Most used vs best studied

We ranked work activities by how often people bring them to AI, then checked whether the studies in our ratings test each one directly. Most of that work has no direct study in our ratings, or too few to rate.

• Points to a gain
• No clear effect or mixed
• Early signals only
• Too few studies to rate
• No direct study in our ratings

• It does not mean AI fails at this work. Too little evidence is different from bad evidence. Plenty of this work may well benefit; the studies have not settled it.
• It does not mean no study exists. We count only studies where people worked with AI on real or realistic work, with a comparison. Many others test AI on its own or test members of the public. Each page lists the ones we set aside and why. Our search could also have missed some.
• It is not a judgement on any tool. The usage data comes from two products, but the ratings come from studies of many tools.

• Usage. Anthropic Economic Index (Anthropic, CC BY): Claude.ai, April and May 2026, work conversations only, worked out from each activity’s share and its share of work use. Working with AI (Microsoft Research): Bing Copilot, January to September 2024, US, by what users were trying to do. It has no work-only cut. We never merge the two: they classify conversations differently and are years apart.
• Mapping. Two coders each matched 55 activities to our use cases: our editorial pass, which had seen the ratings, and an independent one, which saw neither the ratings nor the first answers. They gave the same label on 43 and agreed whether any use case applies on 51. “Direct” means the studies test essentially this activity. Both coders are AI models, as are the agents that read and checked the papers, working to our published rules. No person outside the team has reviewed this.
• Ratings. Computed by published rules, from studies graded on 5 November 2026 or earlier. Where an activity links to several use cases, the strongest rating counts, which favours the evidence.
• Stress tests. “At most one of 30”, and both percentage ceilings, hold if we drop our downgrade for older AI models, and if we also count the lab and public studies we set aside. Counting looser matches raises the share a lot, and the size then depends on who did the mapping, so we publish no figure for it.
• Limits. Both usage sources are vendors’ own products with their own users. Some papers could not be opened, including a police report quality study from the same research group as the trial above. We have not yet run every search in our protocol for these activities.
• Disclosure. Anthropic makes Claude, one of the two usage sources, and this site uses Claude in its research and writing. No AI company paid for this analysis or had any say in it.