Audit Agent
The difference between a control that operated and one that is asserted to have operated is whether anyone can show it, months later, without rebuilding the story.
Every decision can be re-read later: what was checked, what was found, who released it.The trail is written as the checks run, immutably, so evidence is a record rather than a reconstruction.
What it checks
On every invoice, not a sample.
- Every check that ran against the transaction, and its result
- Who approved, who overrode, and the reason given
- The document version the decision was made on
- That the trail cannot be edited after the fact
What it reads, what it writes
Reads
Every agent's output · Approvals and overrides · Document versions
Writes
An immutable record per transaction · Evidence packs that can be handed to an auditor as they are · Control adherence per step
Where it sits in the team
Agents do not work alone. Each one hands its result to the next, so a finding raised here shows up as context downstream rather than being re-derived.
Hands to it
It triggers
When it isn't sure
Nothing is omitted for being inconvenient. A check that failed and was overridden is recorded exactly as that, including who decided and why — a trail that only holds the clean cases is not a trail.
Any check that runs on every transaction will meet cases it cannot settle. What makes a control trustworthy is not that it never hesitates — it is that hesitation produces a named question for a person, rather than a silent pass or a silent block.
See what this one finds in your last 90 days.
One export. No connector. We come back within a working day.
$Check your savings→