annex III, such as credit scoring, CV screening or exam grading, or is the AI a safety component of a product covered by Annex I? A listed use is high-risk unless it doesn’t materially influence decisions about people, and profiling people always counts (Article 6). Record the reasoning either way.
    Owner: Product with Compliance · Keep: classification record
  • Decide whether you’re provider or deployer. Putting your name on a vendor’s high-risk AI, substantially changing it, or repurposing an AI for a high-risk use can make you the provider (Article 25).
    Owner: Product with Legal · Keep: role decision
  • Ask your AI vendor the right questions, in writing: purpose, risk class, documentation, test results, logs, oversight, change notices. Vendors supplying parts of a high-risk system must help you comply (Article 25(4)).
    Owner: Procurement with Product · Keep: supplier due diligence
  • Map risks, misuse and harm scenarios, and keep the map up to date (Article 9, risk management). High-risk.
    Owner: Product with Design · Keep: risk log
  • Support AI literacy for the people who build, operate and oversee it (Article 4).
    Owner: People team with Product · Keep: training records
  • If you deploy a high-risk system: tell workers before using it at work, and check whether you need a fundamental rights impact assessment (Articles 26 and 27). High-risk.
    Owner: Legal and HR with Product · Keep: notices, impact assessment

2 · Design

  • Map every AI touchpoint: where AI talks to, or decides about, a person.
    Owner: Design · Keep: touchpoint map
  • Tell people they’re dealing with AI at the latest at the first interaction, unless it’s obvious, in a clear and accessible way (Article 50). Disclose deepfakes too. Don’t rely on the terms and conditions.
    Owner: Design · Keep: AI label and notice components in the design system
  • Design meaningful oversight: review, override, stop and escalate, with the time and information a reviewer needs to judge (Article 14). High-risk.
    Owner: Design with Product · Keep: oversight design and decision points
  • Design explanations and a way to challenge a decision. People affected by a decision based on a high-risk system can ask for an explanation (Article 86).
    Owner: Design · Keep: explanation and redress flows
  • Check every AI screen for accessibility (European Accessibility Act; Article 16 for providers of high-risk systems).
    Owner: Design · Keep: accessibility test results

3 · Build

  • Mark generated content in a machine-readable way if your system generates audio, images, video or text (Article 50(2)). New systems since 2 August 2026; systems already on the market by 2 December 2026.
    Owner: Engineering · Keep: marking method and tests
  • Build in automatic logging (Article 12), and keep the logs for at least six months (Articles 19 and 26). High-risk.
    Owner: Engineering · Keep: logging specification
  • Set acceptance criteria and test against them: accuracy, robustness, security and bias. Required for high-risk systems, including data governance (Articles 10 and 15); good practice for everything else.
    Owner: Engineering and Data, with Product · Keep: test results
  • Write the technical documentation as you build, and finish it before release (Article 11 and Annex IV). High-risk.
    Owner: Engineering with Product · Keep: technical file
  • Run it inside a quality management system that covers design, testing, changes and records (Article 17). High-risk.
    Owner: Compliance with Engineering · Keep: quality management records

4 · Release gate: don’t ship until…

  • the prohibited-practice check, risk classification and provider or deployer decision are recorded
  • AI labels and notices are in place, visible and accessible
  • generated content is marked, if the system creates it
  • the change impact check for this release is done
  • open questions for Legal have answers
  • High-risk: a reviewer can really override and stop the system, and that has been tested; logs are on and retrievable
  • High-risk providers: conformity assessment done, EU declaration of conformity signed, CE marking applied and the system registered in the EU database (Articles 43, 47, 48 and 49)

Owner: Product signs off, with Legal and Compliance · Keep: release sign-off record

5 · After launch

  • Monitor what matters: overrides, complaints, errors and drift, not just conversion. High-risk providers need a post-market monitoring plan (Article 72).
    Owner: Product with Data · Keep: monitoring reports
  • Re-check the classification whenever the feature changes significantly. For a high-risk system, a substantial modification means a new conformity assessment (Article 43(4)).
    Owner: Product · Keep: change log
  • Report serious incidents within 15 days at most, and sooner in the gravest cases (Article 73). Deployers tell the provider first (Article 26). High-risk.
    Owner: Compliance with Engineering · Keep: incident records
  • Keep the documentation for 10 years after the system is placed on the market (Article 18). High-risk.
    Owner: Compliance · Keep: technical file, quality records, declaration of conformity

An example: same feature, different obligations

A spend-management app uses AI to set company card limits. If the company repays the card, the AI is assessing a business, so it isn’t a listed high-risk use: the general items still apply, such as the prohibited-practice check, AI literacy and, wherever the AI talks to people, the notices. If employees repay the card personally, the AI is now assessing a person’s creditworthiness, a use listed in Annex III. The same feature is then very likely high-risk, and every item marked high-risk applies, from oversight and logging to technical documentation and incident reporting.

An illustrative example, not a client case. It depends on who is being judged.

Want help running this on your product?

Start with a readiness review of one product, or a workshop with your teams.