Most AI use is augmentation. Design for it.
Anthropic’s first Economic Index found more AI use augments people than automates them. So AI products should optimize for time and margin.
Most AI use today augments people rather than replacing them. That’s the headline of Anthropic’s first Economic Index, published February 10, and it should change what AI products are designed to optimize.
What did the Economic Index find?
Anthropic classified about a million Claude conversations, with a privacy-preserving tool so no person read them, and mapped them to the tasks in the US government’s O*NET occupation database. It found:
- 57% of use was augmentation, where AI works with a person: checking their work, iterating on a draft, helping them learn. 43% was automation, where AI performs the task directly.
- Software work was the largest share by far, at 37.2% of conversations. Arts, design and media came to 10.3%.
- About 36% of occupations used AI for at least a quarter of their tasks. Only about 4% used it for three quarters or more.
The paper has the method and its limits: it covers one assistant’s free and paid consumer users, not business or API use.
Why it matters for product design
A lot of AI product design assumes the goal is to remove the person. The demo shows a task done end to end, untouched. The metric is how much was automated.
If most use is augmentation, that’s the wrong scoreboard. The person isn’t leaving. What they want back is time, margin for error and the confidence to make the call themselves.
So the questions change. Not “how much of the task did the model do?” but:
- How much time did the person get back, end to end, including checking the output?
- Did they make fewer mistakes, or just faster ones?
- Do they understand the result well enough to defend it?
- Are they better at the job a month from now, or more dependent?
The pattern isn’t new
A tool I designed at Uber worked this way long before anyone called it AI. Before Uber Studio, testing a trip meant phones on desks, personal accounts and real cars driving real blocks to reproduce a bug. Studio let an engineer spawn a rider and a driver and walk a trip through its real states from a chair. It didn’t decide whether their change was right. It gave them the afternoon back, and a way to see the failure before a driver did.
That’s the job for most AI products right now: hand people back their hours and make them surer of the call, not take the call away.