What should an AI agent audit log record?

An AI agent audit log should record every action the agent takes, not only what it said. Each entry answers six questions: which agent acted, for whom, what it did and in which system, which rule allowed it, who approved it if approval was needed, and how it turned out. Log refused and failed actions too, and keep the log somewhere the agent cannot change it.

A chat transcript is not an audit log. It shows the words, but it does not show that the agent issued a refund in your billing system, which limit applied or who signed off. When a customer disputes a charge, or an auditor asks how a decision was made, the action record is what you need.

What fields should every AI agent audit log entry have?

Record these fields for each action. Tools name them differently; the content should be the same.

Field	What it records	Example
When	Time in UTC, to the second	2026-10-30 14:02:11 UTC
Agent	Which agent, and which version of its instructions	Billing agent, v14
On behalf of	The customer or employee it was serving	Customer 4821
Action	What it did, in which system, on which record	Refund issued, billing, invoice 3307
Inputs	What it based the action on	Ticket BIL-311, refund policy v3
Rule	The rule that allowed, held or blocked it	Refunds over $200 need finance
Approval	Who approved and when, or "within rules"	Alex, Finance, 14:01
Result	Succeeded, failed or refused, with the error	Succeeded, $240
Links	The conversation and the previous step	Conversation C-9921

Two of these get skipped most often. The first is refused actions: a run of blocked attempts is often the earliest sign that an agent is working outside its job. The second is the instruction version, which tells you whether a mistake happened before or after you changed the agent's instructions.

Who reads an AI agent audit log?

• Operations, to see why a booking moved or a ticket closed.
• Finance, to check every payment, refund and discount against its approver.
• Support leads, to see what the agent told a customer and what it did next.
• Security and outside auditors, to review access, denied actions and rule changes.

Each of them should be able to filter by customer, action type and date without asking an engineer. If only the person who built the agent can search the log, it will not get read.

How does an audit log settle a dispute?

Here is a fictional example. A customer of a home services company says the agent promised a full refund for a missed visit. The log shows the conversation, the rule that applied (refunds over $200 need finance), the request sent to the finance lead, her approval of a partial refund, and the message the agent then sent with the amount. The disagreement is over in a few minutes, because everyone can see what was offered, who approved it and under which policy.

It works the other way too. If the agent did promise something outside policy, the log shows which source or instruction led it there, so you fix the cause instead of arguing about the transcript.

What changed recently?

A common format for agent logs is being drafted. "Agent Audit Trail," an Internet-Draft updated on October 26, requires each record to name the agent, the principal it acted for, the action and the outcome, and to carry a hash of the previous record so that edits and gaps show up (IETF Datatracker). It is an individual draft with no formal IETF standing, but its field list makes a good checklist.

The case for logging got sharper as well. On October 16, India's CERT-In said AI model providers had disclosed that their agents reached live systems at Indian organizations and took actions that "exceeded the authorisation of their assigned tasks." The organizations had not detected this themselves (TechCrunch).

Frequently asked questions

What should an AI agent audit log not record?

Passwords, API keys, full card numbers and other secrets. Store a reference or a masked value instead, and keep full message contents in a separate store with tighter access and a shorter retention period.

Can an AI agent edit its own audit log?

It should not be able to. The system around the agent writes the log to append-only storage, so neither the agent nor an administrator can quietly change an entry. Chaining each entry to the one before it makes any edit visible.

How long should you keep AI agent audit logs?

At least as long as the business records the actions touched, such as invoices or appointments, and through your dispute window. In the EU, the AI Act requires high-risk AI systems to allow automatic logging over their whole lifetime (European Commission). Check your industry's rules with counsel.

Next steps

Decide your AI agent audit log fields before the agent goes live, then test them. Pick one action from last week and see whether someone outside the project can explain it from the log alone. Aurora writes every action to one audit log, with the rule that applied and who approved it, and your team can take over any conversation with its full history.

See how Aurora handles rules, approvals and the audit log, write an AI agent approval policy to go with it, read when an agent should hand off to a human and how an AI customer service agent solves a request, or learn what MCP connects on the way to your systems.