Skip to content
56North

AI Act compliance

AI Act compliance: which evidence to build, and who produces it

Complying with the AI Act is not enough, you have to be able to show it: with evidence that is dated, attributed to its author and intact since it was recorded, covering your own uses rather than the vendor's product.

Who is in scope?

The regulation does not sort companies by size or sector, but by use.

  • Every size. There is no exemption for SMEs. They get lighter requirements, such as simplified documentation and lower fine ceilings, not a waiver.
  • Beyond the EU. The regulation applies to deployers established in the Union, and to providers and deployers established elsewhere as soon as the output of the system is used in the Union (Article 2). A Swiss, Mauritian or US company that recruits, assesses or serves people in the Union is in scope.
  • Every sector. High risk is defined by use (Annex III): biometrics, critical infrastructure, education, employment and worker management, access to essential services (credit, health and life insurance, public benefits), law enforcement, migration, justice. A hotel chain that screens applicants with AI is in scope just like a bank. On top of that comes AI built into products that are already regulated (Annex I), such as machinery or medical devices.
  • Out of scope. Purely personal use, defence and national security, research carried out for its own sake.

The vendor's compliance is not yours

The regulation separates the provider, who develops an AI system and places it on the market under its own name, from the deployer, who uses it under its own authority (Article 3). The vendor documents its product. You answer for how you use it. One party's evidence does not cover the other's obligations.

This page is about evidence. The obligations themselves, platform by platform, are covered in our guide EU AI Act: what companies using Copilot, Agentforce or Joule must do now.

The evidence file, article by article

Each obligation of a company that deploys AI calls for a specific piece of evidence. The file is built as you go, in the order the obligations apply.

  1. 2 February 2025

    In force

    Every company that uses AI

    • Art. 4

      Support the AI literacy of staff

      In the fileThe measures taken, who benefited, on what date

  2. 2 August 2026

    In force

    Transparency

    • Art. 50(3)

      Inform people exposed to emotion recognition or biometric categorisation

      In the fileThe information text and the date it was put in place

    • Art. 50(4)

      Disclose deepfakes and generated text published to inform the public

      In the fileThe disclosures shown, and the human review rule where it applies

  3. 2 December 2027

    Upcoming

    High-risk systems (Annex III)

    • Art. 26(1)

      Use the system according to its instructions

      In the fileThe instructions received, the measures taken to follow them

    • Art. 26(2)

      Human oversight by competent people

      In the fileThe named assignment, the training followed

    • Art. 26(5)

      Monitor, report risks and incidents

      In the fileFindings, reports, suspensions

    • Art. 26(6)

      Keep the logs

      In the fileThe logs, for at least six months

    • Art. 26(7)

      Inform workers and their representatives

      In the fileThe information given, before putting into service

    • Art. 26(11)

      Inform people subject to a decision

      In the fileThe information text and where it appears

Every piece in the file is

  • dated
  • attributed to its author
  • intact
  • tied to a system and an article
  • scoped

The information required by Article 50 must be given at the latest at the first interaction or exposure. Bodies governed by public law, private entities providing public services and certain other deployers must also carry out a fundamental rights impact assessment before deploying a high-risk system (Article 27).

A company that puts an agent into service under its own name may also be a provider. It then answers for Article 50(1): people must know they are dealing with an AI.

What makes evidence hold up

Evidence is judged by how it holds up over time. For an inspector to accept it, it must be:

  • dated: what was true, and on what date;
  • attributed: who established it, by name;
  • intact: nothing can alter it without it showing;
  • attached: to a system in the AI register and to the article it documents;
  • bounded: its scope is stated, including what it does not cover.

Screenshots in a shared folder meet none of these conditions.

Who produces which evidence

Evidence Who produces it
Instructions for use, system documentation The provider, who hands them to the deployer
Register, risk class, named owner The company, in the Cockpit
Cost and usage measurements The Cockpit, from measured calls
Testing and review by people Human in the Loop, an offer activated on engagement
Monthly flight report, frozen and sealed The Cockpit, published by a person

In the Cockpit, every piece of evidence carries a recording date and its author. The monthly flight report is frozen when published and sealed with a fingerprint a third party can check.

What 56North does not do

We do not issue certifications. We prepare an evidence file ready for an inspection, covering the obligations of a company that deploys AI systems, in particular Articles 26 and 50. We never call it a certification.

What missing evidence costs

Breaches of deployer obligations (Article 26) and transparency obligations (Article 50) carry fines of up to €15M or 3% of worldwide annual turnover, whichever is higher, and whichever is lower for SMEs. For prohibited practices, the ceiling is €35M or 7% (Article 99). To see these deadlines in context, read AI governance.

Do you know which AI systems run in your company, and who answers for them?

Request an assessment

A free 30-minute first conversation, no commitment.

Frequently asked questions

Does the vendor's compliance evidence cover us?

No. The vendor, a provider under the regulation, documents its product. The company that uses it is a deployer and answers for its use: human oversight, logs, informing people, transparency.

How long must the logs of a high-risk AI system be kept?

At least six months for the automatically generated logs under the deployer's control, or longer if the system's purpose requires it or another law provides for it (Article 26(6)). The obligation applies to Annex III systems from 2 December 2027.

Is there an AI Act certification for a company that uses AI?

56North does not issue certifications. We prepare an evidence file ready for an inspection, covering the deployer's obligations, in particular Articles 26 and 50.

Sources

Sources opened on 12 October 2026.

  1. AI Act, Article 3: definitions (provider, deployer)
  2. AI Act, Article 4: AI literacy
  3. AI Act, Article 26: obligations of deployers of high-risk systems
  4. AI Act, Article 27: fundamental rights impact assessment
  5. AI Act, Article 50: transparency obligations
  6. AI Act, Article 99: penalties (text reproduced by artificialintelligenceact.eu)
  7. European Commission, AI Act Service Desk: implementation timeline
  8. EUR-Lex: Regulation (EU) 2026/1744 (Digital Omnibus on AI)

Updated 12 October 2026