Use the current timeline

As checked on October 2, 2026, the European Commission records that the AI Omnibus entered into force on July 27, 2026. Its timeline places Annex III high-risk system requirements at December 2, 2027 and high-risk systems embedded in regulated Annex I products at August 2, 2028. Earlier obligations and the general application date have their own rules. The previous blanket statement that high-risk duties begin in August 2026 was outdated. See the Commission’s implementation timeline and consolidated regulation.

Identify the role and intended use

Classify each system before answering a customer questionnaire. A model provider, downstream application provider, deployer, importer and distributor have different responsibilities. Using an external model does not settle your role or make every system high risk. Review the intended purpose, placement on the market, affected people, Annex categories and relevant exceptions.

An AI sales assistant and an employment-screening system can use similar technology but have very different legal and operational contexts. Document the analysis for the specific workflow and obtain legal review where classification is uncertain.

Explain AI interaction and data use clearly

Evaluate applicable transparency duties for the feature, including whether people need to know they are interacting with AI. Keep that explanation near the interaction. A cookie toggle by itself does not explain model behavior, human escalation, retention or third-party processing. Review the Commission’s AI Act guidance alongside privacy obligations.

Build evidence that engineering can maintain

  • Inventory models, providers, versions, agents, retrieval stores and tools.
  • Map customer data and identities through retrieval, memory, context and tool execution.
  • Test tenant boundaries, indirect prompt injection, unauthorized actions and leakage using authorized test data.
  • Record limitations, findings, remediation and retest results.
  • Define who reviews a major model, provider, tool or architecture change before release.

These are practical assurance activities. A security assessment does not decide the legal classification or provide a conformity assessment, CE marking, SOC 2 attestation or ISO certification.

Match the engagement to the trigger

For one procurement questionnaire, start with the Security Gate Sprint. For a production AI feature, scope an AI Application & Agent Security Assessment. For changing agents and tool integrations, Continuous AI Assurance defines a recurring testing and review cadence.

For engineering detail, read the MCP security testing guide and prompt injection guide. Treat statutory deadlines and customer contract deadlines as separate entries in the release plan.