If an AI agent acts and writes its own log, who is the record for?

Ashish S K asked a question on LinkedIn that cuts to the point: if an AI agent can take an action and write its own audit record of that action, who is the record actually evidence for? ^[1]

It's a fair question, and it gets sharper the moment you imagine the action going wrong.

Why it matters

AI agents are already doing things: booking meetings, answering customers, placing orders, updating records. The more autonomous they get, the more often someone will ask what the agent did, on whose instruction and when.

The agent's own log is the obvious place to look. The trouble is, the log is kept by the same organisation that ran the agent. If there's a dispute—say a customer claims the agent charged them twice, or changed an order they didn't approve—the operator's own record is the first thing that gets questioned.

It's like the banknote counter at the bureau de change keeping its own ledger. Useful for the counter, less convincing for the person who just handed over cash.

The Article 12 question

Under the AI Act, high-risk AI systems must technically allow for the automatic recording of events over their lifetime. ^[2] Article 12 says what to log. It doesn't say how to protect the log, or who gets to see it.

For high-risk systems under Annex III, the requirement applies from 2 December 2027. ^[2] For systems under Annex I, it's 2 August 2028. ^[2]

So here's the question for anyone building or running an AI agent: when the law asks you for the logs, and a customer or a supervisor asks whether the logs have been touched, what do you show them?

What this means for you

If you're building AI agents, or letting them loose on customer-facing tasks, three questions are worth asking now:

1. Who keeps the record? If the agent logs to a database your team controls, that's fine for operations. It's less fine as evidence the day someone challenges what the agent did.
2. Can the record be checked? A log you can prove hasn't changed is stronger than one you simply assert is correct.
3. How reversible is the action? Booking a meeting is easy to undo. Instructing a payment, changing a contract term or placing a medical order is not. The harder the action is to reverse, the more the audit trail matters.

The first step is deciding what to log: the instruction the agent received, the tool it called, the result it got and the time. The second is deciding how to make that record hold up when it's questioned later.

Separating action from record

One way to answer Ashish's question is to separate the actor from the witness.

Here, the thing to seal is each step the agent takes: the instruction it received, the tool call it made, the result and the timestamp. The fingerprint of that record goes to a neutral third party—in this case, WitniumChain ^[3]—at the moment the step happens. The content of the instruction stays with you.

Anyone you share the record with can later check that it's the same one that was sealed, by calculating the fingerprint again and comparing it. Change a single character in the record and the fingerprint won't match.

What it shows is that the record existed at that time and hasn't changed since. What it doesn't show is that the agent's action was authorised, correct or lawful. Those are separate questions, and they still need answering. But at least the argument starts from an agreed version of what happened, rather than whose record to trust.

How it would work

1. Your agent performs an action and writes a log entry.
2. Your system sends a fingerprint of that entry to WitniumChain through the MCP server and gets a witness ID back.
3. When the action is questioned, you share the log entry and the witness ID. The recipient calculates the fingerprint and checks it against the sealed one.

What to watch

Article 12 of the AI Act sets the logging requirement for high-risk systems: 2 December 2027 for Annex III and 2 August 2028 for Annex I. ^[2] It doesn't say how to protect the logs, but that question is already coming up in tenders and audits.

For AI agents, the question is simpler and harder at the same time: when someone asks what your agent did, whose version of events will they believe?