Four surveys measured AI adoption in the American economy over the past year and returned four different answers. The Census Bureau’s Business Trends and Outlook Survey put it at about 18% of firms at the end of 2025. The Real-Time Population Survey put work-related generative AI use at about 41% of the workforce in November. The Atlanta Fed’s Survey of Business Uncertainty, which asks senior executives rather than firms or individuals, estimated that 78% of the labor force works at a company that has adopted AI. A Census household survey fielded in March 2026 found that 55% of workers had used AI on the job for at least one of 11 listed tasks.
The instinct is to ask which number is right. Jeffrey Allen of the Federal Reserve Board, who assembled the first three in an April 2026 note, treats that as the less useful question. Each survey has a different unit of analysis and a different weighting method, and he assigns each one a job: the Business Trends and Outlook Survey is the best estimate of the share of U.S. businesses that have adopted AI, the Real-Time Population Survey is the best estimate of the share of the labor force using generative AI at work, and the executive survey is, in his words, “a good upper bound on the scope of access to AI tools at work.”
Read that way, the numbers stop competing, and they also stop being comparable to one another. Only two rest on the same base. The executive survey’s 78% and the employment-weighted 32% reported in an April Census working paper both describe shares of the labor force: the first counts people who work somewhere that has adopted AI at all, and the second counts people who work somewhere that reports AI in use within a business function. The 46-point difference between them, on one denominator, is what this piece calls the translation gap.
The surveys measure different things accurately
A spread this wide between estimates of the same phenomenon usually indicates a measurement problem. In this case the surveys are precise, and they are measuring different things.
One caution belongs up front, and the Fed note states it plainly. In November 2025 the Census Bureau revised its question from AI use “in producing goods or services” to use “in any of its business functions,” which breaks the survey’s own series at that date. Every Business Trends and Outlook figure in this article comes from after the revision, so the figures are comparable to each other but not to the pre-revision readings that most 2025 commentary relied on.
Allen is direct about the deeper measurement problem, and his statement of it is the most precise short description of the gap I have found: the accessibility and general-purpose nature of modern generative AI tools “have made it difficult to distinguish experimental, incidental or otherwise insignificant AI usage at work from more meaningful adoption patterns.” A worker who asks a model to rephrase an email has used AI at work. So has a firm that has encoded its monthly reporting process end to end. Both answer yes, and only the second has changed how the business runs.
The household data describe the first kind of use. In the Census Bureau’s March 2026 Household Trends and Outlook Pulse Survey, among workers who used AI on the job in the prior week, 31% said it saved them one to two hours of task time and 25% said it saved them less than an hour. Another 15% reported three to four hours saved, 15% more than four hours, and 10% none, while 3% said using it cost them additional time.
Those figures deserve to be read in both directions. Saving an hour or two on a task is useful. It is also roughly what a capable person achieves by finding a useful tool and applying it to their own work in their own way, without any change to how the company operates.
Performance correlates with breadth
The most important finding is in a Census Bureau working paper released in April by Kathryn Bonney, Cory Breaux, Emin Dinlersoz, Lucia Foster, John Haltiwanger, and Aditya Pande, drawing on a new AI supplement to the Business Trends and Outlook Survey fielded between November 2025 and January 2026.
The authors find a robust positive correlation between a firm’s commercial performance and the breadth of its AI integration, measured across functional deployment, task-level use, and operational investment. The correlation attaches to breadth rather than to the fact of adoption.
Most adopters lack that breadth. Among firms that had adopted AI, 57% had integrated it into three or fewer business functions, and 65% limited worker use to three or fewer tasks. Among users, 66% reported using it only to augment existing work rather than to change how the work is organized. Firm-level adoption was 18%, rising to 32% on an employment-weighted basis, and 23% of firms reported workers using AI in work-related tasks.
The performance correlation attaches to breadth of integration, and breadth comes from documenting how each function’s work is done.
Taken together, the two findings are unusually clear for survey work. Depth of integration across the business correlates with performance, and most adopters run AI in three functions or fewer. The 43% running it in four or more are the firms the performance finding describes. The two groups have access to the same technology; what separates them is how far they have integrated it.
The functions AI reached first carry little of the payroll
The same working paper reports where inside the business AI was deployed. Among adopting firms, the most common deployments were Sales and Marketing at 52%, Strategy and Business Development at 45%, and IT at 41%. At the worker level, the leading generative AI uses were writing, document analysis, and information search. The household survey found the same ordering from the worker’s side: searching for information or technical help at 37%, writing communications or documentation at 32%, generating ideas at 32%, interpreting or summarizing at 31%, and administrative tasks at 27%.
Compare that list with a privately held company with $5 million to $50 million in revenue. Sales and marketing is often one person, or part of the owner’s role. Strategy and business development usually belongs to the owner. IT is often an outside vendor on a service agreement. In a company of that size, the three functions where AI has spread furthest carry some of the smallest payroll.
The payroll in those companies is concentrated elsewhere: production scheduling and dispatch, order entry, shipping coordination, service documentation, billing and collections, permit and compliance paperwork, the monthly close, and reports sent to customers on a fixed schedule. Those hours repeat and follow a recognizable sequence, and in the companies we review they usually account for more labor cost than the functions where AI has been deployed. They are also hours that a general-purpose tool cannot perform correctly without instruction, because doing them correctly depends on how this specific company does them.
The size data are consistent with this. As of May 2026, 37% of firms with at least 250 employees reported using AI, against a national rate of 19.8% and under 20% among firms with four or fewer employees. The smallest firms are the least likely to have their processes written down in a form a general-purpose tool can follow.
Translation has no natural owner
Between a general capability and a specific company’s work there is a step that nobody is assigned to do. We call it translation.
The step is mostly writing. Take one recurring piece of work and write down how it is done, including the parts that are done badly. Write down what a good version looks like, specifically enough that someone could grade an attempt. Write down the inputs it needs and where they are kept, and what must never appear in the output. Once that is encoded, the work runs the same way each time and can be reviewed against a written standard instead of the preferences of whoever reviews it.
Few providers sell this step, for a structural reason. Software vendors sell capability, because a product is a capability. Consulting firms sell large engagements, because large engagements are what their staffing model supports. Translation is specific to one company, cannot be packaged as a product, produces little that can be demonstrated, and needs a few uninterrupted days from someone who understands both the work and the tool. Where a license is already in place, the constraint is rarely technology or budget; it is that those days are never scheduled.
This also explains the adoption statistics. A survey can detect access, which is now broad, and contact, which is common. It has difficulty detecting translation, because translation looks like documentation rather than technology, and it appears in the data only later, as breadth.
What follows for your own decisions
If you own a company in this range, start with your team’s hours before evaluating any technology. Take one week and sort the hours into two columns: hours in which a person decided something, and hours in which a person carried out a known sequence. The second column is your list of translation candidates. Building it requires no purchase, and it shows where a tool would have work to do.
If you advise these companies, the breadth result is the finding to bring into client conversations. A client who reports using AI has told you little about whether it has changed their economics. The follow-up questions are how many functions, how many tasks, and whether any of it is written down. A firm with three encoded processes and a review standard is in a different position from a firm where nine people have each saved an hour on their own.
If you allocate capital into privately held businesses, the translation gap is a diligence question with an answer. Two companies with identical financials can differ substantially in how much of their labor cost sits in pattern hours that have never been written down. That difference does not appear in a quality of earnings report, and I have not yet seen it reflected in a purchase price. Measuring it takes about a week.
Better models will improve each process that has been documented, and the number of documented processes is what moves a company from the 78% toward the 32%. The first one can be written this quarter.
Works Cited
- Monitoring AI Adoption in the U.S. Economy | Jeffrey S. Allen, FEDS Notes, Board of Governors of the Federal Reserve System, April 3, 2026
- The Microstructure of AI Diffusion: Evidence from Firms, Business Functions, and Worker Tasks | Kathryn Bonney, Cory Breaux, Emin Dinlersoz, Lucia Foster, John Haltiwanger and Aditya Pande, Center for Economic Studies Working Paper CES-26-25, U.S. Census Bureau, April 2026
- About a Third of Workers Who Used AI in the Last Week Said They Completed Tasks One to Two Hours Faster | Renee Stepler, America Counts, U.S. Census Bureau, August 11, 2026
- Large Firms With at Least 20 Employees Biggest AI Users | Adam Grundy, Cory Breaux and Dhanapati Khatiwoda, America Counts, U.S. Census Bureau, May 26, 2026
- Business Trends and Outlook Survey | U.S. Census Bureau