What's this about?
A study of how sophistication in generative AI use varies across the back-office workforce of a large firm, using proprietary data covering 713,564 employee prompts and the corresponding model responses.
Why is prompt data hard to get?
Because it is the most sensitive corpus a company has — it contains what employees actually asked, in their own words, including everything they didn't know and everything confidential they pasted in. Almost all research on AI use relies on surveys, which measure what people say they do. This measures what they did, at a scale that survey work cannot reach.
Why does this matter to you?
- Field evidence from a real firm is rare enough to be worth reading even if the conclusions are unsurprising.
- "Sophistication varies" is the finding most organisations under-plan for — deployment assumes a uniform user who doesn't exist.
- Pairs with the Bocconi study in Lab Updates, which found training and tool access produce different, separable effects.