The Average Firm Spends $2,068 Per Employee on AI. Ask What It Got Back.

What firms actually spend on AI, and where the number comes from

Start with the source, because it is unusually good. NBER Working Paper 34836 surveys nearly 9,000 CFOs, CEOs and senior executives at firms in the US, the UK, Germany, Australia, Canada, Mexico and South Africa, in the summer of 2026. The same research team, economists from the Bank of England, the Atlanta Fed, the Bundesbank and Macquarie University, had already run the identical questions in the winter, and both waves are folded into the revised paper. It is a working paper, not peer-reviewed, and its numbers are what executives report about their own firms rather than something an auditor checked. Treat it as the best available map, not as the territory.

The headline figures, from the paper's own abstract:

• 78% of firms use AI.
• US firms expect to spend about $2,000 per employee on AI-related expenditure in 2026.
• That implies roughly $300 billion of aggregate annual spending across the sample.
• 2% of firms account for more than half of the total.

The detail behind the average is more useful than the average. In a companion piece written by three of the paper's own authors, published on the Atlanta Fed's macroblog in May 2026, the per-employee number is broken out by year and by sector. The survey question deliberately bundles everything: software, subscription services, hardware, worker training and IT staff support. So when you read "$2,068", that is not a licence bill. It is the whole AI line of a firm's budget, IT payroll included.

Two things stand out. First, spending is rising fast, roughly 50% in one year, which tells you this is still a budget being built, not one that has settled. Second, the sector spread is enormous: a law firm or a consultancy is putting four times per head into AI what a factory is. If your business looks more like the factory, the services number is not a target, it is somebody else's problem.

The aggregate passes an independent sanity check too. The paper's bottom-up estimate of about $280 billion lines up with the Stanford HAI AI Index 2026, which puts US firm spending at $285 billion. Two teams counting in different ways landing within 2% of each other is the closest thing to a fact in this area that anyone has.

The average is not the number to copy

Here is the trap in that table. It says "average", and the underlying distribution has a very long tail. In the untrimmed data the per-employee figure is $3,108; the researchers winsorise at the 99th percentile (capping the biggest spenders), and the number drops to $2,068. A third of the average is being carried by a handful of firms whose AI budget is measured in millions.

The cleaner statement is the concentration figure: 2% of firms account for more than half of all AI spending. That is not a small-firm phenomenon, and it means the "average firm spends $2,068" sentence is, strictly, describing nobody. There is no median firm sitting at $2,068 with an unremarkable budget; there is a large group spending well under it and a small group spending enormously over it.

For a business of five to fifty people, that changes the question from "am I keeping up?" to "what is my number, and can I justify it?". Those are different questions and only the second one has an answer you can act on. Adoption, meanwhile, is already close to saturated: 78% of firms report using AI, and in the earlier winter wave the same team found 69% across four countries, with the US at 78% and Australia at 59%, and text generation the single most common application at 41% of firms (NBER Digest summary, May 2026). Using AI is not a differentiator any more. Spending on it wisely might be.

None of that argues for spending less out of principle. It argues for refusing to treat a number pulled upward by a few giants as a target for your own budget, and for building the number from your own process instead.

Spending moves in months; productivity moves in years

Now the part that decides whether the money was worth it. The same executives who forecast the spending were asked what AI had actually done to their firm's labour productivity over the previous three years, and what they expected over the next three. The reported answer is 0.2 percentage points of annual labour productivity growth over the last three years, against an expectation of 0.6 points per year for the next three. In the earlier wave of the same survey the expected three-year productivity gain was 2.25%, with a predicted 1.2% reduction in headcount.

Read that as a forecast of a forecast. The executives are not lying: 0.2 points is a real, positive effect, and it is consistent with what independent econometric work finds. BIS Working Paper 1325, published in January 2026, matched data on more than 12,000 EU and US non-financial firms and put the productivity gain from AI adoption at 4% of the labour productivity level, with no adverse effect on employment in the short run. Four percent of a level and 0.2 points of a growth rate are different animals, but both say the same thing: the gain is real, and it is not the doubling that the sales deck implies.

The uncomfortable detail is in the BIS paper's second finding. The productivity benefit "concentrates in medium and large firms". The firms big enough to have someone whose job is to make the tool fit the process are the ones banking the gain. A ten-person firm buys the same licences and captures less, because there is no spare person to do the fitting.

So the budget question and the timing question are the same question. If you approve a monthly subscription on the promise of a productivity jump, you are buying into a curve that historically takes about three years to bend, and that bend is hardest for firms with an implementation capacity you may not have. The firms that close the gap are rarely the ones that bought the most licenses. They are the ones that changed a process and left it changed.

The arithmetic for a ten-person firm

Benchmarks are useless until you put your own numbers into them. Take a ten-person building firm: a tiler, two apprentices, an office manager, a buyer and the owner. Assume the AI budget question is the one the owner is actually asking: is it worth putting a tool in front of the quoting process, which is where the firm loses the most evenings.

Assume, stated plainly as assumptions and not as findings: the firm produces three quotes a day between the office manager and the owner; each quote takes 40 minutes of measuring notes, product lookups and typing, so about two hours a day of the firm's time; the blended cost of an hour of that time, counting salary, taxes and the owner's own hour priced honestly, is $35. That is $70 per working day, or roughly $15,400 a year on quoting alone.

Now price the alternative. At the survey's $2,068 per employee, a ten-person firm's "average" AI budget would be $20,680 a year. That is more than the entire cost of the process it is supposed to improve. This is the single most useful comparison the benchmark gives you: the average firm is spending, per head, more than most small firms lose on any one process, which is only rational if the spending is spread across many processes at once.

Run the other direction and count what a tool would have to save. A quoting assistant at $30 per seat, used by two people, is $720 a year. To be worth it at a 3:1 ratio it has to free up about $2,160 of the office manager's time: about 62 hours, or 37 working minutes a week. That is the honest bar: an hour a week, not "a revolution". If the tool saves 15 minutes a week, it is not a failure and it is not worth the switching cost either.

And keep the whole AI line in view, because the survey's definition of AI spend includes training and IT support. A $720 subscription with a two-day setup and a person half a day a week keeping it working is not a $720 decision. The number to bring to the budget conversation is the subscription plus the hours, and the hours are usually the larger half.

How to set your own number

The benchmark is a reference point, not a plan. Build the plan in this order, and do it in one sitting rather than over a quarter.

1. Pick one process, and write down what it costs today. Count the hours and the people, not the frustration. Two hours a day of quoting is a cost; "quoting is a nightmare" is not.
2. Set the bar before you shop. Decide the minimum saving that makes the tool worth it, say 30 minutes a week of real time, and write it down. This is the sentence that protects you from a persuasive demo.
3. Price the full line, not the licence. Subscription plus setup plus the recurring minutes someone spends keeping it fed with context and checking its output. Add the checking honestly; it is usually the item that kills the return.
4. Compare against the sector number, not the headline. If you work in manufacturing or trades, the $900 figure is the relevant one, and being under it is normal, not backward.
5. Give it a review date. One quarter, same process, same measurement. If the saving is not visible, cancel without ceremony. The survey's own finding is that big spenders are concentrated in a small club, and there is no prize for joining it.
6. Only then consider a second process. Spreading AI thinly across five processes is how a firm ends up with a large AI line and an unmeasurable return.

Notice what is not on that list: a target per-employee spend. Working backwards from a benchmark to a purchase is the mistake the table at the top invites. The benchmark tells you whether your number is wildly out of line with your sector. It does not tell you what to buy.

When the right number is zero

Some firms should not be spending on AI this quarter, and the evidence says so more clearly than any vendor will.

If you do not yet have the process written down, do not automate it. A tool wrapped around an improvised process produces improvised output faster, and the first thing you will do is hire back the time as review. The measured version of this is the review burden: the same executives who forecast large gains also report that a large share of saved time goes back into checking the output. That is the cost the $2,068 figure hides.

If you are under about ten people and your work is varied, bespoke jobs, one-off projects, a trade where every site is different, the BIS finding applies with full force. The productivity gains they measured concentrated in medium and large firms, and a firm with no spare implementation capacity is the firm least likely to collect it. Waiting a year is a strategy, not a failure of ambition.

If the only reason you are considering AI is that 78% of firms use it, that is not a reason. Adoption near saturation means the tool itself has stopped being an advantage; what remains scarce is knowing which process to point it at. And if you are a freelancer or a firm of one, be aware that these surveys barely cover you. The sample is firms, with the smallest ones the least represented (NBER Digest, May 2026). The published averages are, for you, an extrapolation across a gap in the data.

There is a version of this where spending nothing is the aggressive choice: keep the money, keep the process manual, and revisit when either the process has stabilised or a tool exists that fits it without a project. The firms that do best with AI are not the ones that started earliest. They are the ones that started on a process that was ready.

Frequently asked questions

How much does the average firm spend on AI per employee?

About $2,068 per employee for 2026 in the NBER et al. survey, up from $1,358 in 2025. The question counted software, subscriptions, hardware, training and IT staff support together, so it is a total AI budget line rather than a licence cost. In manufacturing the figure is closer to $900, and in professional and business services it is $3,470.

Is $2,000 per employee a good target for a small business?

No. The average is pulled upward by a small number of very large spenders: 2% of firms account for more than half of all AI expenditure, and the figure before trimming outliers is $3,108 rather than $2,068. Use it to check whether you are wildly out of line with your sector, not as a spending target to hit.

How quickly should a small firm expect AI to pay off?

Slower than the purchase order suggests. In the same survey executives attributed 0.2 percentage points of annual labour productivity growth to AI over the previous three years, while expecting 0.6 points per year for the next three. Independent work on 12,000+ firms puts the level effect at about 4% and finds it concentrated in medium and large firms.

What should I measure before approving an AI subscription?

Two numbers: what the process costs today in hours and money, and the minimum saving that would make the tool worth it. Then price the full line (subscription plus setup plus the recurring minutes spent feeding and checking it) and set a review date one quarter out. If the saving is not visible at that date, cancel.