Skip to content
EU AI Act

AI Governance Checklist for the EU AI Act's Staged Deadlines

8 min readUpdated 2 May 2026

The EU AI Act's high-risk obligations apply from 2 December 2027 (Annex III systems) and 2 August 2028 (Annex I products). If your company builds or deploys AI systems in the EU, those are the dates to plan toward.

This checklist covers every action a SaaS company needs to complete. Use it to assess your current state and identify gaps.


Step 1 — Inventory and Classify

Before you can comply, you need to know what you have.

  • List every AI feature in your product. Include ML models, scoring systems, recommendation engines, chatbots, AI-generated content, and any third-party AI APIs you call.
  • List every AI tool used internally. Hiring tools, performance monitoring, customer risk assessment — any AI used in operations that affects people.
  • Classify each system against the four AI Act tiers: prohibited, high-risk, limited transparency, minimal risk.
  • Document your classification rationale for each system in writing — not just the conclusion, but why.
  • Check against the prohibited list (Article 5). Remove any prohibited features immediately. The deadline for this was February 2, 2025.
  • Check against Annex III for each system. If you operate in or sell to HR, credit, insurance, education, or essential services sectors, assess carefully.

Step 2 — Address Prohibited Practices Immediately

If you have not already done this, it is urgent.

  • Review your product for any features that could constitute social scoring, real-time mass biometric surveillance, subliminal manipulation, exploitation of vulnerable groups, or prohibited emotion recognition.
  • Remove or redesign any feature that meets the prohibited definition.
  • Document the removal — if you modified a feature, record what changed and why.

Step 3 — High-Risk AI: Full Compliance Programme

For each system classified as high-risk:

Risk Management

  • Implement a risk management system covering the system's lifecycle (Article 9)
  • Identify, analyse, and evaluate risks to health, safety, and fundamental rights
  • Define risk mitigation measures and document them
  • Schedule regular risk management reviews — at least annually and after significant system changes

Data Governance

  • Document the provenance of all training, validation, and testing datasets (Article 10)
  • Evaluate training data for representativeness — does it reflect the population your system will be used on?
  • Conduct a bias evaluation across relevant protected characteristics (gender, ethnicity, age, etc.)
  • Document any known data gaps and the mitigations applied
  • Confirm GDPR lawful basis for any personal data in training sets
  • Implement ongoing monitoring for data drift post-deployment

Technical Documentation

  • Produce full Article 11 technical documentation for each high-risk system
  • Include: system description, intended purpose, design methodology, architecture, training data, performance metrics, known limitations
  • Ensure documentation is sufficient for a conformity assessor to evaluate the system
  • Establish version control — documentation must track to specific model versions

Record-Keeping and Logging

  • Implement automatic event logging sufficient to identify risks and enable post-deployment monitoring (Article 12)
  • Ensure logs capture: when the system was used, inputs, outputs, and any flags or errors
  • Establish log retention period and access controls

Transparency to Deployers

  • Produce Article 13 instructions for deployers — covering intended purpose, performance, limitations, oversight requirements
  • Ensure deployer documentation is provided at product onboarding
  • Update documentation when the system changes significantly

Human Oversight

  • Design the product to enable human oversight — deployers must be able to review, interpret, and override AI outputs (Article 14)
  • Ensure the UI does not discourage override or present AI outputs as final
  • Confirm the system can be stopped or paused by a human operator

Accuracy, Robustness, and Cybersecurity

  • Define and document accuracy and performance metrics for the system
  • Test robustness against adversarial inputs, errors, and inconsistencies (Article 15)
  • Assess cybersecurity of the model against manipulation attempts

Conformity Assessment

  • Complete a self-assessment conformity process (most Annex III systems)
  • For biometric identification and critical infrastructure AI, engage a notified body for third-party assessment
  • Produce a Declaration of Conformity (Article 47)
  • Affix CE marking to relevant documentation

EU AI Database Registration

  • Register each high-risk AI system in the EU AI database before placing on the EU market (Article 49)
  • Keep registration current — update when significant changes are made

Step 4 — Limited Transparency AI (All Interactive and Generative AI)

For chatbots, AI assistants, generative AI features, and emotion recognition:

  • Add a clear disclosure at the start of every AI interaction informing users they are interacting with an AI system (Article 50)
  • Implement labelling for AI-generated content (text, images, audio, video) presented to users
  • For emotion recognition: inform individuals before assessment occurs
  • Review your product UI — disclosures must be visible and timely, not buried in settings

Step 5 — General Purpose AI (If Applicable)

If you build or fine-tune foundation models (GPAI):

  • Classify the model — is it a GPAI model with systemic risk?
  • Produce technical documentation and a copyright policy (Article 53)
  • Make a summary of training content publicly available
  • For systemic risk GPAI (10^25 FLOPs threshold): complete adversarial testing, incident reporting setup, cybersecurity measures

GPAI obligations apply from August 2, 2025 — earlier than the high-risk deadline.


Step 6 — Post-Market Monitoring Plan

High-risk AI compliance does not end at deployment.

  • Implement a post-market monitoring system to collect and analyse performance data (Article 72)
  • Define KPIs for ongoing monitoring — accuracy, error rates, bias metrics
  • Establish a process for reporting serious incidents to national authorities (Article 73)
  • Set a review cadence — when will you re-assess the system after deployment?

Step 7 — GDPR Overlap

Where your AI system processes personal data:

  • Confirm GDPR lawful basis for all processing activities in the AI system
  • Update privacy notices to include AI processing disclosure (Articles 13/14 GDPR)
  • Conduct DPIA if not already done for high-risk processing
  • Ensure GDPR Article 22 (automated decisions) rights are addressed alongside AI Act human oversight requirements
  • Review data processing agreements with any AI sub-processors

Timeline Summary

DeadlineAction required
Now (overdue)Remove prohibited AI practices
2 August 2025GPAI model obligations — documentation and transparency
2 December 2027High-risk (Annex III) systems: conformity assessment, EU database registration, human oversight, technical documentation (2 August 2028 for Annex I products)

Where to Start If You're Behind

If you have not started:

  1. Classify first — without a clear inventory and classification, you cannot plan
  2. Focus on prohibited items — any prohibited AI must be removed, no exception
  3. Identify your one or two most clearly high-risk systems — get documentation started on those before addressing everything else
  4. Treat the conformity assessment as a project, not a document — it requires input from engineering, legal, and product

Most startups can reach a defensible compliance position with 20–40 hours of structured work. The documentation is not technically complex — it requires organisational effort more than legal expertise.

Work out whether you are a provider or a deployer under the AI Act, and what each one owes.

Take the free AI Act role check →