I am an operations manager. How do I review AI work without reading every chat?
Review records in three passes: scan the period's job summaries and step states for gaps and holds, sample a few receipts at the verify page so the team knows checking is real, and open the underlying detail only where a record raises a question. Chats become the exception you consult, not the pile you read.
Chat-reading fails as review for two reasons: volume, and the fact that a chat shows what was said, not what was done. Operations managers who review from records invert both. The record layer already summarises every job into steps with states; a week's work across a team is a scan of summaries, with the anomalies visible as structure: steps missing, jobs held, handovers unaccepted past their age.
Sampling is the discipline that keeps the scan honest. Pick a small number of jobs per reviewer per week, open their receipts at https://zambo.dev/verify/ , and follow one step end to end. Sampling works because it is unpredictable and cheap; a team that knows any job may be opened keeps every job openable, which is the entire control. Findings from samples, a step class that is always vague, a source state always unchecked, become process fixes, which is what operations review is for.
Reserve chat-reading for the flagged few: a job whose record raises a question the record cannot answer. Even then, read with the record beside you, because the chat's job in that meeting is to explain a specific gap, not to substitute for the accounts.
Template: the weekly record review
Pass 1, scan (15 minutes): the period's job list; note missing steps, held jobs, handovers awaiting decision past the agreed age. Pass 2, sample (15 minutes): three jobs chosen at the moment, one step each followed from summary to receipt at the verify page. Pass 3, exceptions: open detail only on flagged jobs; record each finding as a process fix with an owner. Standing outputs: the exception list, the sample results, and one process improvement per week. Rule: a review that ends with "read the chats" has failed; a review ends with findings about records.
Worked example
A Pass 1 scan meets records like https://zambo.dev/handover/bd7e5325-33fd-4c5e-b029-76b781fa2f51 , five steps, acceptance pending, saved over a short window on 2026-10-16: exactly the kind of item the scan exists to surface (open, ageing, decision owed) without reading a single conversation. The manager's action is a chase on the decision, not a literature review.
Frequently asked questions
How many jobs should a sample cover to mean anything?
Three per reviewer per week is enough to change behaviour, because its power is unpredictability, not coverage. If samples keep finding the same defect, that defect is a process problem; widen the sample on that step class until it is fixed.
What do I do with work that has no record to review?
Treat its absence as the finding. Unrecorded work is unreviewable work, and the review's output for it is a routing fix, not a chat-reading marathon. The standard this file repeats applies: client and consequential work runs recorded, or it is not done.
Can the scan itself be delegated to an agent?
The mechanics can: an agent can list the period's jobs, flag states, and draft the exception list. The sampling choices and the findings stay human, because a review whose surprises are pre-computed by the reviewed system is theatre.
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