EU AI Act compliance for AI agents
EU AI Act compliance for AI agents: logging you can prove, not just retain
EU AI Act compliance for AI agents starts with logs that cannot be quietly edited. Article 12 calls for automatic, lifetime event logging for high-risk AI systems. Article 26 makes deployers keep those logs for at least six months. Neither article says the logs may be quietly edited. If they can be, they are worth nothing as proof.
DataShield gives each governed agent tool call a scoped yes or no. It also gets a place in a SHA-256 hash chain with Ed25519-signed checkpoints. A check tells you if someone altered, added to, cut, or trimmed a log's end. You can run that check yourself, today, with no signup.
AI agent audit trail compliance: what regulated deployers really need
Logs that prove, not just record
A log that can be quietly altered proves nothing. We hash-link DataShield's audit chain, row by row. We chain the signed checkpoints too, so even a dropped checkpoint shows up. Key rotation fails closed. Designed to map to EU AI Act Art. 12/26 and HIPAA §164.312(b). Verify a sample chain.
Authorization per tool call
Every governed tool call has three gates to pass. Token scope ceiling. Authority tier. A fresh check that no one revoked the session. Then we meter it, dispatch it, and tie it to an agent identity. Break-glass access expires on its own, drops mid-session, and cannot be quietly cut from the log. See Auth.
Data minimization by design
EDPB Opinion 28/2024 names pseudonymization as a GDPR mitigation. DataShield tokenizes identifiers at ingest with deterministic, join-preserving tokens. It rolls up quasi-identifiers. Turning a token back is a walled-off step. We log it. Erasure means crypto-shred. The audit chain lives through it. See Ontology.
EU AI Act Article 12 logging requirements and ISO 42001, stated plainly
Commission fining powers start on August 2, 2026. Fines reach €35M or 7% of turnover. But the Digital Omnibus pushes back two dates. Annex III high-risk grading moves to December 2027. Embedded medical AI moves to August 2028. We won't play scary with the deadline. Here is what already bears weight. Art. 12 lifetime event logging and Art. 26 deployer log retention are in force now. They anchor the audit evidence for each high-risk rollout plan.
ISO/IEC 42001 is turning into a buying rule of its own. Colorado's AI Act makes NIST AI RMF / ISO 42001 alignment an affirmative defense. Both frameworks land on the one question every audit asks. Prove what happened, when, by which agent, against which policy. DataShield's answer is a record. We meter and tie each tool call to an agent, then seal it into a chain. We publish the schema of its check report, word for word, on the architecture page.
Regulated buyers care who owns the referee. Lakera went to Check Point. Portkey to Palo Alto. Prompt Security to SentinelOne. CalypsoAI to F5. The neutral governance layer keeps vanishing into platform vendors. DataShield runs self-hosted, in your own cloud or data center. Or on a dedicated single-tenant server we run. Either way, you hold the keys. See Security.
ISO 42001 needs infrastructure, not just a report
GRC platforms automate the paperwork of AI governance: framework mappings, policy templates, attestation gathering. What they cannot do is BE the control. ISO 42001's Annex A asks for AI event logging, access control and data management. Those controls have to exist in your stack. Only then can any tool attest to them.
DataShield is the layer the report points at. The tamper-evident event log an auditor can re-check without trusting a soul. The per-tool-call access control the checkbox claims. The tokenization and crypto-shred that make the data-management clause true. Keep your GRC platform. Just point it at controls that prove themselves.
EU AI Act compliance for AI agents: buyer questions
Does DataShield make our AI system EU AI Act compliant?
No vendor can. Compliance spans risk management, human oversight and paperwork that belong to you. DataShield brings the technical base for the logging and access-control duties. That means Art. 12-style automatic event logging with tamper evidence. Retention that stays checkable. A record of the yes or no behind each call.
How is this different from shipping logs to our SIEM?
Anyone can edit SIEM logs. An admin or an attacker can change one and nothing flags it. In DataShield's chain, we hash-link each row. We sign and chain the checkpoints too. So a check tells you if someone altered, added to, cut, or trimmed a log's end. Ship to your SIEM as well if you like. The chain is the proof.
Can we satisfy GDPR erasure without destroying our audit trail?
Yes. Actor identifiers are HMAC-committed in the chain, and erasure means crypto-shred. Destroy the key material, and the personal data goes dark for good. The chain still checks out, end to end.
What exactly gets logged per agent action?
Each governed tool call records the agent identity. The token scope it ran under. The authority-tier call. The revocation re-check. The metering. Then it seals that record into the chain. We publish the full dispatch pipeline and the ChainVerifyReport schema on the architecture page.
Lifetime logging with tamper evidence. Authorization on each tool call. Erasure that costs you no proof. All of it self-hosted.
Explore the Demo CenterYou've seen the proof
Ready for a number? Scope your deployment and we'll price it against your own economics.
Get your quote →