This is a working checklist, not legal advice — every organization's obligations depend on its specific use cases, data, and jurisdiction. Treat it as a starting point for a conversation with your legal and compliance teams, not a substitute for one.
Foundational data governance
- Map every system that feeds data into an AI tool — connectors, uploads, and integrations all count, including ones set up informally by individual teams.
- Classify data by sensitivity before it reaches a model: public, internal, confidential, and regulated (PII, health, financial) each need different handling rules.
- Confirm a Data Processing Agreement is in place with every AI vendor that touches regulated data, covering retention, sub-processors, and deletion rights.
- Document your lawful basis for processing personal data through AI systems under GDPR Article 6, and update your Records of Processing Activities.
EU AI Act specific steps
- Classify each AI use case by risk tier (unacceptable, high, limited, minimal) — the obligations attached to a use case scale directly with its tier.
- For high-risk use cases, complete a conformity assessment and maintain the required technical documentation before deployment.
- Implement human oversight for any use case that falls in the high-risk tier — this typically means a named person reviewing and approving consequential outputs before they take effect, not just a theoretical ability to intervene.
- Ensure transparency obligations are met: users interacting with an AI system need to know they're doing so, and AI-generated content that could be mistaken for human-generated content needs to be labeled where the Act requires it.
Security and access controls
- Enforce SSO and role-based access so AI tool access follows the same identity governance as every other enterprise system.
- Confirm PII and sensitive data are redacted or tokenized before being sent to any third-party model provider, unless a specific agreement permits otherwise.
- Test your prompt-injection and data-exfiltration defenses against your actual connector setup, not just in the abstract.
Auditability
- Maintain an immutable audit log of every AI interaction that touches regulated data or takes a consequential action — who asked, what was retrieved, what was generated, and who approved it.
- Confirm the audit log can be exported in a format your compliance and legal teams can actually use during a regulatory inquiry, not just a raw database dump.
- Set a retention policy for audit records that matches your regulatory obligations, which are often longer than your default data retention policy.
Ongoing operations
- Re-run the risk classification whenever a use case's scope changes — a tool that started as an internal drafting assistant can drift into a high-risk use case without anyone updating its classification.
- Train employees on acceptable use, including what data should never be pasted into an AI tool that hasn't been vetted by your security team.
- Review vendor compliance posture (SOC 2, ISO 27001, sub-processor lists) at renewal, not just at initial signing.