The AI gap: what AI could do in your job vs what people actually save
Quick answer
In office and administrative jobs, today's AI tools could do about half of the typical tasks (51/100). In the last four quarters, US workers in those jobs saved 1.6% of their work hours with generative AI – the same as factory workers, whose jobs are a third as exposed. In every one of the 22 occupation groups, what AI could do is far ahead of what people use it for. That gap is the opportunity.
Two numbers that are usually confused
Most "will AI take my job" tools – ours included – show one number: an estimate of how much of a job's tasks today's AI could do. That number comes from rating tasks, not from watching people work.
The other number is measured: how much working time people actually save. The best regular source is the St. Louis Fed's Generative AI Adoption Tracker, a quarterly survey of US working-age adults by Alexander Bick, Adam Blandin and David Deming. It reports, for each occupation group, "the share of total work hours saved by employed adults aged 18-64 … due to the use of generative artificial intelligence at work" – counting everyone in the group, users and non-users (FRED).
The two numbers measure different things, so they should not be divided into each other. But putting them side by side shows something no single number does.
Estimated vs measured, by occupation group
Estimated: average task exposure of the occupations in the group in our catalog (0–100). Measured: share of work hours saved, average of Q4 2025 – Q3 2026.
| Occupation group | Estimated (what AI could do) | Measured (hours saved) |
|---|---|---|
| Computer and Mathematical | 55 | 4.2% |
| Business and Financial Operations | 51 | 3.0% |
| Office and Administrative Support | 51 | 1.6% |
| Legal | 47 | 3.2% |
| Life, Physical, and Social Science | 45 | 2.8% |
| Sales and Related | 44 | 1.6% |
| Management | 42 | 3.9% |
| Architecture and Engineering | 41 | 2.7% |
| Educational Instruction and Library | 41 | 2.4% |
| Community and Social Service | 38 | 2.3% |
| Arts, Design, Entertainment, Sports, and Media | 36 | 2.9% |
| Healthcare Practitioners and Technical | 32 | 1.6% |
| Protective Service | 31 | 0.7% |
| Personal Care and Service | 23 | 1.1% |
| Healthcare Support | 22 | 1.2% |
| Transportation and Material Moving | 21 | 1.0% |
| Installation, Maintenance, and Repair | 18 | 1.4% |
| Production | 17 | 1.6% |
| Food Preparation and Serving Related | 16 | 0.8% |
| Building and Grounds Cleaning and Maintenance | 16 | 0.8% |
| Farming, Fishing, and Forestry | 16 | 1.8% |
| Construction and Extraction | 13 | 1.4% |
What stands out
- Office work is highly exposed but barely saves time. Office and administrative jobs are among the most exposed (51), yet their measured savings (1.6%) are no higher than in production jobs (17). Exposure alone does not turn into saved time.
- Managers save the most outside tech. Management is only mid-table on exposure (42) but second on measured savings (3.9%), close to computer and math jobs (4.2%). People who choose how they work seem to use AI more.
- Nobody is near the estimate. The highest measured value is 4.2% of hours. The lowest estimate is 13 out of 100.
Why the gap is so large
Other research points the same way:
- A Danish study of 25,000 workers in 7,000 workplaces found "modest productivity gains (average time savings of 2.8%)" from AI chatbots, and no measurable effect on earnings or hours (Humlum and Vestergaard).
- When Anthropic asked Claude to estimate time savings from real conversations, the estimate was that "Claude speeds up individual tasks by about 80%" – but the authors note this does not include the time people spend "validating the quality or accuracy of Claude's work" (Anthropic).
Put together: on a single task, AI can be fast. Across a working week, most tasks are not done with AI yet, the ones that are still need checking, and many people do not use it at all. The measured number counts all of that.
What this means for you
If you work in an exposed job, the gap is a head start, not a threat. Most of the time AI could save in your job has not been claimed by anyone yet – including your colleagues. The people who learn to use it well for their own largest tasks, and can show the result, will be the ones who close it.
- See how exposed your job is and where the time could come from: the job list.
- Take the free 11-question test to find the first task in your own week.
Method and limits
- Estimated is our catalog: each occupation's tasks from O*NET 29.3, time shares from the Anthropic Economic Index, and an automation rating that blends a Claude rating with the human labels from GPTs are GPTs (Eloundou et al., Science, 2024). The group value is the average of its occupations.
- Measured is self-reported, US only, and counts everyone in the group. Single quarters move a lot because each group is a small sample, so we use the average of four quarters.
- The two are different measures. A ratio between them would be misleading; the comparison is about direction and size, not a precise "realization rate".
Frequently asked questions
How much time does AI save at work?
In the St. Louis Fed's survey, US workers saved between 0.7% and 4.2% of their work hours with generative AI in the last four quarters, depending on the occupation group. A Danish study found 2.8% on average.
Which jobs save the most time with AI?
Computer and mathematical jobs (4.2% of hours), management (3.9%) and legal jobs (3.2%).
Why is the saving so much smaller than the exposure estimate?
The estimate is what AI could do on each task; the measurement is what people actually save across their whole week, including everyone who does not use AI and the time spent checking its output.
Will the gap close?
The measured savings have been rising, but slowly. How fast the gap closes in your own job depends mostly on whether you and the people around you learn to use AI for the tasks where it is strong.