What does the evidence say about AI and jobs so far?
What the evidence says so far, especially for people starting out.
Automation: the AI does the task for you. "Write the monthly report."
Augmentation: the AI helps you do the task. "Here's my draft. What's missing?"
The difference matters a lot for what happens to jobs. One replaces the work, the other makes the person doing it faster or better.
Anthropic tracks how people use Claude. Between December 2024 and August 2025, the share of conversations where people handed over a whole task (the automation kind) rose from 27% to 39%.
Among businesses connecting to Claude directly through its API, about 77% of use looked like automation.
A Stanford team (Erik Brynjolfsson, Bharat Chandar and Ruyu Chen) studied payroll records from ADP, one of the biggest payroll companies in the US.
AI-exposed job: a job where AI can do many of the tasks, such as customer service or software development.
In their August 2026 update, employment for 22 to 25-year-olds in AI-exposed jobs was down 19% compared with less-exposed jobs. In their first version, in 2025, it was 13%.
The early effect is at the bottom of the ladder.
Junior work is often the most clearly defined: summarise this, draft that, check this list, write the first version.
That's exactly the kind of task AI is good at. Experienced people spend more of their day on judgement, relationships and knowing which problem to solve, which is harder to hand over.
Ethan Mollick, a Wharton professor who writes a lot about AI at work, names four things people bring that models don't easily replace:
Get good at delegating and checking. Briefing the AI well and catching its mistakes is a skill that holds its value.
Use it to learn, not only to finish. Ask it to explain, then do part of the work yourself.
If you manage people, think about how junior staff will build experience if the AI takes their first tasks.