Seats are not usage
The subscription audit has its place: overlapping tools, idle seats, renewal dates. But a seat count cannot answer the questions a manager actually gets asked. What did the AI produce for the Henderson job? Who approved sending it? What did the team spend on agent runs this week, and on what? Which deliverables went out with steps nobody verified?
Those questions are answered by records of work, not records of purchase. If the team's AI work runs through tools that leave receipts, the records already exist. The audit is the discipline of reading them on a schedule instead of after an incident.
The three layers of the record
- Calls. Every tool call leaves a verifiable receipt: the tool, the inputs, the observed result, the time. This is the ground layer, the one disputes and incident reviews are read from.
- Days. The work rolls up by day, so a manager can open one page and see what ran across the team today: Zambo publishes this as the today page, one link per date, instead of a dashboard login nobody remembers.
- Jobs. Client facing work gathers into handovers: the steps, the sources, the decisions, and the client's accept or flag, recorded with its time. This is the layer a client or an auditor ever sees, and it is built from the same receipts.
Spend belongs in the audit, with a cap behind it
Usage without spend is half the picture. The audit should show what the team's agent work cost in the period, by job where the records allow it, next to what the work produced. The stronger move is making spend a control rather than a report: Zambo's spend limit tools let an owner set a daily cap, and calls past the cap are refused, with the refusal itself recorded as a receipt. A cap that fires leaves evidence. A budget that is only reviewed monthly leaves surprises.
The weekly routine
- Open the week's days. Skim the daily roll ups for work you do not recognize. Unrecognized work is the audit's first finding, whatever it turns out to be.
- Sample the jobs. Pick a few client deliverables and read their handovers end to end: steps, sources, the not checked list, the acceptance state.
- Recompute a sample of receipts. Pick calls at random and check their digests. The point is not suspicion. The point is that the team knows the record gets checked, and the record knows it too.
- Review spend against the cap. Refusals, spikes, and the jobs behind them.
- Write the gaps down. Work the team did outside recorded tools, steps nobody checked, clients still sitting on an open acceptance. The list is the agenda for next week, not a confession.
Start with one team's real week
Stand the audit up on live work, not on a policy document. Connect the team's AIs to Zambo at https://zambo.dev/mcp, let a week of real jobs run with receipts under every call and handovers on the client work, then run the routine above on Friday. A $0.99 Day Pass covers a heavy day for one person, and the Zambo Pass at $29 a month is the standing version: private handovers, the approval records as PDF and JSON, and the full record for every job the team sends.
Questions, answered
How do I run a team AI usage audit?
Build it from the work record, in three layers: the receipts under every call, a daily roll up of what ran across the team, and the handovers behind client facing jobs. Then run a weekly routine: skim the days for unrecognized work, read a sample of handovers end to end, recompute a sample of receipt digests, review spend against its cap, and write down the gaps, including work done outside recorded tools. On Zambo the layers exist by default: every call leaves a verifiable receipt, the today page rolls work up by date, and handovers carry the client record.
What is the difference between a usage audit and a subscription audit?
A subscription audit counts what the company pays for: seats, overlapping tools, idle licenses, renewal dates. A usage audit reads what the team's AIs did: which jobs ran, what they produced, what each deliverable rests on, what the work cost, and what nobody checked. The first trims the bill. The second is the one that answers a client, an auditor, or your own incident review when something goes wrong.
What records does a manager actually review in a team AI usage audit?
Three: the daily roll up, to see what ran across the team on a given date; a sample of handovers, to read real jobs end to end, steps, sources, decisions, unchecked items, and the client's recorded accept or flag; and a sample of individual receipts, recomputed to confirm the records are intact. Add the spend record: what the period's agent work cost and whether the cap fired. Anything beyond that is usually reporting about reporting.
How do spend caps fit into a team AI usage audit?
As a control with evidence, not just a budget line. Zambo's spend limit tools let an owner set a daily cap for agent work. Calls that would pass the cap are refused, and the refusal is recorded as a receipt like any other call. The audit then reviews three things: total spend for the period, the cap level against the work the team actually does, and every refusal, because a refusal means real work hit the ceiling and someone should know which work and why.
What does a team AI usage audit not cover?
Any AI work done outside recorded tools: a disconnected app, a personal account, or a chat that never calls a recorded tool leaves no receipt and appears nowhere in the audit. It also does not judge quality. The audit establishes what ran, what it produced, what it cost, and what was never checked, with records a third party can recompute. Whether the work was good is still a management judgment, made honestly only when the unchecked list is on the table next to the deliverables.