How do I run a Q4 audit of everything my AI agents did this year?
Run five passes in Q4: inventory every agent and connection still live, sample the year's job records and recompute a few receipts, test that consequential outputs carry recorded approvals, reconcile agent spend against statements, and write the gap page, what ran unrecorded, so January starts with fixes, not findings.
A year of agent work feels unauditable because it happened everywhere at once. The records impose the order. Pass one is the inventory: which agents ran this year, through which connections, with which access, and which of those are still live. Anything still connected that nobody owns is finding number one, and it costs nothing to fix in Q4.
Passes two and three sample the substance. Pull a random set of jobs across clients and months; for each, open the record and check the steps, the source states, and one receipt recomputed at https://zambo.dev/verify/ . Then check the approvals: consequential outputs should show a named human decision on the record. A healthy year samples clean; the exceptions, a job with no record, an approval that lived in a chat, become the gap page's entries.
Pass four reconciles money: agent spend by month against statements and the limits in force, with refusals counted as the control working. Pass five is the writing, and it is the pass that pays: the gap page lists what ran unrecorded, by system and period, with the fix and its owner for January. An audit that ends in a page like this is the difference between a team that has records and a team that has a records system.
Checklist: the five-pass Q4 audit
Pass 1, inventory: agents, connections, access, owners; revoke the ownerless. Pass 2, sample: 10 to 20 jobs across the year; steps read, one receipt per job recomputed. Pass 3, approvals: consequential outputs carry recorded decisions; list exceptions. Pass 4, spend: monthly totals against statements; limits and refusals reviewed. Pass 5, gap page: unrecorded systems and periods, each with a fix, an owner, and a January date. Time budget: one afternoon for a small team whose work was recorded; the gap page is longer when it was not, which is the audit's real output.
Worked example
Pass 2's recompute step, done in public: take any job from the year's records, for example the public record at https://zambo.dev/handover/45434a17-4afc-486a-9646-e689bcee4297 , read its steps, and recompute one receipt at the verify page. Ninety seconds per job is the sampling rate a recorded year makes possible.
Frequently asked questions
Our agents ran mostly unrecorded this year. Is the audit pointless?
The audit becomes the gap page, and that page is the most valuable output available to you: a bounded, dated account of what ran without records, which is the document every fix, client conversation, and next-year plan starts from.
How large should the sample be?
Large enough to cross every client, system, and quarter: 10 to 20 jobs for a small team. Sampling is looking for classes of gap, not certifying every job; the classes announce themselves quickly when they exist.
Who should run the audit?
Someone other than the person whose work is sampled, where the team size allows. The records make self-audit possible, but a second reader finds the ownerless connection and the unrecorded system that familiarity skips.
By Brennan Zambo, founder of Zambo.
Last verified: 2026-10-17.
Try a handover
A handover gives your client one proof link with verifiable receipts to review, then Accept or Flag.
try it: say 'use zambo to hand this over to <client>' in your ai