More Costly Than Tokens
AI bills are coming due. The real costs never appear on the invoice.
In April, four months into the year, Uber exhausted its 2026 budget for autonomous coding tools. One of the most closely monitored engineering organizations in the world, and it never saw the bill coming.
It exposed the habit underneath. Push your people to use AI, and if you’re like most companies, you won’t measure whether the work got better. You’ll measure how many tokens they burned.
Tokens are the small chunks of text a model processes: the words and fragments that go into a prompt and come back as a response. They’re also the unit a finance director can count. And once you can count something, you can budget it, rank it, reward it, and ration it.
The Wall Street Journal named the practice in an April 2026 article, “Why Some Companies Say AI Tokenmaxxing Is Key to Survival.” Tokenmaxxing became a flashpoint: whether burning more tokens produces useful work, or just a bigger bill.
The bills came due. Meta killed its token leaderboard, Microsoft canceled its Claude Code licenses, and Salesforce started reining in the spending it had gamified months earlier.
But rationing treats the bill, not the force behind it. And that force is older than AI. In 1865, William Stanley Jevons noticed that more efficient steam engines didn’t cut Britain’s coal use. They raised it. Cheaper power made coal worth burning for more things.
Jevons proved that cheaper doesn’t mean less, so rationing is fighting a structural tide, not fixing a leak.
AI is the same loop. The savings get poured back in: more work routed through models, more agents running, more ideas tested, more steps automated. Microsoft CEO Satya Nadella said as much in January 2025. As AI gets more efficient and accessible, he predicted, its use will “skyrocket.”
Conflicting Priorities
The token bill hides a structural problem. Deploying AI carries four priorities, and they don’t all point in the same direction. Three serve the company. The fourth, adoption, serves the vendor. It’s measured in installs and tokens, and it’s the one the labs and their engineers push.
The bill is only the receipt. The question is where waste enters the system.
It starts with who owns the work. AI deployment now has four owners: the executive team for governance, business units for impact, HR for engagement, and IT for adoption.
Treat it as an adoption play, and the failure modes that matter multiply. Agents run ahead of controls. Pilots never reach the P&L. Employees spin up their own tools. Tokens get consumed, but the work still doesn’t ship.
The token bill is the tax you can see. It isn’t the tax that’s killing you. So why do companies fixate on the vendor’s incentive and ignore their own?
To start, adoption has a massive sales force behind it. Look at the glide path into global corporations. OpenAI’s roster includes Bain, Capgemini, and McKinsey. Google works with Accenture, Deloitte, and PwC. The consulting firms once hired for advice are now co-owners of the solutions. The referee and the players share the same financial interests.
Labs will say their teams do more than set up agents. They redesign workflows and improve operations. The best ones do. But there’s a limit to what incentives allow.
A forward-deployed engineer has no reason to tell you that a task shouldn’t be automated, that the work might be better on a competitor’s model, or that the real problem is a senior leader who doesn’t trust the results and treats the tool like a search box.
We call this the Trust Deficit. It was the top problem in our Ground Truth Index last quarter, and no deployment team will solve it.
Agent logic, human logic
In 1987, MIT economist Robert Solow looked at a decade of corporate computer spending and flat output, and said, “You can see the computer age everywhere but in the productivity statistics.”
The same gap is back, faster, and more destructive. In a survey this April from the AI company Writer, 54% of executives said AI adoption was “tearing their company apart.”
Here’s the part no deployment deck will show you: AI is a machine, and the work it lands on is run by people. The two run on opposite logic.
AI goes nonstop, at machine speed, spitting out likely answers it can’t be blamed for. People move in steps, expect the right answer, and maintain a human pace.
You don’t wire those two modes together overnight. Companies smash them and call the wreckage a learning curve. Better models don’t close the gap. They widen it. What closes it is the slow work of rebuilding the organization around what the machine can do. That rebuild has precedent.
When factories first switched from steam engines to electricity, they kept the same setup, using a single central motor to power everything through shafts and belts, just as with steam.
The economic historian Paul David found that it took nearly forty years for productivity to improve, and only after factories were redesigned: a separate motor for each machine, the layout built around the work. The problem was never the electricity. It was the shape of the building.
General Motors ran the same experiment in the 1980s and lost. CEO Roger Smith poured tens of billions into automation, including a showcase plant in Hamtramck designed to operate with minimal staffing. The robots performed worse than the people they replaced. Some welded doors shut. Others painted each other instead of cars.
GM was running the control group, too. With Toyota, it operated a plant in Fremont, California, called New United Motor Manufacturing, or NUMMI.
It rehired the same workforce GM had written off, used fewer robots than Smith’s showcases, and built what Smith never bought: clear roles, teams that owned their work, and the right of anyone to stop the line the moment something looked wrong.
GM bought adoption. NUMMI installed a new operating system.
Absorption limits
Economists have a name for what NUMMI built. In 1990, Wesley Cohen and Daniel Levinthal called it absorptive capacity: a company’s ability to take something new in and put it to work.
Two firms buy the same technology. One compounds it, the other chokes on it, and the difference is never the technology. Michael Kremer’s O-Ring theory later showed why the gap turns brutal.
In real production, quality multiplies; it doesn’t add up. A brilliant model dropped into an organization that can’t take it in is a flawless part bolted to a failed seal. The capability is real. The return is still zero.
Capability is no longer the bottleneck. Intelligence shows up pre-built and abundant, cheaper every quarter.
What can’t be rented is an organization ready to absorb what the technology produces.
So the constraint moves. It drafts off the technology the labs ship in volume and lands on the one thing they can’t ship: the absorption limit of the company writing the check.
Value = Capability ✖️ Absorption
Adoption sits at the bottom row of the priority board. Absorption is in the other three.
Raising that limit is the work nobody’s put up for sale. We call it the Human OS, the work-orchestration layer the labs don’t sell. It isn’t a transformation program with a finish line. Model companies ship a new version every month, so there’s no go-live, only a layer that keeps adapting. That’s why it’s an operating system, not a project.
It draws on more than two hundred failure patterns we’ve logged in the field, a library we call the Grove. Six parts, each best understood by the failure it heads off.
A work atlas, because you can’t redesign work you’ve never honestly mapped, workarounds and all.
Task triage, because skipping it and automating whatever sits closest to the data is a faster version of work someone should have deleted.
Decision rights, because speed without clear responsibility doesn’t remove the liability. It just moves it faster.
Agent governance, because a company running agents across its inbox, its finance system, and its service desk, each with its own memory, hasn’t ended fragmentation. It’s been rebuilt in software.
Role redesign, because the agent takes the task but not the job. Someone still has to own the agent, catch its mistakes, and answer for them. If you don’t define that role, people improvise it badly.
Team belief, because the worker who thinks the agent came to replace her will quietly starve it, and the one who believes it came to make her better will teach it. That belief is the whole return.
Six parts, one job: human requirements the labs leave empty.
When engineers leave the building
Consider a claims executive I’ll call Maria. She ran her group at an insurance carrier for nineteen years. For nine weeks, she watched a team of engineers build an agentic system for her operation. It read claims and drafted decisions. It flagged the odd case for a human to check. Work that used to take four days closed in minutes.
On their last day, the engineers walked her through it, and she asked the only question that mattered: What does my team do on Monday? The answer was runbooks, documentation, and a Slack channel.
Maria saw the gap that the engineers couldn’t. Her company had bought an adoption program and hit its absorption limit the moment the engineers walked out. Nobody owned the agents. Nobody had redesigned the work around them or set the decision rights. Nobody had made the case to her people that the system came to make them better, not to make them unnecessary.
The buy was the easy part. It always is. The hard part was the question Maria asked on the last day, the one the engineers had already packed up to avoid. What does my team do on Monday?
You can count every token a company burns. You can budget them, rank them, ration them. You can’t count whether Maria’s people will teach the system or quietly bury it, and that, not the token bill, is what decides the return.
The token receipt prints on time, to the cent. The bill that matters never prints at all.
It comes due Monday, on the floor where the work gets done.
Chris Perry is the founder of Andus Labs, which builds the Human OS for AI, the work-orchestration layer across people and agents. Ground Truth, the company’s quarterly intelligence release, draws on more than two hundred documented patterns of enterprise AI breakpoints and their fixes.