Where do the numbers in an AI-written case study come from?
Last updated: October 25, 2026
What makes a case study convincing is usually its numbers: time saved, sales up, costs down. That is also where the risk lies with AI-written text. Language models tend to fill missing information with plausible-sounding figures. In B2B, a single invented number makes the whole story look suspect. This post covers what we do about that risk and what you should check before publishing.
In the interview: asking the customer for the number
The best cure for an invented number is a real one. So most of the work happens in the interview:
- If the customer gives a general answer like "we saved a lot of time", the interviewer politely asks for a concrete example or a rough figure in the next question.
- It asks one question at a time. Combining two questions leads people to skip one.
- It doesn't build questions on assumptions the customer never stated. Instead of "How much did your sales grow?", it first asks whether anything changed.
- It doesn't push someone who'd rather not answer. Some figures are confidential, and that's fine.
You can also give the interviewer your own instruction, such as "always ask for a number in the result question".
In the writing: only figures that were said
The rule for writing the story is simple: a figure the customer didn't say in the interview doesn't go into the story. If the customer gave no number, the text doesn't estimate one or add an "approximately". It produces a story without numbers that is still usable, built on concrete description.
A story without numbers can look weaker at first. But a concrete line like "the team no longer builds the report by hand every Monday" is more believable than a "40% efficiency gain" with no source.
Before publishing: the final check is human
No safeguard removes errors entirely; AI can misunderstand too. That's why the last two checks stay with people:
- You: the interview transcript sits next to the story. You can compare every figure with the sentence in the transcript and edit the text and the highlighted figures.
- Your customer: they read the story as it will be published and approve it or ask for changes. We cover this step in the approval post.
A checklist for your own stories
Whatever tool you use, ask these questions about every figure before publishing:
- Who said this number, and in which sentence?
- What time frame does it cover? ("In the first three months" and "in a year" aren't the same.)
- What is it compared against? The previous period, or another tool?
- Did rounding change its meaning? ("About 30%" and "30%" are different claims.)
- Has the customer approved publishing this figure?
For examples of questions that ask for numbers, see the customer interview questions guide.