Reasoning traces aren’t audit trails: what the EU AI Act asks your agents to prove by August

What the August 2, 2026 logging obligation actually requires

The EU AI Act’s provisions for high-risk AI systems became applicable on August 2, 2026, according to the European Commission. Among them is a requirement that high-risk systems keep automatic logs of events over their lifetime — the record of what the system did, not merely what it was designed to do. Retention is generally at least six months, rising to 24 months for certain biometric and law-enforcement uses. This is not a documentation nicety. It is a standing obligation to be able to reconstruct system behavior after the fact, on demand.

The stakes are set by the penalty regime. The Commission notes fines can reach up to €35 million or 7% of global annual turnover for the most serious violations. That figure reframes logging from an engineering preference into a board-level exposure. If your agents make consequential decisions and you cannot produce a trustworthy record of them, the gap is no longer theoretical — it is the exact thing the obligation is written to close, and the number attached to failing it is large enough to notice.

Reasoning traces vs. real audit trails

It is easy to conflate the two because both feel like “a record of what happened.” They aren’t. A reasoning trace is the model’s narration — the chain of thought it surfaced on the way to an answer. It is useful for debugging and for understanding intent, but it describes deliberation, not action. As industry guidance from Apptitude puts it plainly, reasoning traces are not audit trails. An audit trail records the actions an agent actually took: the tool it called, the data it read, the decision it committed, the downstream effect — each one attributable and reconstructable.

The distinction matters most when someone challenges a specific decision. A reasoning trace can tell you the agent “decided to approve the transaction because the risk score was low.” An audit trail proves the agent called the risk service, received that score, and executed the approval at a specific time. One is a paraphrase of intent; the other is the record of events. When a regulator or a claimant asks what your agent did, the paraphrase won’t hold — you need the events, captured completely, at the level of every action rather than only the final answer.

The tamper-evidence gap: a mutable log is a claim, not evidence

Say you clear the first bar — you log every action, not just outputs. There is still a second bar, and it is the one most teams miss. A standard application log is editable. Rows can be dropped, timestamps rewritten, entries appended after the fact. That mutability doesn’t make your team dishonest — but it does mean the log can’t prove its own integrity. As Apptitude frames it, “without cryptographic proof, ‘we logged it’ is a claim, not evidence.” The record and the assertion that the record is intact are the same unverifiable thing.

• A mutable log asks the reader to trust your operational controls; a tamper-evident log lets them check.
• Append-only, hash-chained records make any deletion or edit detectable — the chain breaks if the past is altered.
• Independent verifiability means an outside auditor confirms integrity without access to your internal systems or good faith.

Is your audit trail regulator-ready?

Before you plan a fix, find out where you actually stand. The five questions below map to the gaps that matter under the August obligation: completeness of capture, tamper-evidence, independent verifiability, retention, and replay. Score yourself honestly — the bands tell you whether your trail is evidence, merely logs, or not yet audit-ready.

Five steps to evidence-grade records

The distance between where many teams sit and where the obligation points is smaller than it looks — but only if you close the gaps in the right order. One 2026 analysis (reported secondhand, so treat it as directional) found roughly 33% of enterprises running agents had no audit trail at all, and only about 21% had runtime visibility. If those figures are even roughly right, the baseline is thin. The steps below move you from thin logs toward records that hold up when someone checks them, rather than merely reads them.

• Capture every action, not just outputs — tool calls, data reads, and committed decisions, each attributable to a specific agent and moment.
• Make the log append-only and tamper-evident by hash-chaining entries, so any deletion or edit breaks the chain and shows.
• Enable independent verification — let an outside auditor confirm integrity without trusting your internal systems or your word.
• Retain for the required window — at least six months, and 24 months where biometric or law-enforcement uses apply.
• Prove you can replay a single decision end to end, reconstructing exactly what the agent did from the record alone.