AI Is Everywhere at Work. So Why Is It Doing So Little of the Work?
AI adoption at work is starting to look impressive. According to a new St. Louis Fed report, What Work Does Generative AI Do?, the share of workers using generative AI on the job rose from 33% in August 2024 to 45% by May 2026. AI use is also spread broadly: in more than 80% of occupations, at least one in five workers now uses it.
At first glance, that sounds like deep transformation. Then the Fed looks at the work itself—and the picture changes: fewer than 3% of detailed job tasks have AI adoption above 50%, and none exceed 70%.
That creates a useful contradiction:
AI adoption looks deep when measured by people, but shallow when measured by work.
Or more simply: AI has entered many jobs without yet entering most of the work inside those jobs.
“Do You Use AI?” May Be the Wrong Question
The headline adoption question is usually simple: “Do you use AI at work?” What Work Does Generative AI Do? asks a more revealing question: “What work do you actually use AI for?”
The Fed does this by linking workers to detailed occupational tasks. A worker can truthfully say, “Yes, I use AI at work,” while using it for only a small slice of the job—maybe summarizing documents or drafting emails. The rest of the work may remain largely unchanged.
“45% of workers use AI” does not mean “45% of work is being done by AI.”
Broad Adoption, Shallow Penetration
The most AI-assisted tasks are cognitive and information-intensive: reading documents for technical info, preparing research reports, and analyzing data. By contrast, physical and face-to-face tasks like driving trucks, preparing treatment areas, and assisting with medical procedures show little reported AI use.
What is more interesting is that high occupation-level adoption is not driven by one task everyone automates. Instead, workers use AI for different parts of their jobs. Workplace AI may currently be distributed more broadly than it is embedded deeply.
And That Raises the ROI Question
The Fed does not measure ROI, time savings, output quality, or enterprise productivity. So the right conclusion is not that AI use is superficial. The better question is:
If AI adoption is concentrated in a relatively small share of tasks, how much business value should we expect headline adoption rates to produce?
Company A: 70% of employees use AI occasionally for email, summaries, and research.
Company B: Only 35% use AI—but systems are embedded deeply into underwriting, pricing, customer acquisition, fraud detection, or software development.
The adoption metric says Company A is ahead. The economics might say Company B. That is the gap between usage and value realization.
Usage Is Not Penetration
Faster emails and summaries are useful. But whether those improvements materially affect revenue, margins, throughput, error rates, or labor productivity depends on where they sit inside the operating system of the business.
If AI sits mostly at the edges of work, it can make many employees somewhat faster without fundamentally changing the workflow economics. If it penetrates economically important tasks, the effect could be much larger.
Usage is not penetration.
Penetration is not productivity.
Productivity is not ROI.
What AI Can Do Is Not What Workers Will Use It For
Measures of AI “exposure” predict which tasks will adopt AI, but only up to a point. Some occupations use AI far less than predicted because privacy, sensitive records, and the cost of errors create barriers. Others adopt much more than predicted because workers find uses in planning and communication even inside partially manual jobs.
What AI can do is not the same as what organizations and workers will actually use it to do.
The Shallow Pattern May Not Last
Workers who have used generative AI longer tend to use it across more tasks. Once workers learn how to use AI successfully in one domain, they begin finding uses elsewhere. AI may already be widespread across the labor market while still working its way deeper into actual workflows.
If that happens, the next meaningful question will be:
Which economically important tasks are actually changing because of it?