A number from Uber this week, reported by TechCrunch, is worth sitting with. The company ran through its entire annual AI budget in four months. Not one department. The whole company, the whole year, gone by April.
The response was a cap. Fifteen hundred dollars per employee per month, per coding tool, tracked on an internal dashboard, with overrides available. Uber had spent the previous year telling staff to use AI as much as possible and ranking usage on internal leaderboards. The reversal took one budget cycle.
What the Cap Actually Admits
The interesting part is not the overspend. Companies overspend on new categories all the time. The interesting part is the reason Uber gave for not being able to justify it. The COO said it is very hard to draw a line between AI usage and measurable productivity gains.
Read that again. This is Uber. An instrumented company, a data culture, every ride a row in a table. If Uber cannot connect AI spend to output, the median company running Copilot and three other tools has no chance of doing it either.
The honeymoon assumption of the last two years was simple. More AI equals more productivity, so spend more. That assumption is now being audited, and the first audits are coming back inconclusive. The spending was real. The proof never arrived.
There is a second-order problem hiding in the leaderboard. The moment you rank people on AI usage, usage becomes the goal, and people optimize for the number. You get more tokens consumed and more tools adopted, which looks like momentum and costs like momentum, with no guarantee that a single deliverable improved.
The timing tells you something too. Four months is one quarter plus a scramble. The budget did not erode slowly as teams found value. It got consumed fast, in the rush to be seen using the new thing, which is the signature of adoption driven by fear of missing out rather than by a return anyone measured.
The Metric Problem Nobody Wants
The reason the line is hard to draw is that most companies measured AI adoption, not AI return. Adoption is easy. Seats, logins, tokens, lines of code accepted. Return is hard. Did the work get better, did it ship faster, did it make money.
I wrote before that your ROAS is lying to you once the auction is run by AI. The same failure mode is now inside the building. Teams optimized for the metric that was easy to read, and the easy metric was usage.
Building GEOflux and Madison AI, the question I get from operators is never whether people are using the tool. It is what the tool returned. Those are different questions, and only one of them shows up on a leaderboard.
Here is the version I see most often in client work. A team adopts an AI writing tool, output triples, and everyone celebrates the productivity. Six months later, nobody can point to a revenue line that moved, because the constraint was never the volume of content. It was the quality of the thinking, and the tool scaled the part that was already cheap.
The companies hitting Uber's wall are the ones that priced AI as a capability expense without a return model. Vendors priced per seat and per token, which rewards consumption, not outcome. I argued that SaaS pricing does not work for AI products because the cost scales with usage while the value does not. Uber just lived the buyer side of that mismatch.
What to Do Before Your Own Cap
The move is not to stop spending. It is to instrument return before the finance team forces the question. Pick the three workflows where AI is deployed most heavily. For each, write down the baseline before AI and the single metric that would prove it worked: cycle time, error rate, revenue per head, whatever is real for that workflow.
Then read the cold data. If you cannot find the line in three workflows you chose on purpose, you will not find it across forty you did not. That is not a reason to cut everything. It is a reason to concentrate spend where the return is legible and stop subsidizing the rest.
One more discipline. Separate the experiments from the infrastructure. Some AI spend is research, you are buying information about what works, and it should be capped and time-boxed on purpose. Some AI spend is production, it runs a workflow that earns its keep, and it should be defended. Most companies mix the two into one line called AI, then panic when the line is large.
Split it, and the cap stops feeling like a retreat. The research line is supposed to be spent and learned from. The production line is supposed to pay for itself. Uber's real mistake was not spending. It was never separating the bet from the engine.
Uber did not cap AI because AI failed. It capped AI because nobody could prove it worked. Those are different problems, and only one of them is yours to fix before the next budget cycle.