Artificial Intelligence in Business: What the Numbers Show

Artificial intelligence in business is narrower and slower than the coverage suggests. As of 3 May 2026, 19.8% of U.S. businesses reported using AI in a business function, according to the Census Bureau, and use is concentrated in large firms: 37% of businesses with 250 or more employees, against under 20% of businesses with fewer than 20. The measured gains are real but task-level. Customer support throughput rose 15% in one large study, and professional writing time fell 40% in another.

Key takeaways

• U.S. business AI use held between 17% and 20% from December 2025 to May 2026, according to the Census Bureau’s Business Trends and Outlook Survey, which samples about 1.2 million businesses every two weeks.
• Adoption is a firm-size story: 37% of firms with 250+ employees use AI, against under 20% of firms with four or fewer employees.
• In a 2025 Quarterly Journal of Economics study of 5,172 customer-support agents, AI assistance raised issues resolved per hour by 15%, with the largest gains among the least experienced staff.
• A 2023 Science experiment with 453 professionals found writing time fell 40% and rated output quality rose 18%.
• A widely cited 2025 MIT report put the share of generative AI pilots with no measurable profit-and-loss impact at 95%, a figure that sits awkwardly beside the task-level results and deserves reading with care.
• The Lifoholic AI Payback Test prices one use case per month: in the worked example below, a six-person support team nets about $2,550 a month, but only if the freed hours are redeployed.

How many businesses actually use AI

Artificial intelligence in business means using machine-learning systems, including large language models, inside a business function such as marketing, customer service, finance or software development. That is close to the definition the U.S. Census Bureau now uses, which matters because the headline number depends entirely on the question asked.

The best ongoing measurement is the Census Bureau’s Business Trends and Outlook Survey, which covers about 1.2 million businesses on a two-week cycle. Between December 2025 and May 2026 the survey put overall AI use between 17% and 20%, with 20% to 23% of businesses expecting to use it within six months. AI use in the data collection period ending 3 May 2026 was 19.8%.

Size drives almost everything. Census figures for that period show 37% use among firms with at least 250 employees and 32% among firms with 100 to 249 employees, while firms with four or fewer employees stayed below 20%. Use rose among firms with at least 20 employees over the six months and did not change significantly among smaller ones. The gap is widening rather than closing.

Source: U.S. Census Bureau, Business Trends and Outlook Survey, May 2026 release.

Headcount changes the picture again. A Census Bureau working paper on the 2026 AI supplement found 18% of firms using AI in a business function during November 2025 to January 2026, rising to 32% on an employment-weighted basis. Most firms have not adopted AI. Most employees work somewhere that has.

Different surveys give different answers, and the gap is instructive. A Federal Reserve note published in April 2026 compared three sources: Census data showing about 18% of firms adopting by the end of 2025, the Real-Time Population Survey showing about 41% of individuals using generative AI for work as of November, and the Survey of Business Uncertainty estimating that 78% of the labour force works at a firm that has adopted AI. Firm-level counts, employment weighting and individual self-reports are three different questions, and quoting one as “AI adoption” without saying which is how the same market gets described as both 18% and 78% adopted.

Where AI measurably improves work

The strongest evidence for artificial intelligence in business comes from task-level field experiments, not from vendor case studies. Three results are worth knowing by heart because they set realistic expectations.

Brynjolfsson, Li and Raymond’s study in the Quarterly Journal of Economics, published in May 2025, tracked 5,172 customer-support agents through the staggered rollout of a generative AI assistant. Issues resolved per hour rose 15% on average. The distribution matters more than the average: less experienced and lower-skilled agents improved in both speed and quality, while the most experienced agents saw small speed gains and small declines in quality.

Noy and Zhang’s 2023 experiment in Science assigned incentivised writing tasks to 453 college-educated professionals and gave half of them ChatGPT. Average time fell 40% and rated output quality rose 18%. As in the support study, the weakest performers gained the most, which compressed the gap between workers.

Source: as listed. The consulting row is a qualitative summary; see the sourcing note for its status.

The third result is the corrective. In a field experiment with consultants, researchers described a “jagged frontier”: AI helped substantially on tasks inside its capability and made performance worse on tasks just outside it, with users unable to tell which was which. [LINK NEEDED: Harvard Business School working paper 24-013, Dell’Acqua et al. (2023), “Navigating the Jagged Technological Frontier”] The practical implication is that a single company-wide rollout averages good and bad use cases together, which is the fastest way to get a result of roughly zero.

Why most AI projects show no return

The most quoted number in this field is 95%. A July 2025 report from MIT’s NANDA initiative, The GenAI Divide: State of AI in Business 2025, concluded that 95% of enterprise generative AI pilots produced no measurable impact on profit and loss, against 5% that delivered significant value. [LINK NEEDED: MIT NANDA, “The GenAI Divide: State of AI in Business 2025”, July 2025 — link to the report itself, not press coverage]

Treat that figure as a signal rather than a measurement. The report is preliminary and was not peer reviewed, and published descriptions of its sample differ between accounts: some report 150 executive interviews and 350 employee surveys, others 52 interviews and 153 leaders surveyed, with 300 public deployments analysed in both versions. A number that cannot be pinned to a stable method is not in the same evidence class as a randomised experiment with 5,172 agents.

What the report says about causes is more useful than the headline. It attributes failure to a learning gap rather than to model quality: tools that do not retain context, do not improve from feedback and are bolted onto workflows rather than built into them. That is consistent with the jagged-frontier finding and with the Census data. Buying access is easy. Changing a workflow is the expensive part, and it is the part that produces the return.

The two bodies of evidence are less contradictory than they look. Task-level experiments measure what happens when a specific tool is placed inside a specific, measurable workflow. Pilot surveys measure what happens when an organisation buys a general tool and hopes a workflow forms around it. The difference between 15% and zero is usually scope.

The Lifoholic AI Payback Test

The Lifoholic AI Payback Test prices a single use case, monthly, before you commit to it. One use case, one number, one month.

• Monthly value = hours saved per week × 4.3 × loaded hourly cost
• Monthly cost = seat licences + supervision hours × supervisor cost + (setup hours × cost ÷ 12)
• Payback = monthly value − monthly cost

Worked example: a six-person customer support team, each handling chats 35 hours a week. Apply the 15% throughput gain measured in the QJE study. The same volume of work takes 182.6 hours instead of 210, freeing 27.4 hours a week. The Bureau of Labor Statistics put the mean hourly wage for customer service representatives at $20.92 in its May 2023 survey; adding 30% for employer costs gives a loaded rate of $27.20.

Source: Lifoholic calculation. Throughput gain from Brynjolfsson, Li and Raymond (2025); wage from BLS OES May 2023. Licence price, 30% employer loading, supervision and setup hours are stated assumptions, not sourced figures.

One rule keeps this test honest: hours freed are not money saved until something changes. If the team handles more volume, the gain is real. If everyone works the same hours and the queue simply empties earlier, the saving exists on paper only. Write down before you start which of the two you intend, because that decision is what separates a measurable return from a pilot that quietly ends.

Run the test on the workflow you can measure most easily, not the one that sounds most impressive. Deep, uninterrupted attention is what a serious evaluation needs, which is the argument behind Lifoholic’s guide to structuring focused work blocks.

Which functions to start with

Start where the work is text-heavy, repetitive and already measured. The Census Bureau’s updated AI supplement tracks use across 15 business functions, including finance, human resources, customer service, marketing, IT and research and development, and its sector figures show where adoption has actually gone: 39.7% in the information sector and 33.9% in finance and insurance as of 3 May 2026, against about 14% in retail trade.

Three questions decide whether a function is ready:

1. Can you state the current baseline in a number, such as tickets per hour, drafts per week or days to close the books?
2. Does a person already check this work before it leaves the building?
3. Would a wrong output be caught, or would it reach a customer, a regulator or an account?

A function that fails question one cannot produce a payback figure, only a feeling. A function that fails question three is the wrong place for a first deployment, whatever the vendor demo shows.

Training matters as much as tooling, because the support study found that the newest staff gained the most from AI assistance while the most experienced gained little. If you are rolling AI into a team, sequence the learning rather than dropping in a tool and a login; Lifoholic’s explainer on how sequential learning works covers why order changes retention.

[EXPERIENCE INSERT NEEDED: the author describes one real AI deployment with numbers — the function chosen, the baseline metric before and after, the monthly licence and supervision cost, and what the measured payback turned out to be. Include one use case that was tried and abandoned, and why.]

What artificial intelligence in business actually costs

The licence is rarely the largest line. Four costs show up repeatedly in the evidence and in practice.

• Supervision. Every output that reaches a customer needs a human check until error rates are known. In the payback example, supervision costs nearly twice the licences.
• Rework. The jagged-frontier result means some tasks come back worse. Rework time belongs in the cost column, not in a footnote.
• Integration. MIT’s report attributes most pilot failures to workflow integration rather than model quality, and integration is engineering time.
• Quality drift at the top. The support study found the most experienced agents saw small quality declines. Senior staff are often the last group who should be told to use the tool by default.

When this advice does not apply

• Regulated decisions. Lending, hiring, insurance pricing, medical and legal advice carry disclosure and fairness obligations that vary by country and state. Treat the payback figure as irrelevant until compliance is settled.
• Businesses under five employees. Census data shows adoption below 20% in this group, and the payback maths rarely works when there is no repeated, measurable workflow to improve.
• Physical work. The measured gains cited here come from text-based tasks. They do not transfer to trades, logistics floors or hands-on services.
• Confidential or client data. Where contracts or data protection rules restrict where data is processed, the tool choice is constrained before the business case is.
• Markets outside the U.S. All adoption figures here are American. Rates, sector mixes and rules differ, and the U.K., EU and India each measure this differently.
• Anything you cannot baseline. No baseline, no payback, no way to know whether the pilot worked.

Frequently asked questions

What is artificial intelligence in business?

Artificial intelligence in business is the use of machine-learning systems, including large language models, inside business functions such as customer service, marketing, finance, IT and software development. The U.S. Census Bureau measures it as a business’s use of AI in any of its business functions, which is the definition behind the widely quoted adoption rates.

What percentage of businesses use AI?

About 19.8% of U.S. businesses reported using AI in a business function as of 3 May 2026, according to the Census Bureau’s Business Trends and Outlook Survey, within a six-month band of 17% to 20%. Weighted by employment the figure is higher, around 32%, because larger firms adopt more.

Does AI actually improve productivity?

In measured settings, yes, by task-specific amounts. Customer-support agents resolved 15% more issues per hour with an AI assistant in a 2025 study of 5,172 agents, and professionals completed writing tasks 40% faster in a 2023 experiment. Gains concentrate among less experienced workers and shrink or reverse on tasks outside the model’s capability.

Why do most AI projects fail?

A 2025 MIT report put the share of generative AI pilots with no measurable profit-and-loss impact at 95%, attributing failure to weak workflow integration rather than model quality. The pattern in that report and in the research is consistent: general tools bought without a defined workflow and a baseline metric produce no measurable return.

How much does it cost to use AI in a small business?

Seat licences are usually the smallest line. In the worked example here, six licences cost $180 a month while supervision and amortised setup added $477. Budget for review time, rework and integration work, then compare all of it against the hours the tool actually frees.

Which business function should use AI first?

Pick the function with a baseline you already measure, an existing human review step, and low consequences for a wrong output. Text-heavy work such as support replies, drafting and internal documentation fits those criteria. Census data shows the information and finance sectors adopting fastest, and retail trade slowest.

What to do this week

Pick one workflow, write down its current number, and run the Payback Test on it before buying anything. If you cannot state the baseline in a sentence, that is the first task, not the AI. Decide in advance whether freed hours become more output or fewer hours, because that choice is what makes the result countable. Run it for a month and compare the figure to the one you wrote down. More on the numbers behind everyday business decisions is in the Business archive and the Technology archive.