What belongs on an AI agent compliance checklist?

Per-call execution records, tamper-evident storage, independent verifiability by third parties, input and output data lineage, human oversight records, and defined retention with fast retrieval. Policies and training cover intent; these six cover evidence, which is what auditors accept.

Why do agent compliance programs fail audits?

Two usual reasons. First, records are summaries rather than per-call evidence: a dashboard line like 14 tool calls cannot answer what each call did. Second, records live only in the vendor's console and were written by the agent itself, so they are neither tamper-evident nor independently checkable.

What does the EU AI Act require for AI agent records?

Article 12 requires high-risk AI systems to keep automatically generated logs enabling traceability, including periods of use, reference databases consulted, and input data. Enforcement is underway, and logs an agent could silently rewrite will not satisfy auditors who test the trail.

How do verifiable receipts help with compliance?

AER-1 verifiable execution receipts, published as an IETF Internet-Draft, are minted per tool call by the execution layer with tool, caller, timestamp, canonical inputs and outputs, a SHA-256 commitment, and sequence position. One primitive delivers per-call records, tamper evidence, and independent verifiability at once.

What is the fastest compliance win for an agent team?

Start minting verifiable receipts for every tool call before the next audit cycle. It is the one change that turns every downstream checklist item from a policy claim into checkable evidence.

AI Agent Compliance Checklist: The Evidence You Actually Need

The checklist

Per-call execution records. Every tool call and model invocation logged with who, what, when, and the exact inputs and outputs. Not sampled, not summarized: per call.
Tamper-evident storage. Records the agent cannot rewrite. Minted by the execution layer, with cryptographic commitments, append-only.
Independent verifiability. A third party, auditor, regulator, or client, can check the records without your infrastructure or your permission.
Data lineage. Which data went in and what came out, in canonical bytes, so a decision can be traced to its inputs.
Human oversight records. Where a human approved, overrode, or reviewed, with timestamps. The absence of oversight should be visible, not assumed.
Retention and retrieval. Defined retention windows, and records retrievable by run, by agent, by time, and by tool, fast enough for an audit.

Where teams usually fail the checklist

The common failure is item one: records exist, but they are summaries. A dashboard that says task completed, 14 tool calls is not a per-call record. When an auditor asks what the 14 calls did, the team exports a log the agent wrote itself, which fails item two as well.

The second common failure is item three. Records live in the vendor's private console, checkable only by logging in. That is access, not verifiability. An auditor wants to recompute, not to browse.

How verifiable receipts cover the hard items

AER-1 verifiable execution receipts, published as the IETF Internet-Draft draft-zambo-aer1, mint a receipt per tool call at execution time: tool, caller, timestamp, canonical inputs and outputs, SHA-256 commitment, sequence position. That single primitive covers items one, two, and three at once:

Per-call: one receipt per call, by construction.
Tamper-evident: minted by the execution layer, commitments anyone can recompute.
Independently verifiable: no account, no API key, no trust in the operator.

Items four through six are policy and plumbing around the receipts: store them with retention windows, index them for retrieval, and log human decisions the same way. The receipts are the evidence layer everything else references.

Hand this to your auditor

The Auditor Packet is the printable version: what a verifiable receipt is, a real sample to check, and the checklist in auditor language. Get the Auditor Packet.

Regulations this maps to

The EU AI Act's Article 12 requires automatically generated logs for high-risk systems that enable traceability: periods of use, reference databases, input data. AER-1 receipts are that logging duty in executable form. US state rules, like Utah's AI disclosure requirements, similarly turn on provable records of what the system did.

None of these regulations accept a log the agent could have rewritten. The checklist item that matters most to every regulator is tamper evidence, because everything else is commentary on records that might be fiction.

Start the evidence chain today

You can generate a real verifiable receipt in under a minute, free, no signup. That receipt is the unit your whole checklist is built from.

Mint your first compliance-grade receipt.