Anyone can log what an agent did.
Prove what it didn't.
Disputes are rarely about what your AI agent did. They are about what someone says it did — and an audit log cannot rebut that, because a log with nothing in it is equally consistent with 'it never happened' and 'we deleted it'.
The three allegations a log leaves you defenceless against.
“Your agent kept contacting me after I opted out.”
Certify that no action involving that person occurred in the window. Matching is on digests, so the certificate never contains their phone number or email — it stays shareable with a regulator without disclosing the customer.
“Someone accessed this record.”
Certify that no action touched that object between two dates, with the chain shown to be contiguous across the whole window — no gap where a receipt could have been removed.
“The AI did that on its own.”
Certify that nothing was taken under standing policy or a schedule — that every action in the period was authorised by a named human. This is the claim operators most often have to make to a regulator.
Three conditions, all of which must hold.
- The chain verifies. Every action produces an Ed25519-signed receipt, hash-chained to the one before it. Editing any receipt breaks its signature; the whole chain is re-verified on every read.
- The window is contiguous. Sequence numbers run without gaps. A deleted receipt leaves a hole that every subsequent link exposes — which is the failure a normal log cannot detect at all.
- The period was anchored. Chain heads are published so the past is frozen before any dispute arises. We could still lie; we could no longer lie undetectably, which is the only property worth having.
The certificate refuses far more often than it grants. If no receipt exists after the window closes, it refuses — a record that simply stops inside the period is indistinguishable from one that was truncated. A certificate that is easy to obtain is worth nothing in the argument it exists for.
What it cannot tell you.
An action taken entirely outside Deskwise leaves no receipt, and no certificate can speak to it. That assumption is printed on the certificate rather than buried in a footnote. These receipts are also signed by us — so they prove our record was signed by the holder of the published key and has not been altered, not that the world independently agrees. The published chain head is what turns that from a promise into something checkable.
We say this plainly because an auditor who discovers a limitation themselves discounts everything else you have told them.
What people ask.
Why can't an audit log prove something didn't happen?
Because an empty log is equally consistent with two very different stories: the action never occurred, or the record of it was removed. Nothing about the log itself distinguishes them. Proving a negative needs the record to be provably complete across the period in question, not merely to contain no entries.
What makes a certificate of non-action different?
Three conditions have to hold at once. The receipt chain must verify end to end, so nothing was altered. The sequence must be contiguous across the window, so nothing was deleted from the middle. And the window must be bounded by chain heads published before the dispute arose, so the record could not have been reconstructed afterwards. Miss any one and the certificate is theatre.
What can it not prove?
An action taken entirely outside the system leaves no receipt, and no certificate can speak to it. That completeness assumption is printed on every certificate rather than buried. Deskwise also signs its own receipts, so a party extending no trust at all should rely on the published chain head, which is what makes tampering detectable rather than merely unlikely.
Can an auditor verify it without trusting Deskwise?
Yes. The chain exports with its public key, and the verifier is a standalone file that shares no code with Deskwise and imports only Node's crypto module. An auditor runs it on their own machine against a file we cannot alter after handing it over.
When would a business actually need this?
Opt-out and do-not-contact complaints, access disputes over a record, chargebacks where an agent acted on a customer's behalf, and regulator questions about whether a decision was autonomous. In each, the allegation is of an action that usually never occurred — and a log cannot rebut it.