Business workflow guide
ChatGPT dots spreadsheet review: check before changing data
AI-assisted guidance prepared for Blue Canvas from the official sources linked below. The examples are proposed workflows, not reports of firsthand product testing.
Start a spreadsheet task with an exceptions report, not an instruction to clean everything automatically. Agree what each column means, preserve the original and ask for the row references behind every finding. That gives the reviewer something concrete to accept or reject.
Decide what the review should answer
OpenAI lists data analysis among a dot's capabilities. That does not establish that your particular workbook, formula, macro or spreadsheet application is supported. Confirm what the task can read before relying on its result.
A manageable first question is 'which rows in this fictional equipment list need attention before we use it?' Choose a harmless sample with a few known problems. Avoid payroll, customer records and business-critical calculations for the first exercise. Do not upload live records merely to test whether a connection works.
Name the permitted input and output separately: read the sample file, then return a proposed exceptions table in the conversation. The task should not overwrite a shared spreadsheet, remove rows or repair formulas. Those are different decisions with different consequences.
Product source: OpenAI's official documentation, checked 31 October 2026.
Write a small data dictionary first
Column labels rarely explain all the rules. 'Quantity' might mean items available, items ordered or individual packs. A blank can mean unknown rather than zero. Give the reviewer the intended meaning, accepted format and exceptions before asking it to judge the data.
For our fictional equipment list, the agreed rule is one row per item ID, a non-negative whole-number quantity and a location chosen from an approved list. These are example business rules, not universal spreadsheet requirements. A repeated item ID is an exception under this rule, not automatic evidence that the second row should be deleted.
| Column | Example rule | Do not assume |
|---|---|---|
| Item ID | Required and unique in this sample | Similar names identify the same item |
| Quantity | Whole number, zero or greater | A blank quantity means zero |
| Location | One of the supplied location names | An unfamiliar name is a spelling error |
| Last checked | A date in the agreed format | An ambiguous date can be silently converted |
Build a test with answers you already know
Create a small synthetic sheet with one repeated item ID, one missing quantity and one location outside the approved list. Keep a separate expected-findings list. Include at least one unusual but valid row so you can see whether the review invents an error when the rules are actually satisfied.
Ask for a stable row identifier alongside the visible row number. Sorting can move a record, and row numbers alone can become misleading. Keep the input unchanged during the test and record which sheet or export was reviewed.
Compare the reported findings with your expected list. Check both missed problems and false alarms. An accurate count of three exceptions is insufficient if the report names the wrong three rows. Inspect the actual values and the stated rule for each finding.
Ask for exceptions, not silent corrections
Keep observations and proposals in different columns. 'Quantity is blank' describes the input. 'Ask the stock owner to confirm it' is a proposed next step. 'Set quantity to zero' would alter the meaning and needs evidence you may not have.
The output should also say which checks could not run. If only an exported table was read, do not present the result as an audit of hidden sheets, formulas or workbook behaviour. Ask for the exact scope examined and retain the original for comparison.
Adapt this example to your approved task
Review [synthetic file and sheet] without changing it. Use only [data dictionary and approved rules]. Return each exception with item ID, row reference, observed value, rule, reason and proposed next step. Separate confirmed exceptions from uncertain cases. Report input rows examined and anything you could not inspect. Do not fill blanks, delete duplicates, change formulas, edit shared files or send the result elsewhere. Stop if the intended input or meaning of a column is unclear.
Check access and repeatability before real use
OpenAI says access through a plugin depends on its account, permissions and execution environment. Local work also depends on the connected computer being available with the ChatGPT app open. A file existing somewhere in your business does not establish that this task can read it.
For a second test, change one known sample value and ask for a fresh review of that named revision. Check that the report reflects the change rather than repeating the first result. Keep the two test reports distinct so the differences are inspectable.
Only then consider a bounded business sample approved by its owner. Agree who can review exceptions and who can authorise corrections. This process is not a financial audit, a security assessment or proof that every possible spreadsheet defect will be found.
Product source: OpenAI's official documentation, checked 31 October 2026.