B2Shift · Published · Updated

This EU AI Act deployment checklist is an engineering aid for teams building or using AI systems. Use it as an AI Act checklist to organise the questions, owners and evidence to take into a qualified legal or compliance review. This EU AI Act compliance checklist prepares evidence for that review; it does not determine whether a system is compliant, replace the Regulation, or provide legal advice. The Commission’s implementation guidance and the applicable legal text should be checked again before deployment because timelines and guidance can change.

Start with the system and the business use

Create one inventory entry for each AI system and materially different use. Record the business process, intended purpose, affected people, deployment countries, model and provider, connected data, automated outputs, system owner and any consequential actions. A general label such as ‘internal assistant’ is too broad: answering policy questions, ranking job applicants and approving credit are different uses with different risk questions.

Capture the real workflow around the model. Note what information it receives, who sees its output, whether a person can meaningfully review it, what systems it can change, and what happens when it is wrong or unavailable. Include the expected user-facing disclosures and the evidence retained for monitoring.

Identify your role and the provider’s role

Document whether your organisation is acting as provider, deployer, importer, distributor or in more than one role for the specific system and use. Record who supplies the model, who integrates it into a product or process, who determines the intended purpose, and who operates it. Contract language alone does not settle every role question; the facts of the use and the Regulation matter, so obtain legal review where responsibilities are unclear.

For each external provider, collect the system description, intended-use limits, instructions, known limitations, update and incident notices, data handling terms, and technical information needed for your own assessment. Keep the provider version and date so a later model or configuration change can be compared with the one you reviewed.

Screen for prohibited practices and high-risk use

Before a pilot, compare the actual use against the Commission’s current guidance on prohibited practices. Do not rely on a vendor’s category label or a model card alone. If the use may fall within a prohibition, stop the deployment path and obtain specialist legal advice before testing it with people or production data.

Next, assess whether the system falls within a high-risk category under the Act, including the relevant Annex III use cases or an AI system that is a safety component of a regulated product under Annex I. Record the use-specific reasoning and the person who approved the classification. Do not assume that a tool is low risk because it is general-purpose or because a person remains nominally in the process.

When an AI system influences employment, access to essential services, education, safety or another consequential decision, assess the actual decision pathway and applicable sector rules with counsel. This checklist does not classify a system for you.

Put operational controls around the use

For any system approved to proceed, set controls that match the consequences of failure. Define least-privilege access, permitted inputs and outputs, the human review point, prohibited actions, escalation owner, fallback process and rollback path. A reviewer needs time, context and authority to reject an output; a button that only records approval is not meaningful oversight.

Test the system against representative and difficult cases before release. Include missing, conflicting and out-of-distribution inputs; measure the errors that matter for the use; and define who can pause the system when quality, security or operating conditions change. Keep test results, known limitations, approvals, configuration versions and material changes together in an evidence record.

Make user-facing transparency part of the product design where required. The Commission states that Article 50 transparency obligations apply from 2 August 2026. For systems placed on the market before that date, its FAQ describes a limited extension only for Article 50(2) marking and detection duties for AI-generated content until 2 December 2026; it does not defer the other Article 50 duties. Check whether the narrow transition and its conditions apply to your system rather than adding a generic AI label and assuming that resolves the requirement.

Prepare for monitoring, incidents and changes

Name the system owner and the people responsible for technical operation, user support, security incidents and compliance questions. Record the channels through which a user or affected person can report a problem. Define what is logged, who can access logs, how long they are retained, and how an incident is escalated and investigated.

Treat model, prompt, data-source, workflow and provider changes as release changes. Reassess the intended purpose and risk when a change expands the users, decisions, data or actions. Keep a dated change log, repeat the tests affected by the change, and document the approval to resume normal operation.

Check the dates against current official guidance

As of 15 October 2026, the Commission’s current AI Act pages state that the Act generally became applicable on 2 August 2026, with earlier application for prohibited-practice and AI-literacy provisions from 2 February 2025 and governance and general-purpose AI obligations from 2 August 2025. The AI Omnibus entered into force on 27 July 2026; the Commission’s resulting timeline gives 2 December 2027 for high-risk use cases under Annex III and 2 August 2028 for high-risk AI embedded in regulated products under Annex I. Confirm the dates and transitional rules that apply to your system directly against the latest official materials before setting a project deadline.

Evidence to bring to an expert review

Prepare the system inventory, workflow diagram, role analysis, intended-purpose statement, provider documentation, risk classification rationale, data and access map, test results, human-oversight design, transparency copy, monitoring plan and change log. Ask the reviewer to identify the applicable obligations, the evidence gaps, the responsible legal entities and the release conditions. Do not describe a completed engineering checklist as certification or a guarantee of compliance.

This is general operational information, not legal advice. For data protection controls that may apply alongside the AI Act, see our guide to safe, GDPR-aware AI workflows.

Build an evidence pack that changes with the system

Keep an inventory for each AI-enabled system and the workflow in which it is used. Record the intended purpose, the components and external services involved, the business owner, who operates it, who receives its output and whether that output can affect a decision or trigger an action. Map the boundary between the system provider, deployer and other suppliers from the actual arrangement; do not infer legal roles only from who bought a licence or wrote the integration. A role or risk classification may require qualified legal review.

For each material risk or control, keep a short evidence entry: the use scenario, the issue being addressed, the control, the accountable owner, the artifact or test result and when it was last reviewed. Evidence might include a representative evaluation set, access-control configuration, human review procedure, incident runbook, user-facing notice or log sample. Keep an unresolved question visible with an owner and target review date instead of describing it as completed.

Treat changes as a reason to revisit the assessment. A new model or provider, changed data source, different user group, new integration, expanded purpose or greater ability to act can change the system boundary and its risks. Record the version deployed, what changed, who approved the change and which checks were repeated. The Commission's current framework, enforcement and AI Omnibus pages should be rechecked before relying on a date or scope statement; application timelines and supporting guidance can change.

This evidence pack helps engineering, product and legal reviewers work from the same system description. It is preparation for expert review, not a finding that a system is high-risk, exempt or compliant. Keep legal conclusions, the facts reviewed and the responsible decision-maker separate from engineering test results.

Engineering evidence to prepare for a qualified AI Act review
TopicUseful evidence to collectOwner to involve
System and purposeVersioned diagram, intended use, users and affected workflowProduct and business owner
Roles and suppliersContracts, instructions, system boundary and responsibility mapProcurement and legal
Data and evaluationData sources, test cases, limitations and observed errorsEngineering and data owner
Human oversightReview steps, intervention authority, escalation and pause routeOperational owner
Monitoring and incidentsLogs, incident path, review cadence and change historySecurity and service owner
Open questionsIssue, owner, evidence needed and review dateQualified legal/compliance reviewer

Make evaluation evidence reproducible

For each important test, record the scenario, expected behaviour, observed output, source or model version, relevant configuration, reviewer and date. Include ordinary cases and deliberately difficult ones: missing or conflicting information, an unsupported request, an unavailable tool, a permission denial and a case that must stop for human review. Preserve enough context for another reviewer to repeat the check without retaining more sensitive input than the organisation needs. When a provider, model, data source, prompt, connected tool or intended workflow changes, identify which tests must run again and compare the result with the last accepted version. This is engineering evidence for review; it does not itself determine a legal category or compliance outcome.

Sources

European Commission — AI Act and application timeline

European Commission — AI Act enforcement framework and timeline

European Commission — AI Omnibus enters into force

European Commission — Navigating the AI Act

European Commission — Transparency obligations under Article 50

Discuss your workflow