My AI says the spreadsheet is finished. How do I check the work behind it?
Check the job record behind the file, not the file: the sources the data came from with their check states, the transformation steps in order, and the validation steps with their results. Review the receipt and actual deliverable before accepting a handover.
Spreadsheets are where plausible numbers go to be believed. The file in front of you contains results and formulas; it does not contain the decisions that matter: which extract was used, which rows were dropped and why, which currency conversion, which version of the price list. Asking the assistant "is this right" returns a summary of its own work, which is the one witness with an interest in the answer.
The recorded job gives the checker real material. Source steps name the inputs with their states, so a stale extract or an unchecked source is visible before the numbers are. Transformation steps show the pipeline in order, each committed as it ran, so the dropped-rows step and the deduplication step either exist or conspicuously do not. Validation steps carry their own results: totals tied to the source, spot checks against known figures, the reconciliation that either happened or was skipped.
Checking then becomes sampling with a map: recompute one column from its recorded sources, trace one surprising figure back through its steps, confirm the validation steps ran and passed. Fifteen minutes of that finds more than an afternoon of staring at the grid, because the grid was never where the errors lived.
Checklist: auditing the work behind a spreadsheet
1. Sources: every input named on the record, with its check state; any "not checked" source downgrades the whole file. 2. Transformations: steps in order; look specifically for the filtering, joining, and deduplication steps and their stated rules. 3. Validations: tie-out steps with results (totals reconciled to source, known figures matched). 4. Sample: recompute one column and trace one surprising cell to its steps. 5. Gaps: any stage with no step is unverified; require it rerun recorded or accept the risk in writing. 6. Sign-off: record the check on the job when it passes.
Worked example
The source-state discipline on a live record: https://zambo.dev/handover/48479a33-d996-4419-a9c8-d1b9a55e62b9 carries sources as {1 resolved, 1 not_resolved, 1 not_checked}. Applied to a spreadsheet job, that line alone tells the checker where to dig before trusting a single total: the not_checked input is the audit's first address.
Frequently asked questions
The spreadsheet came from a chat with no record. What now?
Treat the file as unverified and choose by stakes: low stakes, sample the numbers manually against sources you trust; high stakes, have the job rerun as a recorded job so the checking material exists. What you cannot do is verify a history that was never saved.
Which single check catches the most errors?
The tie-out: workbook totals reconciled to the source system's totals by a validation step, with the result saved. Most spreadsheet disasters are quiet omissions (missing rows, stale extracts), and a tie-out is designed to hear exactly that silence.
Should the assistant's explanations in chat count as the record?
No. Chat explanations are narration, produced after the fact and editable by regeneration. The record is the steps saved while the work ran, with receipts a checker can recompute without asking the assistant anything.
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