My spreadsheet AI demo needed a second prompt. I left it in.
I recorded an invoice-checking demo for GetSheetAI, the Excel and Google Sheets tool I’m building. The workbook had 3,000 synthetic invoices.
The first request was simple: find duplicate invoices and incorrect totals, put them on a new sheet, leave the original alone.
It found the records, but the output still needed work. Dates looked like Excel serial numbers, and “duplicate” lumped together exact copies and invoices with different amounts.
So I asked a second time: format the dates, separate those cases, and show the expected totals and differences. Much more useful.
Not quite the “one magical prompt” demo I had in my head. Apparently even my demo needed a follow-up meeting.
I kept both requests in the recording. The result was 19 records worth reviewing—not 19 proven bad payments.
For anyone else building AI tools: do you show that back-and-forth in demos, or does it lose people before they see the result?
The short demo: https://youtube.com/shorts/J-ajXdq5kVo
(Synthetic data, real Excel recording, AI presenter and narration.)
Keep it. The second prompt proves that the outcome is reviewable instead of magically certain. I’d frame the demo around a two-step “find → verify” flow: first produce candidates, then turn candidates into a decision sheet with a clear distinction between exact duplicates and anomalies. The metric to show is not prompts used but minutes saved and false positives avoided. A shorter polished version can lead; a longer uncut version can sit beside it for buyers who need to trust the workflow.