1. Start with an approved export, not a pile of live invoices
Use this guide to check a proposed invoice table, not to approve bills or send money. Obtain permission for the records you use and remove names, account numbers and unrelated private details that the calculation does not need. OpenAI's data-analysis guide supports spreadsheet and structured file analysis. For a first test, use a small fictional CSV with one row per invoice.
2. Give the table a checkable structure
Use columns for invoice ID, vendor alias, invoice date, due date, currency, original amount, recorded payments, open amount and dispute status. Keep currencies separate. Specify the as-of date and whether your approved policy treats an invoice due today as current. Do not let the assistant invent that policy. If you only have scanned invoices, text extraction can miss rows, digits or visual information; reconcile the extracted table back to the original before calculating.
3. Run this fictional three-row test
As of September 30, 2026, use USD only: A is $500 due September 20 with $200 recorded paid; B is $400 due October 5 with no payment; C is $100 due September 10 and disputed. Ask ChatGPT to compute open balances and days past due, retaining disputed status. The expected total is $800: A $300, B $400 and C $100. A is 10 days overdue, B is not yet due, and C is 20 days overdue but still disputed. A dispute does not erase the balance.
Analyze this fictional table only. Recompute open amount as original amount minus recorded payments. Use September 30, 2026 as the as-of date. Show a row-by-row calculation, separate not-yet-due items, preserve disputes and flag missing data. Do not infer a payment approval or contact a vendor.
4. Check the analysis, not only the answer
Ask for the calculation steps or code and inspect them when the interface offers analysis details. Check the dates, amount types, subtraction and treatment of missing values. OpenAI says generated analysis can be inspected; a plausible chart is not evidence that the uploaded rows were interpreted correctly. Verify the three expected open amounts and the $800 total against the source table.
5. Add failure cases before real data
Repeat A's identifier on a fourth row: it should be flagged for review, not silently treated as another legitimate bill. Delete B's due date: B's aging becomes unknown rather than a guessed date. Change C to EUR: the result should show a separate currency subtotal, not add dollars and euros. A negative open balance should be an exception requiring source review, not automatically a vendor credit you can spend.
6. Review the real output with the responsible owner
If the test passes and account/privacy rules allow the real export, rerun using the approved as-of date. Request a source-row reference for each exception and compare totals with the bookkeeping system. Keep the source export version and extraction date. This is an analysis aid, not a substitute for an accountant's review, a payment instruction or proof that all bills were imported.
File and account limits can change
OpenAI's upload documentation sets limits by file type and account and notes that document-image handling varies by feature and plan. Check the current upload rules for your account. If an upload fails, do not assume the file has been analyzed. Keep confidential records out of public/shared conversations, and verify your organization's settings before using client ledgers.
Checked September 30, 2026. Documentation-checked steps and a fictional arithmetic test, not a live private-ledger test. Our Search Console showed 10 impressions for "invoiced to chatgpt" during Sep 12-27; no fresh independent monthly-volume result was obtained. Sources: OpenAI data analysis and OpenAI File Uploads FAQ. For reusable review instructions see our Accounts Payable Aging Review. AI Agent Skills is independent of OpenAI.