AI for business: where the money shows up and where it stalls

Roughly one in five US businesses uses AI, and that share barely moved over six months. It’s the honest starting point for any plan built on AI for business, and it comes from the Census Bureau’s own survey rather than a vendor deck. The same data says where the exceptions sit: large firms, information, finance and insurance. Underneath the headline, the value is measurable in a few narrow places and thin nearly everywhere else.

So the useful question isn’t whether AI works. It’s which number you’re entitled to use as evidence, because the surveys disagree by a factor of four and each one counts a different thing. Everything below comes from pages we opened in August 2026, with the collection period stated wherever the figure moves with time.

One in five US firms use AI, and company size explains most of it

The Census Bureau’s Business Trends and Outlook Survey asks businesses whether they’re using AI in their operations. Between 14 December 2025 and 3 May 2026, the reported rate sat between 17% and 20%. Between 20% and 23% said they expected to be using it within the next six months. Expected use ran only about three points above current use across the whole period, which is a quieter finding than it looks: expectation isn’t converting into use very fast.

Size is the variable that the national average hides, and it hides a lot. As of the collection period ending 3 May 2026, 37% of businesses with at least 250 employees reported using AI, and 32% of those with 100 to 249 employees. Firms with fewer than 20 employees showed no significant change across the period, and among businesses with four or fewer employees the rate stayed under 20%. Adoption rose among firms with 20 or more employees and stalled below that line.

Sector splits the same way, which is why a single national figure travels so badly. Here are the three the Census Bureau called out, with the share expecting to adopt within six months alongside.

Sector	Currently using AI	Expects to use AI in 6 months
Information	39.7%	42%
Finance and insurance	33.9%	39%
Retail trade	14%	17%

Read the retail row against the information row and you get the shape of the whole market. The Information sector is nearly three times as likely to use AI as retail trade. That matters because most published advice on AI for business is written by and for the first group, then sold to the second.

The adoption number you pick changes the answer by four times

A Federal Reserve note published on 3 April 2026 by Jeffrey S. Allen put three US surveys side by side, and the spread is the most useful thing in the AI adoption debate. His FEDS Note reports firm-level adoption at about 18% at the end of 2025, work-related generative AI use by individuals at 41%, and an employment-weighted firm adoption rate of around 78%.

Survey	What it counts	Rate	Period
Business Trends and Outlook Survey	Share of firms adopting AI	about 18%	December 2025
Real-Time Population Survey	Share of individuals using generative AI for work	41%	November 2025
Survey of Business Uncertainty	Firm adoption, weighted by employee headcount	around 78%	November 2025

None of those numbers is wrong. They’re answers to different questions, and Allen says so plainly: “The different sampling distributions combined with adoption heterogeneity across firm size classes likely drives a considerable share of the gap between the BTOS and SBU estimates.” Weight by headcount and you count the big employers, which is where adoption already sat. Count firms one by one and the millions of small businesses dominate.

Intensity is the check that deflates all three numbers, because using a tool and depending on it aren’t the same thing. In the individual-level survey, 12% reported daily use and 35.2% reported use at least weekly as of November 2025. So even inside the 41% figure, most people touching these tools aren’t touching them most days.

Eighteen percent, 41% and 78% are all true at once. Whoever picks the survey has already picked the conclusion.

Rundowns AI

Where it pays: customer support and code

The strongest field evidence for a return is still a customer support study. Erik Brynjolfsson, Danielle Li and Lindsey R. Raymond tracked the staggered rollout of a generative AI assistant across 5,179 support agents, published as NBER working paper 31161. Access raised issues resolved per hour by 14% on average, and the average conceals the point. Novice and low-skilled workers improved 34%. Experienced and highly skilled workers saw minimal impact.

That distribution matters more than the average, because it tells you which teams the money actually comes from. The gain lands on the people furthest from the top of the skill curve, and it looks like compressed onboarding rather than a lift for everyone. A function with high turnover and a long ramp is where that arithmetic works hardest.

The second place with real volume is software. Anthropic’s Economic Index report sampled one million Claude.ai conversations and one million first-party API transcripts from 13 to 20 November 2025. The most common task on both was modifying software to correct errors, at roughly 10% of API records and 6% of Claude.ai usage. Business API traffic is concentrated, too: the top ten tasks account for 32% of it, up from 28% in the previous report.

The back-office automation that dominates conference agendas barely registers in that sample. Building and maintaining invoice processing systems came to 0.24% of API traffic. Classifying emails into predefined labels was 0.23%, and calendar scheduling and meeting coordination 0.16%. Generating personalised B2B cold sales emails was the largest of that group at 0.47%. Those are the workflows most often promised in enterprise pitches, and they’re rounding errors in actual usage. We made the same argument from a different angle in why enterprise AI keeps failing at deployment rather than capability.

Where it stalls: the trial that measured people who felt faster

Internal AI business cases tend to rest on self-reported time saved. There is one randomised controlled trial that tested exactly that, and the result cuts the wrong way. METR recruited 16 experienced open-source developers and ran a randomised trial across 246 real issues in their own repositories, projects averaging more than 22,000 stars and over a million lines of code. When AI tools were allowed, the tasks took 19% longer.

But the perception gap is the part that should worry anyone running a pilot. Those developers forecast a 24% speedup beforehand. After finishing the tasks slower, they still believed AI had sped them up by 20%. A survey of your own staff would have recorded a win that the stopwatch says didn’t happen.

METR is careful about how far this travels, and so are we. The authors state: “We do not claim that our developers or repositories represent a majority or plurality of software development work.” The setting was mature codebases, expert maintainers, and tools available in early 2025. The finding isn’t that AI slows everyone down. It’s that felt productivity and measured productivity came apart in the one place somebody measured both.

The autonomy story stalls in a similar way. Stanford’s 2026 AI Index reports that “AI agent deployment was in the single digits across nearly all business functions,” even as organisational AI adoption rose in 2025 to 88% of surveyed organisations, with generative AI used in at least one business function at 70% of them. So the pattern is broad adoption and narrow autonomy. That’s a distinction worth holding onto whenever a roadmap promises agents next quarter.

What it costs before anyone measures anything

Seat licences are the part of AI for business you can price exactly, so here it is from the two vendors’ own pages, read in August 2026.

Product	Annual billing	Monthly billing	Notes
Microsoft 365 Copilot Business	$21.00 user/month	$25.20 user/month	Add-on, needs a qualifying Microsoft 365 licence, up to 300 users
Microsoft 365 Copilot Business, promotional	$18.00 user/month	not offered	1 July to 30 September 2026, first year only, annual commitment
Claude Team, standard seat	$20 per seat/month	$25 per seat/month	For teams of 2 to 150
Claude Team, premium seat	$100 per seat/month	$125 per seat/month	Higher usage limits
Claude Enterprise, self-serve	$20 per seat/month	not printed	Seat price plus usage at API rates

Run the arithmetic on a mid-sized rollout and the number is small enough that it rarely gets a business case at all. A hundred Microsoft 365 Copilot Business seats at the annual rate is $25,200 a year, which is our calculation from the printed $21.00 figure. That’s less than one salary, which is precisely why so many of these deployments never got measured properly in the first place.

The catch sits in the row that doesn’t print a ceiling. Claude’s self-serve Enterprise tier is $20 a seat plus usage at API rates, so the seat is the floor and the token bill is the variable. Anyone budgeting from the seat price alone is budgeting the smaller half. We worked a full example of that second half in what it costs to run an AI feature.

The vendor line that appeared once and then went away

Services firms are the cleanest public proxy for corporate AI spending, because they book it as revenue. On 18 December 2025, reporting its first quarter of fiscal 2026 ended 30 November 2025, Accenture printed a key metric reading “Advanced AI new bookings of $2.2 billion” beside total new bookings of $20.9 billion. By our calculation that’s about 11% of bookings in the quarter.

The line didn’t appear in the key metrics of either quarter that followed. The second quarter release of 19 March 2026 led on record new bookings of $22.1 billion. The third quarter release of 18 June 2026 reported new bookings of $19.32 billion, down 2% in US dollars, on revenues of $18.7 billion, up 6% in US dollars. Neither carried an advanced AI bookings figure in its key metrics.

Demand for large-scale reinvention remains strong … we are seeing more large-scale AI transformation programs, while executing our strategy to capture new areas of growth.

Julie Sweet, Chair and CEO, Accenture, third-quarter fiscal 2026 results

We can’t tell you why the disclosure stopped, and we’re not going to guess. What we can say is that only the first of those three quarters carried a hard AI number, while the qualitative language kept going after the quantitative line stopped. The same third-quarter release counted 104 client bookings of $100 million or more year to date, up 13%, without saying how many were AI. That’s the pattern to watch in any supplier’s reporting, and it’s the same reflex we described in four questions that check whether a company really does AI.

What would change this conclusion

Three things would move it, and all three are observable from outside a company. The first is the Census series breaking out of its 17% to 20% band for two consecutive collection periods, with small firms driving the move rather than large ones. The band has held since December 2025, so a break would be genuinely new information.

The second is a randomised trial in a non-software function that finds a speedup and survives replication. Right now the measured wins concentrate in support and in code, and the one trial that timed expert developers found the opposite of what they felt. A clean result in finance, legal or operations would widen the map considerably.

The third is agent deployment climbing out of the single digits in Stanford’s next index, because that’s the threshold where the spending thesis stops being about assistance and starts being about headcount. Until then, the defensible version of AI for business is narrow: two functions with evidence, a seat cost you can approve without a meeting, and a usage bill that nobody has capped.