I asked AI to chase a late invoice, and it left blanks where the amount should be.
What the Reel says
I asked AI to chase a late invoice, and it left blanks where the amount should be.
I only gave it the client, the job and how late it was.
Claude's email had blanks for the amount, the due date and the invoice number.
GPT-5.5 left the numbers out completely.
Either way, your client has to look the invoice up, and that's one more reason to wait.
But AI can't fill in numbers it can't see: it only knows what you type.
So I pasted in the invoice and told them to repeat each figure exactly.
This time, both wrote INV-0142, $1,850.00 and October 31, 2026, exactly right.
AI works with what you hand it, so the real invoice gets you a reminder your client can actually pay.
Sources and assumptions
- Anthropic: Prompting best practices (Claude docs): 'Be clear and direct': 'Claude responds well to clear, explicit instructions. Being specific about the output you want can help enhance results.' and 'Think of Claude as a brilliant but new employee who lacks context on your norms and workflows. The more precisely you explain what you want, the better the result.' (checked 2026-11-13)
- Xero: Chasing outstanding invoices (US guide): A payment reminder should 'Include the invoice number, amount owed, original due date, and a clear request for payment.' (checked 2026-11-13)
- QuickBooks: Past-due invoices guide: A past-due reminder letter should include the invoice number, issue date, due date, payment terms, amount owed including any late fees, payment instructions, and your phone number and contact information. (checked 2026-11-13)
- Our test: Claude Sonnet 5.5 and GPT-5.5 (OpenAI API), run 2026-11-13 Pacific (2026-11-14 UTC): Real outputs in record.tests: before-claude, before-chatgpt, after-claude, after-chatgpt (the URL is a placeholder required by the schema; the test itself is in tests.json) (checked 2026-11-13)
Assumptions:
- The invoice, freelancer (Sam Ortiz Design), client (Dana Reyes, Harbor Lane Bakery), amount ($1,850.00) and dates are fictional sample material.
- 'Today' in the test is November 13, 2026, so the invoice due October 31, 2026 is 13 days overdue; the vague prompt said 'about two weeks late'.
- Each prompt was run once per model. Another run may word things differently, so the post shows these outputs as one real test, not a guarantee.
- The models were run through Claude (claude -p, Sonnet 5.5) and the OpenAI API (GPT-5.5), not the ChatGPT or Claude apps.


