AI wrote "Not specified," then filled the field anyway.
Fictional call notes in, eight CRM fields out: Contact name, Job title, Company size, Budget, Decision maker, Competitors considered, Target go-live date and Next step.
Tested in GPT-5.5 (OpenAI API), Oct 2026
Most of us would write this. I ran it on fictional notes. With real ones, remove names, account numbers and personal details first, and check the tool's data policy or your employer's approved tools.
Real output · GPT-5.5 (OpenAI API) · Oct 2026
It wrote "Not specified" in two fields, then filled them anyway. Dana also got a made-up job title, "Wholesale lead."
Each field was filled anyway
The notes only say Dana "runs the wholesale side," so "Wholesale lead" is made up. Vans aren't a company size, and a pricing request isn't a budget. A filter for empty Budget would miss it.
Tested in GPT-5.5 (OpenAI API), Oct 2026
A field asks to be filled, so the model tends to reach for the closest fact. Anthropic's docs suggest telling it to use only your information and letting it say it doesn't know.
Source: Anthropic, Claude Docs: Reduce hallucinations
Real output · GPT-5.5 (OpenAI API) · Oct 2026
With the added line, Company size and Budget now say "Not mentioned," and Job title quotes the notes instead of inventing a title. Both runs did this.
Still check every field
It still isn't clean CRM data: Job title is a description, Decision maker names no person, and the go-live date is "Before the Valentine's rush." Anthropic says these steps reduce made-up answers, not eliminate them.
Try it: Add the line to your next call-notes prompt, then count the Not mentioned fields.
Source: Anthropic, Claude Docs: Reduce hallucinations
Sources and assumptions
- Anthropic, Claude Docs: Reduce hallucinations: 'Allow Claude to say "I don't know": Explicitly give Claude permission to admit uncertainty. This technique can drastically reduce false information.' and 'External knowledge restriction: Explicitly instruct Claude to only use information from provided documents and not its general knowledge.' Also the Note: 'although these techniques significantly reduce hallucinations, they don't eliminate them entirely. Always validate critical information' (checked 2026-10-29)
- Own test: engine/ai-test.js, GPT-5.5 (OpenAI API), record.tests ids before-gpt, after-gpt, before-gpt-run2, after-gpt-run2: Both prompts were run twice on the same fictional notes. Both before runs made up the job title "Wholesale lead" and filled Company size and Budget with van and pricing details. Both after runs wrote "Not mentioned" for Company size and Budget (checked 2026-10-29)
Assumptions:
- The call notes (Dana Whitfield, Larkspur Floral Studio, 2 vans, Valentine's rush) are fictional, written for this test
- Tested in GPT-5.5 through the OpenAI API, Oct 2026, not the ChatGPT app. The Claude runs (ids before, after) failed to execute in the research sandbox, have no output and must not be quoted or named as tested
- The technique source is Anthropic's documentation written for Claude. The post applies the same instruction to GPT-5.5 and shows only GPT-5.5 results
- Two runs per prompt. Results may vary between runs and model versions
- "Unsupported" means not stated in the notes: "Wholesale lead" is a made-up title, and vans or a pricing request are not a company size or a budget
The short version
- Tell AI to use only the notes
- Empty fields should say Not mentioned
- A confident guess hides what you still need to ask
- Check each filled field against the notes


