CORE IDEA: Generative AI has reached almost every corner of the labour market, but hardly anyone uses it for most of their work. It is spreading wide rather than deep. What now decides whether AI is used is not the task. It is the person doing it. |
For three years the debate about artificial intelligence has been about capability. What can the models do? Which jobs are exposed? Which tasks could be automated? A new working paper from the Federal Reserve Bank of St. Louis asks a simpler question. Not what generative AI might do, but what work people are actually using it to do.
That difference matters. Exposure studies show where AI could be useful. Chat logs show what today’s users ask for. Neither shows how many people doing the same work have never started. This study does, and its answer is both encouraging and uncomfortable: adoption is wide but shallow.
What the research measures
The researchers surveyed a representative sample of American workers across four waves, from August 2025 to May 2026. Each person listed the tasks they do regularly and the tasks for which they use generative AI. The result shows real use, not theoretical exposure.
By May 2026, 45% of workers used generative AI for work and 55% of adults aged 18 to 64 used it outside work. In total, 62% used it somewhere. But the headline numbers hide the real picture: in most occupations and most tasks, fewer than half of the people involved use the technology at all.
This confirms an argument we have made throughout this series: access is not adoption. A licence, an approved tool or a capable model changes nothing on its own. People need experience, confidence and a reason to work differently.
Where adoption is highest and lowest
The pattern is consistent: AI is used where work is made of information, and avoided where work is made of physical presence. Computing and professional roles sit at the top; personal services and manual trades sit at the bottom. The same split appears task by task. Reading, research, analysis and planning lead. Driving, cleaning, clinical care and equipment work trail.
Highest-adoption occupations
Rank | Occupation | Rate |
|---|---|---|
1 | Computer and information research scientists | 87.3% |
2 | Information security analysts | 85.4% |
3 | Network and computer systems administrators | 82.4% |
4 | Computer programmers | 81.2% |
5 | Public relations specialists | 80.4% |
6 | Personal financial advisors | 79.7% |
7 | Chief executives | 79.7% |
8 | Software developers | 78.8% |
9 | Market research analysts and marketing specialists | 78.7% |
10 | Computer and information systems managers | 78.2% |
Adoption rate: the share of people in that occupation or task who report using generative AI regularly. The same definition applies to all four tables.
Lowest-adoption occupations
Rank | Occupation | Rate |
|---|---|---|
1 | Stockers and order fillers | 14.8% |
2 | Janitors and building cleaners | 13.8% |
3 | Other installation, maintenance, and repair workers | 13.1% |
4 | Maids and housekeeping cleaners | 11.3% |
5 | Fast food and counter workers | 11.3% |
6 | Maintenance and repair workers, general | 11.2% |
7 | Painters and paperhangers | 10.4% |
8 | Licensed practical and licensed vocational nurses | 10.4% |
9 | Receptionists and information clerks | 7.6% |
10 | Animal caretakers | 5.3% |
The field matters, but it does not decide
It is tempting to see these rankings as a simple split between office jobs and manual jobs. But the study shows that this is only part of the story. The type of occupation explains less than 20% of the difference in AI adoption, and the task itself explains only about 11%. Most of the difference—around 70%—comes from people doing similar work but choosing to use AI differently.
This is the true adoption gap: two people can hold the same job, tackle the same task, and have access to the same tools, yet choose entirely different approaches. Skill, personal belief, willingness to experiment, and immediate team culture all play a role. The data shows that while job role sets the backdrop, individual habit dictates adoption—task is not destiny.
Highest-adoption tasks
Rank | Detailed task | Rate |
|---|---|---|
1 | Read documents to gather technical information | 61.3% |
2 | Prepare research reports | 60.7% |
3 | Design integrated computer systems | 59.5% |
4 | Analyse data to identify trends or relationships among variables | 57.5% |
5 | Update knowledge about emerging industry or technology trends | 57.1% |
6 | Conduct quantitative failure analyses of operational data | 55.8% |
7 | Prepare data for analysis | 55.4% |
8 | Develop marketing plans or strategies | 55.0% |
9 | Collaborate with others to determine design specifications or details | 54.8% |
10 | Design websites or web applications | 54.6% |
Lowest-adoption tasks
Rank | Detailed task | Rate |
|---|---|---|
1 | Arrange delivery of goods or services | 0.0% |
2 | Assist practitioners to perform medical procedures | 0.0% |
3 | Collect biological specimens from patients | 0.0% |
4 | Drive trucks or other vehicles to or at work sites | 0.0% |
5 | Inspect electrical or electronic systems for defects | 0.0% |
6 | Maintain medical equipment or instruments | 0.0% |
7 | Measure the physical or physiological attributes of patients | 0.0% |
8 | Prepare patient treatment areas for use | 0.0% |
9 | Present food or beverage information or menus to customers | 0.0% |
10 | Record service or repair activities | 0.0% |
Note: Task-level adoption reflects generative AI used directly to perform that specific activity. It does not capture peripheral administrative or supportive tasks performed around the core activity.
Experience changes the equation
The most telling evidence concerns learning. Among people who already use generative AI, those with at least six months of experience apply it to about 9% more of their tasks than newcomers do. A typical worker has around five core tasks, so this is roughly one extra task. The step is small, and that is the point: adoption does not arrive all at once. It spreads one task at a time, as confidence grows.
The same pattern shows up elsewhere. People who use AI for one task tend to use it for others, and those who use it at work tend to use it at home. The authors are careful not to treat correlation as cause, but every result points in one direction: experience makes the next step easier.
This is why the first useful interaction matters so much. Once someone sees where the tool helps and where it fails, trying it again costs almost nothing. In most organisations, AI use travels through people long before it appears on any process map.
From capability to participation
Not every occupation will—or should—adopt generative AI at the same rate. Physical labor, caregiving, safety-critical roles, and high-liability fields carry inherent constraints, while other jobs simply gain little value from the technology today. Low adoption is not inherently a failure.
Yet the broader implication is clear: the next phase of AI integration will be driven less by model performance and more by practical workflow integration. For years, the core question was whether AI could do the work. Today, the real question is whether people can learn to work alongside it. Technology creates capability, but experience builds confidence—and confidence is what drives true participation.
────────────
References
Alexander Bick, Adam Blandin, David J. Deming and Tyler Schumacher, “What Work Does Generative AI Do?”, Federal Reserve Bank of St. Louis Working Paper 2026-017A, current version 25 August 2026.
────────────
Bridging the Gap: How We Help Teams Adopt AI?
Interested in expanding AI adoption across your company? Get in touch. We help teams move from isolated experimentation to confident, everyday use through practical coaching, carefully chosen use cases and a clear adoption plan.