The AI Invoice Means It's Working. Is It Buying Anything?

In brief. AI spending per employee is rising about 50% a year, and a few companies are spending most of it. How fast the invoice grows is set by the vendor's defaults, not the company's plan. Most companies cannot say what the spending was supposed to produce. And every surprise AI invoice with a company's name on it arrived after adoption worked, at the companies that have gone furthest first.

The detail

AI spending is rising. U.S. companies spent $1,358 per employee on AI in 2025 and expect $2,068 in 2026, up 52%. More than half expect to spend no more than $200 per employee; the top 10% plan $2,800 or more, 14 times as much. The gap is not two kinds of company. It is how far one company's spending rises once its people use AI heavily.

The vendor sets the pace. In April, Anthropic changed which model answers by default in Claude Code, its tool for software developers, and its own published cost estimate went from $6 to $13 per developer per working day. A spokesperson: "There was no pricing or product change." An independent measurement found the same work cost about 40% more after one model update at the same advertised price.

Most companies cannot say what the spending was for. Four in five executives at large companies say AI is producing measurable results; less than one dollar in five of their AI usage spending can be tied to a measured financial result. In the U.S., 82% of senior leaders at companies investing in AI are concerned about usage costs; 64% say their company tracks usage with clear limits. More are worried than have controls, and more have controls than can answer the board's question: what was this supposed to do for us?

What it looked like. Uber used up its 2026 budget for AI coding tools about four months into the year; Priceline's renewal for one tool came back at four to five times the price; Salesforce told investors in August that AI usage costs were part of why it did not raise its profit forecast. Those invoices arrived in the spring. By autumn they had reached filings, forecasts and the surveys boards read, which is why the question is live in budget season. The cases, and the evidence that cuts the other way, are in section 05.

In brief. Set two numbers before the rollout: the target outcome, and the cost per result the company will accept. Then control the cost per result rather than who may use the tool, and keep checking both for as long as the AI runs. That is how the return gets collected. It is not a case for restricting AI.

The detail

The two problems in this issue are one problem. A company that never wrote down what the AI was supposed to produce has nothing to judge the invoice against. A company that did, but left the vendor's defaults in place, watches the running cost grow with success until it eats the result. The companies that kept their value did both halves: a target, and a cost per result, held together and rechecked as models, prices and users changed. They came out using more AI, not less.

No company in the evidence put the controls in before the rollout; every published account runs roll out, surprise invoice, then controls. The advice here comes from what that order cost them, and it works whether a company is planning its rollout or already has tools running.

What would prove this wrong. A company (roughly 1,000 to 10,000 employees) that gave AI tools to most employees with no cost-per-result target and still showed a measured financial benefit without restricting access. Evidence that typical spending stays flat as use deepens. A vendor default change that lowered the cost of the same work as often as it raised it.

In brief. The companies that kept their value did one thing the others had not: they put a single control layer between the company and every AI vendor and ran their cost control from there. Large companies built it. A company can buy it, and choosing it well is the most consequential decision in the plan. From that layer they controlled a cost per result (per software change, per booking, per document), changed which model answers by default and how much it reads rather than taking the tool away from people, made one person responsible, and showed the invoice to the finance team every month. Every one did it after the invoice arrived, and every one came out using AI more, not less: four times the users on the best tools, 30% more software shipped, usage near a company high at half the spend. Two software companies published the numbers; no company that only buys AI has published both halves of its own, which is why the two numbers have to be set in-house.

The detail

The companies that kept their value reduced cost by changing the settings, changing which model they used, or accepting lower profit, rather than by restricting who could use the tool. Uber and Priceline added spending caps, and Priceline relaxes its limits for work that brings in revenue. One company reversed, in one division. These are the practices, in their words, starting with the one that makes the others possible. The invoices they were responding to are in section 05.

They put one layer between the company and the AI vendors. The companies with the clearest results did it from one place: a piece of software the company controls, sitting between its people and every AI vendor, through which every request passes. Coinbase's engineers built a gateway that "routes everything" through a single point: it chooses the model, reuses material already read, keeps a record of who spent what, and applies redaction, logging and cost limits "before anything reaches a vendor." Man Group built its around the firm's data and a different model for each kind of task. OutSystems, a software company with about half its customers in Europe, built a gateway its chief executive calls an LLM router for its own engineers after its token spending peaked in June and July; by October the company was spending less than it had forecast for the quarter, "because of optimizations largely," and he said most of what it had been doing with AI "we could do just as well, a lot cheaper with a better harness." The arXiv paper on AI agent software explains why the layer matters: the software the vendor bundles already decides which model answers, what it reads and how much it reuses, so "it picks the rate on the price sheet and sets the volume bought at it," and a company that keeps the defaults "inherits these choices and their bill."

Most companies have only part of one. KPMG's global survey in September found security and data-access controls the most common capabilities in companies' AI control layers, at 43% each, and model routing, "which decides which model handles which task," the least common, at 23%. KPMG's reading: "Most organizations appear to have built a gate before they built a switchboard. A gate lets AI be deployed. A switchboard decides where work goes, and that is where orchestration and consumption economics begin." Among companies reporting an established return, 41% have the switchboard.

Large companies built theirs with engineering teams the size of Coinbase's. A company does not have to. The vendors bundle their own version of the layer with their products, and the arXiv paper proposes a ladder for deciding whether to rent that, buy an independent one, or build. This is where the evidence ends and the position begins. Done well, one layer gives a company four things at once: its data kept off public AI services, through the redaction and data-access controls that KPMG found are the most common capabilities companies build first; the freedom to switch models as prices and capabilities change, without rebuilding what sits on top; AI agents that can be moved to a different model when the market moves; and one place where the cost of every agent and every team is visible against a limit. Without the layer each of those is a separate project; with it they are settings. The wrong choice locks a company into one vendor's models and prices for as long as the AI runs. The right one is only as useful as the numbers it is set to enforce, which is why it comes after the two numbers in the plan, and why which layer, built or bought, set against which targets, is a different answer for every company.

They controlled the cost of each result, not the number of people allowed to use the tool. Booking Holdings, with roughly 9,000 technology staff, measures its software teams on total technology cost per finished software change, one number combining staff salaries and AI usage charges. Its chief financial officer told Fortune that AI spending can go up as long as the cost of each software change that goes live comes down, and those teams have increased the changes going live by about 30%, counting only changes that passed quality checks. That is the rule: one cost-per-result figure, staff plus AI, with total spending free to rise as long as the cost of each result falls. The evidence is one conference talk with no dollar figure, and Booking is far larger than a company; it is the approach to copy, not a comparable company. Uber's chief technology officer, four months after the company had used up its annual budget for AI coding tools: the price Uber paid per token fell while the number of employees using the most capable AI tools quadrupled, "Not because we've restricted access, but because we've treated efficiency as an engineering problem rather than a budget problem." Fortune's caveat stands: Uber did not say whether the total invoice fell.

They changed the default settings instead of lowering the spending caps. Coinbase's AI spending fell by nearly half in a quarter while usage stayed near a company high. Its engineering lead: "91% of employees weren't hitting their usage caps. Instead of lowering caps, we set cheaper model defaults to cut spend." The share of material reused rather than re-read rose from 5% to 60%. Accenture found that 54% of AI requests go to a more expensive model than the task needs, and that 90% of more than 9,000 tasks it analyzed did not need the most capable model; its advice is to have central software decide, "not as guidance left to each developer's judgment."

They made one person responsible and showed the AI invoice to the finance team every month. Smartsheet made the team that manages its technology budget responsible for total AI spending, set warnings before people reach a limit, and published cost reports by department and by manager. Bristol Myers Squibb's technology chief prepared the CFO and board in advance to expect "pretty high token consumption," telling the Journal it would be "orders of magnitude higher than what we spend today"; the only case in the record where the cost was forecast rather than discovered.

The cheapest model is not always the cheapest. A cheap model that is bad at a task can cost more than an expensive one that finishes it. On long tasks with many steps, the most expensive model cost less in total than the next cheaper one; researchers confirmed this by recalculating the cost of about 10,000 real work sessions. At Docusign it ran the other way. And a model can be chosen for neither reason: Salesforce rebuilt its AI assistant on models made for specific jobs and published speed, not cost, about 26 thousandths of a second to decide how to handle a request against about 1,446 for an all-purpose model, saying outright that the result "isn't a smaller stack, or a cheaper one." A slow answer holds up the work that depends on it. So "can we afford the more expensive model here?" has an answer only against the financial result that AI project was set up to produce, and the answer changes as prices, models and the work change.

What cuts against the pattern. The order of events. Salesforce deliberately let its developers use Anthropic's AI without spending limits for six months, with a fallback plan ("worst case scenario, we would pull back"), and only then began choosing a model per task. It presents "roll out first, control later" as deliberate. That order worked for a company earning 34.3 cents of profit on every dollar of sales, and it is still the reason Salesforce gave for not raising its profit forecast (section 05). It says nothing about a company with a fixed budget and no AI product to sell.

Two companies published the numbers. Docusign, 7,044 employees, processes more than a million agreements a day; software decides which size of model handles each document and cuts what the model reads from the equivalent of a long book to a few pages. In August it published the results for the smaller model it trained for the job: roughly a twentieth of the cost of the large model it replaced, about eight times the documents in the same time, and 7 percentage points more accurate. Cost went down and accuracy went up at the same time; ratios with no dollar figures, from the company's own blog. Figma, 1,886 employees, told the Securities and Exchange Commission that the direct cost of delivering its product rose $32.6 million in one quarter, "primarily due to a $27.1 million increase in technical infrastructure and hosting costs relating to AI and increased usage of our platform by paid users," and the same report is the first to include a sentence about managing that cost: changes "intended to manage usage and optimize the efficiency of our AI-related spend." The $27.1 million covers both AI costs and the cost of serving more paying customers; the report does not say how much is each. Both companies sell AI, so a rising AI invoice is matched by sales income. A company has the opposite situation, an invoice for tools its own employees use with no sales income tied to it, and no company of that kind has published that invoice as a number; where it appears in a report to regulators, it is buried inside an existing expense line.

The number nobody has. How often does running cost overtake the result at a typical U.S. company under 10,000 employees? No source measures it, and this issue does not estimate it. The companies that publish a cost attribute no result to it; the companies that publish a result publish no cost. And mid-sized companies are the weakest at proving a result at all: 14% of companies with $100 million to $1 billion in revenue can show their chief financial officer a measurable result from AI spending, against 34% of smaller companies, 17% of those at $1 billion to $10 billion, and 23% above $10 billion. Global survey; countries not disclosed. Until a company puts both halves in one document, nobody knows, and a company that sets both numbers for itself will be the first to.

The same choices, made before the invoice. The companion guide to this issue scores three invented AI ideas at a 1,200-person professional services company this way: customer support assistance, with a result already defined and a cost that grows with every request; reporting automation, cheap to run but with no baseline measured; and help writing sales proposals, worth eight people's selling time once the pricing rules are fixed. The order of funding came out of the scoring, before a vendor was in the room. A company that does not score its ideas this way discovers the cost, and the lack of a result, when the invoice arrives.

In brief. Five questions to ask about one AI investment before it gets budget, and seven steps that keep the two numbers together for as long as the AI runs. The two numbers are not a one-time approval check; the model, the price, the users and the work all keep changing, so the comparison runs from the first day and is redone on every change. Kept that way it produces one figure the board has not had: what AI returns after its running cost. The CEO names the target outcome and the owner, the CFO sets the cost per result and receives it monthly, the IT leader chooses the control layer and the settings that hold it, in that order. The comparison starts the day the AI goes live, but a project is judged against its cost per result only once it is out of its development phase, with a dated end and a spending cap, because the cost is real from day one and the result takes time to stabilize. Seven steps, two numbers, and a one-page monthly report.

The detail

The test. A company that can answer all five has a result it can defend and an invoice it can explain.

A company that can answer the first question and not the second has a result the invoice may wipe out. A company that can answer the second and not the first has a cheap tool and no idea why it bought it.

Development first, then production. Two things make the division harder than it sounds. The invoice stacks a variable charge on a fixed base (fees per user, the platform, the control layer itself), so a project is charged only the variable cost it causes; the base is the cost of having AI at all, reported once. And the result is unstable while a project is being tuned, while the cost is real from the first request, so a project measured in its second month looks like it is losing money even when it is on track. So every project gets a development phase with a dated end and its own spending cap, during which the cost per result is measured and reported, not judged. From the date it moves to production, the three verdicts in step 4 apply. How long that takes differs by project, so the plan sets the date per project. Nothing in the evidence says how long a typical project takes to reach a stable cost per result; this is the position, not a finding.

Why the numbers have to be kept, not set. A new default model, ten times the users, an agent that now reads twice as much: each moves the cost after the plan was approved, so the two numbers are compared from the first day and rechecked whenever anything changes. Kept that way, the comparison produces one figure the board has not had: what AI is producing for the company minus what it costs to run, by project, added together, month by month. Beside it, what the company is doing with that value: some new revenue, some costs eliminated, and most of it, in the evidence so far, hours saved, which the company has to reassign to other work. Reassigned to what, and was that measured? That is the part of the value most often lost.

Who does what. 75% of board audit committees rank oversight of AI among their top three priorities, and it is the area they feel least confident overseeing, at 56%. Finance is responsible for AI costs and returns at only 5% of companies. Companies where one named person is responsible for the costs and returns of AI are 3.7 times more likely to be able to show their chief financial officer a measurable result; among the 12% with no such person, none could. So the order is by seat. The CEO names the target outcome and the owner in the same sentence, before the budget is approved. The CFO defines the cost per result at the planned level of use and receives it monthly next to the result. The IT leader chooses the control layer and sets the defaults that keep the cost per result within the limit. A company that starts with the settings ends up with a smaller invoice and no result to point to; a company with tools already running starts with whichever is missing.

The plan. Every company in this issue that kept its value did some version of this after the invoice arrived. The plan is to do it before, or to wrap it around what is already running; step 7 is where that starts.

The answer the evidence in the next two sections gives to the question that opened this issue: at the companies that have gone furthest, the invoice did eat into the value where nobody had written down the two numbers, and the ones that kept their value controlled the cost per result, not who could use the tool, and kept expanding. Nobody has measured the return for a typical company, which is why the two numbers have to be set in-house. Seven steps, two numbers, one page a month. It is less work than a surprise invoice, and it is the difference between AI that costs money and AI that pays.

In brief. These are the invoices the practices in section 03 were a response to. Every one with a company's name on it came after adoption worked: a budget used up by April, a renewal that came back at four to five times the price, AI usage costs given by one company as a reason it did not raise its profit forecast for the year. Across the market, one large company in three uses up its annual AI usage budget before the year ends. Two things are true that sound like opposites. Companies invested less in AI overall than they predicted a year ago; the big plans came in small. But where they did put AI to work, the usage charges ran past the budget. The surprise invoices hit the companies that have gone furthest first, in the spring; by autumn they had reached filings, profit forecasts and board surveys. And cost is not the reason the typical company has held back; most have not reached the point where the invoice matters yet.

The detail

Uber used up its 2026 budget for AI software-writing tools about four months into the year, and a $1,500 monthly cap per employee per tool followed. The harder problem was not the money. It was that Uber could not say what the spending had produced. Its president and chief operating officer, Andrew Macdonald, in a podcast interview reported by Fortune: "Maybe implicitly there's more that is getting shipped, but it's very hard to draw a line between one of those stats and 'Okay, now we're actually producing like 25% more useful consumer features.'"

Priceline. When its contract for an AI coding tool came up for renewal, the new price was four to five times higher, an employee told TechCrunch; the vendor's usage reports did not match the company's own data, and usage limits on certain teams followed. "With AI, you're putting the credit card in the hands of the end user," its senior director of IT finance told the Wall Street Journal. "If you have no control over that, or if the end user is not educated enough, they're going to run up that tab." Microsoft cancelled AI coding tools it was buying from another vendor for one division, the only case here where a company took a tool away entirely; cost is the stated reason, and Microsoft sells a competing product.

Salesforce is the one company where AI usage cost affected a figure investors track: its profit forecast. Its deputy chief financial officer told an investor conference on August 27 that about six months earlier the company "unleashed Claude in our R&D cycle," giving its product developers Anthropic's AI without limits, and: "It is part of the reason we did not raise margin guidance on the year is because we are covering some of the token spend that we have got going." The results announcement filed the day before kept the full-year target for profit as a share of revenue at 34.3%.

Across the market. "Every CFO got the token bill in February and March," the chief executive of OutSystems said in an October podcast interview. "It was an oh shit moment for the whole world." One in three large companies use up their annual AI usage budget before the year ends; only 53% can tell which team or person generated the usage. One chief technology officer, in an account relayed to TechCrunch by the head of Faros AI, described an engineer who spent $40,000 on AI usage in a month and said he genuinely did not know whether to stop him or tell everyone else to be like him. That is the problem in one sentence. Without a stated financial target, a large invoice and a large success cannot be told apart.

How many pulled back. In KPMG's mid-2026 global survey, 49% of companies had scaled back, paused or questioned projects using AI agents (AI programs that carry out multi-step tasks on their own) once the expected cost began to outweigh the value. A survey of what people say, from a sponsor that sells governance work; it confirms the direction, not the figure.

The evidence on the other side. Total AI investment came in below plan: a year ago 35% of U.S. senior leaders at companies investing in AI expected to be spending $10 million or more by now; only 23% are. Companies did less than they planned, and the part they did ran over. At companies already paying for usage-billed AI tools, more say the cost led them to use AI more (37%) than to cut back (15%). And cost is not why the typical company has held back: among U.S. businesses not planning to use AI, "too expensive" ranks eighth of eleven reasons, at 6.9%. The invoice has hurt the companies where the most employees use AI the most. The named cases arrived between February and June; from August the same costs reached a quarterly filing (Figma), a profit forecast (Salesforce), and the surveys boards read in budget season. It reaches the rest in the order they adopt.

In brief. AI products charge a small fixed fee per user and a large variable charge for usage. Four things drive the usage charge: the default model, how much background material the AI reads each time, how much extra work each new vendor release makes it do, and AI agents left to work without a person watching. The usage included in the monthly fee runs out earlier in the month as adoption grows, and overage starts; a background agent can do this with nobody watching. The price per token is falling and the invoice still rises, because the price cuts apply to older models, newer models do more work per request, and companies use the savings to do more with AI. The only way to judge the numbers is one AI project at a time: what that project costs to run against the result it was measured to produce.

The detail

Where a vendor charges a monthly fee per user, it also charges for usage on top, and the fee is the small part: Anthropic's business plan is $20 per user per month plus usage at standard rates.

Four things drive the usage charge, and the company controls none of them unless it changes the defaults.

The default model. The software that runs an AI agent, breaking a job into steps and sending each to the AI, is usually supplied by the vendor. It decides which model answers, what it reads, whether it reuses material already read, and which extra AI helpers it starts. The vendor's software chooses both the rate and the volume, and a company that leaves the defaults in place accepts both.

How much the AI reads. In AI coding tools, reading is where the cost is: re-reading the company's existing code accounts for about 83% of the spending, while the new code the AI writes is 0.4% of what it processes.

Volume, driven by the vendor's releases. Companies in Accenture's survey expect the amount of AI they use to rise 78% over two years while the price per token falls 19%. Our arithmetic on those two figures: the invoice ends up roughly 44% higher if nothing else changes.

Agents left running. In the longest 1% of AI coding sessions Jellyfish measured, sessions with little human supervision produced about eleven times as much usable code as sessions of the same length where a person checked each step. Unsupervised AI work produces the most output, and runs up usage charges fastest.

Caps, allowances and overage. A spending cap does not reliably stop spending. OpenAI's pricing page says the AI stops answering once a monthly budget is reached, then adds: "There may be a delay in enforcing the limit, and you are responsible for any overage incurred." Google's Gemini accounts are moved automatically to a higher spending level, so a $250 limit becomes $20,000 to $100,000 once an account is 30 days old and has paid $1,000; Google confirmed this. Most plans bundle a set amount of usage into the monthly fee and charge for everything beyond it; Anthropic's plans draw all of a user's activity "from the same pool," reset on a rolling five-hour window with weekly limits on top, and at the limit the choices are to wait, move up a plan, or "turn on usage credits to keep working at standard API rates," that is, pay for overage. The vendor sizes the allowance; adoption uses it up.

Nobody is watching the agents. An AI agent running in the background, with no one at a screen, has no one to notice its overage. Anthropic bills its managed agents at $0.08 per hour of active runtime on top of token charges, so the clock runs whether or not the work is useful. One company's prepaid credits ran out silently and $30,141 of charges accrued before its cost-alert tool said anything. A background agent that quietly adds an overage charge every month is the most ordinary form this takes.

Why falling prices do not fix it. The price of a fixed level of AI performance falls roughly 13-fold a year, and the price U.S. businesses actually paid per million tokens fell 41% between March and September 2026. The invoice still rises, for three reasons. The cuts apply to last year's level of performance, so the newest model, which is what vendors sell today, has not been discounted yet. The work the AI does per task grows faster than the price falls: the cost of running the best available model rises 3 to 18 times a year. And companies spend the savings: nine in ten executives say cheaper usage prices would lead them to use AI for more things rather than spend less.

So a pilot's invoice predicts nothing about the invoice after a rollout: one model, a few people and short sessions, against the vendor's default model, all employees, and long sessions nobody is watching. The only way to judge it is project by project: what this AI project costs to run, against the result it was measured to produce. That is what cost control means in this issue. It does not mean deciding who may use the tool.