There's a story moving through the enterprise world right now, and Uber is its poster child. The company rolled AI coding tools out to roughly five thousand engineers in December, watched adoption climb past eighty percent by March — and exhausted its entire 2026 AI budget by April. Four months. Not because the tools failed: seventy percent of Uber's committed code was coming from them. Because nobody could draw a line from the tokens to the value. The COO said it plainly — that link is not there yet. And Uber isn't alone. Microsoft revoked internal AI coding licenses months after enabling them. Amazon killed an internal usage leaderboard after employees gamed it with junk prompts. A consultant told Axios one client burned half a billion dollars in a single month after nobody set usage limits. The word for the moment is "tokenmaxxing," and the mood is betrayal.

The numbers underneath the mood are just as real. MIT's NANDA initiative found that 95% of enterprise generative AI pilots deliver no measurable P&L impact — against $30 to 40 billion in spending. BCG puts it at 74% of companies unable to achieve and scale value.

The vendor's side of the table

Let me say what everyone in the room is thinking. Of course the vendor doesn't know your business. His job is to sell capability. Tokens, seats, licenses — the name changes by wave. The incentive doesn't.

I can tell you this with some authority, because I spent years on that side of the table. I was at PTC when computer-aided design was the wave. I was at Autodesk when design platforms, data management, services portfolios, and subscription economics were changing how customers bought software. I've sat in those rooms.

The vendor will tell you a great story about partnership, and sometimes he means it. He may even have a services team that gets your pilot working. But that team is built from technical people, and the vendor can make the product work — he can't make your company work differently. That was never his job. The salesperson across the table isn't a bad person. He's a person with a quota, and the quota is not your outcome.

I know because I sat in the deal reviews. I watched engagements arrive scoped by the sales team and rubber-stamped by delivery — custom code with no development discipline behind it, no testing plan, no answer for the customer's IT organization when it asked about compliance or security. I watched enterprise capability get sold on a product that couldn't yet carry it at enterprise scale, and I took the call from the executive who found out. And here's the thing: nobody on our side was lying to him. Most of us couldn't see the problem either. The opacity ran all the way up.

That was the story at PTC. It was the story at Autodesk. It is the story inside the companies selling AI capability now.

The fifth time I've watched this movie

But here's where I get off the bandwagon, because the backlash is making the same mistake the buying did.

This is the fifth time I have watched this movie. ERP created real leverage where the work could be standardized — finance, procurement, inventory, reporting, the order book. It did not, by itself, make the shop floor, planning discipline, data governance, or exception handling work differently. CAD changed how products were designed; the engineering-change process, release discipline, and downstream handoffs often lagged behind it. SaaS and cloud did the same: real leverage where the work changed with the platform, disappointment where the platform was laid over the old company. Every wave was real. Every wave was sold as more pervasive than it was. And every wave punished buyers who bought capability, kept the old operating system underneath it, and never priced the long-term carry. Then, when the return didn't show up, they blamed the vendor.

This time, the industry has put two numbers on the mistake — and together they're damning. BCG's research, since corroborated by other major firms, found that successful AI deployment is 10% algorithms, 20% technology and data, and 70% people and processes. And in December, Deloitte's chief technology officer reported the spending side: companies are pouring 93% of their AI budgets into technology and 7% into the people expected to use it.

Put those two numbers next to each other. Success is 70% people and process. Spending is 7% people. The market is funding the part the vendor sells and starving the part only the business can change — and then calling it betrayal when the spend doesn't convert. It isn't betrayal. It's an operating failure with a purchase order attached. Deloitte's CTO called it buying the ingredients while ignoring the recipe. He's being polite. Every board deck gets excited about the 10%. The 70% isn't in the pitch because it was never the vendor's to sell.

Here's the mechanism, and it explains the 95% better than any incentive story. The pilot succeeds because the pilot is designed to succeed. You give it your best people. You feed it your cleanest data. It has executive attention and organizational goodwill, because everyone wants the demo to work. Then you take it to the general company — to the average team, the messy data, the manager who didn't ask for it, the process that's been quietly held together by one person since 2019 — and it dies. Not because the model got worse. MIT's own diagnosis was exactly this: the failure isn't model quality, it's integration — the gap between what the tool can do and what the organization can absorb. The pilot tested the technology. Production tests you.

You can watch it in the workforce data right now. Deloitte's own research found that while access to generative AI keeps rising, actual usage fell 15%. Forty-three percent of workers admit they bypass the approved tools for unapproved ones they say simply work better. And the one group that bucked the trend — workers given real hands-on training — reported trust 144% higher than those without. The tools are present. The organization isn't absorbing them. That is the 7% bill coming due.

This isn't an innovation problem. It's an operating problem, and eventually it becomes an EBITDA problem. The dashboard says usage is up. The board deck says transformation is underway. The income statement says nothing changed. That gap is the work.

The vendor was never going to know your operations — that was never for sale. Outcome-based pricing won't fix that. A vendor can't price outcomes inside operations he can't see.

Before the indignation gets too comfortable, look at how you run your own business. Your sales team is paid on bookings, not on whether delivery holds margin. Your departments are measured on utilization, not on what converts. Uber didn't just get caught by the meter — it ranked its engineers on internal leaderboards by how much AI they consumed, while the teams driving adoption were never the teams answerable for the bill. The buyer gamified the vendor's revenue metric inside its own house. The vendor optimizing for his own metric isn't a betrayal of how business works. It is how business works — including yours. Don't be naive about it in others and blind to it at home.

The part your budget hasn't priced yet

Pilots look cheap while production usage compounds. A workflow that called a model once becomes a chain of calls — agents, retrieval, review, exception handling, monitoring — which is how a full-year budget disappears in four months. Unit costs can fall and the bill can still rise. The tool doesn't have to fail for the economics to break. It only has to be pushed into work that isn't converting — and to produce output faster than the organization can inspect it. For three years, falling unit prices bailed out undisciplined consumption. That bailout is closing.

There's already a number on this. Jellyfish, which measures engineering organizations, found the heaviest AI users were roughly twice as productive as light users — and burned ten times the tokens to get there. Twice the output at ten times the cost isn't a scaling curve. It's a conversion problem wearing one.

So do the arithmetic the vendor will never do for you. Take your AI run-rate — the real one, fully loaded. Divide it by your operating margin. That's the new top line you need just to pay for it. Two million dollars of AI spend at a 10% margin is twenty million dollars of new revenue. Is the work the tokens are doing producing that? Are you measuring it?

For thirty years, the cost of capability outrunning structure was invisible. It showed up late and laundered — as margin erosion, delivery slippage, modernization that landed but didn't convert. You paid for the absorption gap in every wave; you just never got an invoice. This time you do. The token bill is the first itemized, real-time statement of where your organization is pushing capability into work that isn't converting. Your predecessors in every prior wave would have killed for that signal. You're getting it monthly and treating it as a billing dispute. Read it as a map instead of a bill — Uber's COO, staring at a spent budget and asking where the features are, is reading the map. Most of the market is still arguing about the bill.

How you fix it in this wave

Find the places where the capability actually converts. AI can compress cycle time, extend scarce judgment, improve decision quality, and remove work that shouldn't require a person. It can also create noise, rework, false confidence, compliance exposure, and more output than the business can govern. Knowing the difference is the work.

Put a number on the broken process before you put technology on it. If you can't say what the current way of working costs, you can't say what the tool returned. You'll end up measuring usage, because usage is the number the system gives you.

Make absorption someone's job. Not a committee, not a center of excellence that waits for adoption to happen — a person with authority to change how work moves, where decisions get made, and what should be automated, augmented, or left alone. If nobody owns that, the company hasn't deployed AI. It's installed a meter.

The vendor sells capability. Absorption was always yours. It was never in the contract — and no pricing model will ever put it there.