AI audit trail: inspect actions and results
LLM tool call logging: one record per call
Logging an LLM tool call means saving, at call time, the tool name, recorded inputs, the output the system observed, and the time, with a digest that makes a later change fail its public saved-byte comparison. That record is a verifiable receipt, and the ordered receipts form an AI audit trail.
A real receipt you can check
Receipt 91e376b1-37f4-4c5b-a93e-5497da7410ce is a Zambo MCP prompt_shield receipt, saved 2026-10-18T11:36:51.819Z. It is not a CrewAI or smolagents run.
The public receipt page shows the observed result and saved time. Its machine verification response identifies the tool, recorded request context, canonical saved bytes, and output_hash digest. Recompute SHA-256 over the decoded saved bytes and compare that digest with the stored value. Check this receipt or read the machine check.
- Tool
- prompt_shield
- Recorded request
- Inspect the request context preserved by the public receipt. Do not infer missing original inputs from a summary.
- Observed output
- Open the saved prompt_shield result. This is the system's recorded result, not independent validation of the answer.
- Saved time
- 2026-10-18T11:36:51.819Z
- Digest over the saved result
sha256:fc9e6ff14eff679c81abadf439a907ef594dad2050e44c734b48fcd2e9b94cdf
Receipt 5604e79b-bf43-46fb-b4c5-243ecfaed94b is a second prompt_shield tool record, saved 2026-10-18T11:10:55.432Z. The saved receipt does not establish whether a scheduled process initiated it, so it is not presented as a heartbeat. Check this receipt.
Read records per call, per day in Today, and per job in handovers. For the client-facing session-level view, read AI audit trail: show anyone what your AI actually did. A receipt proves the saved result was not changed since it was saved. It never proves the answer was right and never proves an external event happened.
An AI audit trail records observed actions and results, including failed or blocked steps, for later review.
AI agent audit trail definition
An AI audit trail is a record of observed actions, their inputs, and their results. Verifiable receipts help a reviewer inspect those records; a handover collects them into one client proof link.
An AI agent audit trail is a chain of execution receipts: one verifiable receipt per tool call, linked in the order the calls ran.
To audit an AI agent, open each receipt in the trail, check the tool and arguments, and recompute the SHA-256 fingerprint against the observed result.
An audit trail answers what an agent did and what the system recorded at each execution boundary. A receipt is the unit of evidence. The trail links those receipts in order so a reviewer can inspect a multi-step job without reconstructing it from chat history.
What belongs in an AI agent audit trail
Record the tool, receipt ID, timestamp, session or owner context, and status for each call.
Preserve the arguments sent to the tool and the result the implementation observed.
Keep canonical bytes or the declared representation and a reproducible SHA-256 fingerprint.
Distinguish executed work, gateway observations, and reports received from another agent.
Compliance reference
EU AI Act Article 12 automatic event logging obligations for high-risk systems apply from 2 August 2026. This page cites that provision for orientation and is not legal advice.
Live receipt examples
Trail entry: live_priceOpen the public receipt used as one record in a verifiable trail. Trail entry: capability_searchOpen a second receipt and compare its tool and fingerprint fields.How to review a trail
- Open the first receipt and confirm the session or job identity.
- Follow each receipt link in the order the calls ran.
- Check tool names, arguments, statuses, and observed results.
- Recompute each SHA-256 fingerprint over the stated representation.
- Keep unavailable, failed, or blocked states visible in the review.
Related guides
Read the related AI agent audit trail guide, tools that prove what an AI agent did, the verifiable receipt for AI tool calls guide, the execution receipt definition, and the AER-1 open draft.
Definition and attribution
An AI Agent Execution Receipt records one tool execution and what the execution system observed. This definition is by its creator, Brennan Zambo.
What should an AI agent audit trail contain?
An audit trail should identify each execution, tool, timestamp, caller scope, observed result, status, canonical integrity material, provenance, and any real external evidence. A receipt without read-back evidence proves invocation, not effect. Keep proposed, executed, failed, and confirmed states distinct.
Map the fields to the review question: identity for scope, timestamps for ordering, canonical bytes and hashes for integrity, evidence references for upstream confirmation, and redaction rules for privacy. This is informational guidance, not legal advice.
Verification guides
Tools that prove what an AI agent did · Verifiable receipt for AI tool calls · How to verify AI agent work
Keep going with Zambo in your AI
Get receipts for your own AI work. Add one link, then ask in plain English.
Add Zambo to your AITry a handover
A handover gives your client one proof link with verifiable receipts to review, then Accept or Flag.
try it: say 'use zambo to hand this over to <client>' in your ai
What this evidence means
A receipt shows whether the recorded result matches its saved fingerprint and has not changed after the fact. It does not prove the answer is right or authenticate the issuer's identity. AER-1 is an open draft spec.
Explore related work
Questions and answers
What is LLM tool call logging?
It is the call-time record of a tool, its recorded inputs, observed output, time and saved-byte digest. Zambo tool calls return public verifiable receipts. A matching receipt is not proof the answer was right.
Is there a way to audit LLM tool calls?
Yes, for calls that left records. Open each receipt, check the tool and observed output, recompute its saved-byte digest, and separately check important outside outcomes. Today and handovers gather Zambo records at day and job levels.
Add Zambo to your AI
Connect the free MCP at https://zambo.dev/mcp in Claude, ChatGPT, Gemini, Cursor, Grok, Windsurf, or VS Code. No key is needed. The first receipt arrives with the first call through Zambo. Adding a URL does not automatically install a connector; use your client's MCP settings.
More AI agent answers · Try the demo · Current pricing and tool-specific limits