How Are Companies Using AI? What the 2026 Data Actually Shows

How are companies using AI in 2026? Narrowly, and mostly to help people write things. The US Census Bureau puts firm-level use at 19.8% as of May 2026. Among firms that do use it, sales and marketing is the top function at 52%, and 66% of adopters use AI only to augment worker tasks rather than replace them. Meanwhile 88% of surveyed organizations tell Stanford's AI Index they use AI somewhere. Both numbers are correct. They count different things.

That gap is the most useful thing in the data, so this post starts there.

TL;DR

• Firm-level US adoption is 19.8% as of May 3, 2026 (US Census BTOS), but worker-level generative AI use is about 41% and employment-weighted firm adoption is 78%. Four surveys, four units of count.
• Among AI-using firms: sales and marketing 52%, strategy and business development 45%, IT 41%.
• 66% of adopters use AI purely to augment tasks. Only 2% of firms report AI-attributable employment reductions.
• Size is the strongest predictor: 37% of firms with 250+ employees versus under 20% of firms with four or fewer.
• AI agent deployment is still in the single digits across nearly every business function, even at 88% organizational adoption.

The number you cite decides the decision you make

Four credible surveys, run in the same economy, in the same twelve months, produced 18%, 41%, 78%, and 88%. Federal Reserve economist Jeffrey S. Allen laid the discrepancy out directly in an April 2026 FEDS Note: BTOS had roughly 18% of businesses adopted by the end of 2025, the Real-Time Population Survey had about 41% of the workforce using generative AI for work as of November 2025, and the Atlanta Fed's Survey of Business Uncertainty, weighted by employment, put 78% of the labor force at firms that had adopted.

None of these are wrong. BTOS counts firms and asks a strict production question. RPS counts people. The employment-weighted number counts where the jobs are, so one 40,000-person adopter outweighs four hundred small holdouts. Stanford's 88% counts organizations in a survey panel skewed toward large enterprises, and accepts use in a single function as a yes.

Practically: if you are sizing a market, use the firm count. If you are writing an internal policy, use the worker number, because 41% of your staff are already doing this whether or not procurement approved it. If you are deciding whether you are behind, use the employment-weighted figure and accept that you probably are.

What companies actually run AI on

The best function-level data is not a vendor survey. It is the Census Bureau's April 2026 working paper on AI diffusion by Bonney, Breaux, Dinlersoz, Foster, Haltiwanger, and Pande, built on BTOS microdata linked to firm records. Among firms using AI, the leading areas are sales and marketing at 52%, strategy and business development at 45%, and IT at 41%.

At the task level the picture narrows further. Writing, document analysis, and information search dominate. That is a much smaller claim than "AI transformed our go-to-market," and it is the accurate one.

The pattern behind those three functions is consistent. Output is text, a bad draft costs almost nothing, and a person reviews it before it leaves the building. Finance, legal, and operations lag not because models handle numbers badly but because the review cost per output is higher and the failure is expensive. That is the same test I laid out in how to use AI in business: pick the process where a wrong answer is cheap and a review already exists.

Augmentation is the default, not the cover story

The same Census paper found that 66% of AI-adopting firms use it exclusively to augment worker tasks, and that employment reductions attributable to AI appeared in only 2% of firms. That is 33 augmenting firms for every one cutting headcount.

Run the arithmetic on what that means for a business case. If 19.8% of firms use AI, and 2% of all firms report AI-driven employment reductions, then even generously attributed, roughly one in ten adopters is touching headcount at all. A plan whose payback depends on removing a role in year one is betting against nine out of ten of its peers.

The realistic payback is time inside an existing role. If AI saves a 10-person marketing team four hours a week each, that is 40 hours a month against a seat cost of maybe $300. The math works. It just does not show up on a headcount line, which is why measuring it is its own problem. I wrote the measurement version of this in how to measure AI ROI.

Size predicts adoption better than anything else

Census figures as of May 3, 2026: 37% of firms with at least 250 employees, 32% of firms with 100 to 249, and under 20% of firms with four or fewer. Between December 2025 and May 2026, use rose among firms with at least 20 employees and did not move significantly below that line.

By sector the same story repeats with different labels. Information sits at 39.7%, finance and insurance at 33.9%, retail trade at 14%. Very large firms in information, professional services, and finance run 50% to 60% unweighted.

The cost that gates a deployment is fixed, not variable. Someone has to pick the process, write the evals, get data access, define the review gate, and own the thing when it breaks. A 400-person firm assigns that to a role. A 6-person firm assigns it to whoever has a spare Thursday. API pricing is not the constraint. Staffing the ownership is.

Agents are where the talk and the deployment separate

Stanford HAI's 2026 AI Index reports 88% of surveyed organizations using AI in at least one business function and 70% using generative AI specifically, while AI agent deployment remains in the single digits across nearly all business functions.

Hold those two numbers next to each other. Near-universal assistive use. Near-zero autonomous use. The measured productivity gains in the same report land where the human stays in the loop: 26% in software development, 14% to 15% in customer support, and 50% on marketing output volume, with smaller gains on tasks that need deeper reasoning.

That is not evidence agents do not work. It is evidence that the expensive part of an agent is not the model. It is evaluation, permissions, failure handling, and an owner. If you want the distinction sharpened, AI workflow vs AI agent is the post that draws the line, and most teams asking for an agent want a workflow.

What to do with this

Four moves, in order.

1. Find out what your staff already run. If 41% of the national workforce uses generative AI for work, a chunk of your team does too, unsanctioned. Ask before you write policy.
2. Start where sales and marketing started. Text output, cheap failure, existing review step. Not because it is exciting, but because it is where 52% of adopters proved the pattern.
3. Assume augmentation. Budget against hours returned inside a role, not a headcount line, and instrument it on day one.
4. Build the review gate before the autonomy. The single-digit agent number is a staffing and evaluation gap, not a model gap.

For the function-by-function version of where this lands on a P&L, AI in business covers it. If finance is your first candidate, the Claude finance agents pack has the working builds, and the AI daily driver stack is what I actually run day to day.

The bottom line

How are companies using AI in 2026? One in five US firms uses it in production, mostly for writing and document work in sales, marketing, strategy, and IT. Two thirds of adopters use it purely to augment people, and 2% of firms have cut jobs over it. Big firms are roughly twice as likely to have deployed as small ones, and that gap is about who owns the project, not what it costs. Agents are still single digits. The companies getting value are not doing something exotic. They picked a text-heavy process with a cheap failure mode, put a person on the review, and measured the hours.

The next post in this thread goes out to the newsletter first. Join it here, then grab the AI daily driver stack and pick your one process this week.