In July, Accenture published guidance for companies alarmed by their AI bills. Avoid blanket usage cuts, it said. Don’t tell employees to use less AI. Route work to the cheapest model that does the job. It also noted that in nearly every deployment it had examined, including its own, fewer than 10% of users and workflows drive the vast majority of the bill.
On Oct. 1, according to an internal announcement obtained by 404 Media, Accenture began centrally funding $29 a month per person toward Claude; anything above that had to be charged to a project. Internal messages seen by the publication show employees who had been spending $500 to $1,000 a month, and one who said their access was revoked entirely. Accenture says reporting that it has put a firm limit on its people’s access to AI tools is incorrect.
Take the company at its word, but the distinction is real, and the exposure is real: fixed-price contracts are more than 65% of Accenture’s bookings, so token costs land on its own margin. But from where an employee sits, the distinction is thin, and Accenture’s own paper records the pressure. In some of its environments, it wrote, requests to lift individual spending limits arrive by the hundreds each day.
Companies everywhere are asking the same question: how many tokens should we give each employee? It is the wrong question, and it is leading managers to decisions they will regret. The right question is what a finished piece of work costs, before AI and after. The answer is usually so lopsided that the argument ends there.
Here is an answer. One of our portfolio companies needed a new website. The agency quote was $60,000. Instead, three people from the business team — none of whom had ever built a website — did it in three days, and the result was better than what the agency proposed. It worked properly on Windows and Mac computers, on iPhones and Android handsets, and on Chinese phones running HarmonyOS — none of which the team knew to ask for. The agents also produced high-resolution versions of the company’s logos, made its marketing material consistent for the first time, and found customer testimonials nobody knew existed, along with awards the company did not know it had won. The tokens cost a few hundred dollars.
Here is a second. Preparing a board presentation on a company we wanted to acquire, we asked AI only to reformat the slides. It did that. It also fact-checked them, which we had not requested, and found several material errors. Then it reached the slide carrying our offer price and told us the price was too high, suggesting we cut it by 40%. We had not asked for its opinion. We checked the reasoning and lowered the bid.
Consider what this does to hiring. We were about to take on several entry-level marketing people at around $60,000 each. We hired one instead and paid considerably more. That one non-technical person does work the whole group could not have done: designing the company’s collateral and building analytics software to measure our customers’ operations against industry peers.
Notice what these have in common. It is not that people worked faster. Three people who could not build a website built one. A marketing hire who does not write software is shipping a software product. The change is not productivity but capability: work that used to require a specialist can increasingly be done by whoever understands the business problem.
We should be honest about what that means. Those are jobs that will not exist. One of us has a close friend, a graphic designer, laid off with two colleagues when new owners decided AI could do what the graphics department did. But a company that refuses these tools does not save those jobs. It loses them later anyway, along with its own.
So why is everyone capping? Because the bills arrived, they were larger than expected, and almost nobody can guess what an afternoon of their own AI use costs.
AI is sold by the token. A token is roughly three-quarters of a word, and everything you send the system and everything it sends back is counted and billed. A million tokens — about 750,000 words, or eight or nine novels — costs between $2 and $8.
A typical office worker asking 30 questions a day uses about 1.3 million tokens a month, roughly $6 of usage. That is why software companies sell $25 seats and still make money. Your average employee is not the problem and never was.
Heavy users are different. When people set AI agents to work — reading files, checking numbers, retrying when something fails — one task can consume 50,000 to 500,000 tokens. Call it $12,000 a year for your most intensive person, which is what the heaviest users in that reporting were spending.
Is that a good deal? Jellyfish, which measures what software teams actually finish, studied 12,000 developers at 200 companies this year. Among the lightest users, one finished job cost 28 cents in tokens; among the heaviest, $89; in other words, the heaviest users spent almost 320 times more than the lightest, which is like comparing the concert ticket sales of Taylor Swift’s Eras Tour to those of JoJo Siwa.
It does not settle the decision. The median developer in that study ran about $52 a month in tokens. The 90th percentile ran $691 — a gap of roughly $7,700 a year, against a salary many times that, for twice the output. The study is observational, so it shows association rather than cause. But the direction is not close. The marginal token is worth less, and still worth far more than it costs.
And the bill understates the cost. One of us modeled AI deployments in four industries — bank fraud detection, pharmaceutical research, chemical discovery and chemical manufacturing — comparing what companies are billed against what those systems cost once repeated attempts, idle capacity, monitoring and staff time checking output are counted. In all four, the token bill was about a quarter of the total. Cut the bill and you cut the quarter you can see, leaving the expensive three-quarters alone.
So measure outcomes, not consumption. The same firm’s published example is a good one: a quarter-end deck that took 15 hours across three analysts now takes two hours for one person, at a token cost of a couple hundred dollars. The token line on that work is trivial. The labor line it replaced is not. Only one of them is visible when you budget per head.
And be careful about writing rules now.
New York offers a warning about writing the rules prematurely. When the state imposed extremely heavy licensing requirements on cryptocurrency firms in 2015, before anyone understood the industry, a wave of exchanges left and never came back. The rules were not badly intended, but they were premature, written by people who would not bear the cost of being wrong. Business, too, should not regulate prematurely.
Companies are now doing this to themselves: approval workflows, restricted tool lists, per-person caps. A cap punishes your heaviest user, who is probably your most productive one. And every one of these is a rule written in 2026 that will still bind in 2028, when the facts have moved.
Universities once banned AI. Now they require it, and they pay for the licenses. Harvard Business School hands its MBA students ChatGPT and Claude. Darden encourages their use on assignments. Pace University — not an elite outlier, a practical school that sends graduates into New York firms — gives every student a Gemini subscription. A survey taken a year ago would have found most schools policing these tools. That is already the old world. They changed because employers did: when we hire recent graduates, proficiency with foundation models is the single most important skill we screen for.
This is a second industrial revolution, and the gains will be of the same order. Overspending is recoverable next quarter. Being left behind is not.
The opinions expressed in Fortune.com commentary pieces are solely the views of their authors and do not necessarily reflect the opinions and beliefs of Fortune.