The ActivTrak 2026 State of the Workspace report says 80% of employees now use AI tools at work. The story beneath that headline is harder: for all the adoption, the financial impact isn't showing up. The companies spending the most have started asking why.
When AI costs more than your engineers
Uber exhausted its annual AI budget four months into 2026 and has begun to question whether it is getting meaningful returns. President and COO Andrew Macdonald, quoted by Jess Weatherbed in The Verge:
"We're going to have to start talking about token consumption and the associated cost versus headcount. So if you're not actually able to draw a direct line to how much useful features and functionality you're shipping to your users, that trade becomes harder to justify."
The numbers underneath that quote are significant. Hillary Remy in TheStreet: "Uber burned through its entire $3.4 billion 2026 AI budget in four months."
A 2025 Mavvrik survey, reported by Lucas Greene in WebProNews, found that 84% of companies report AI spending has eroded their gross margins. Greene's write-up reveals the pattern bluntly:
"AI was supposed to reduce expenses. Instead Microsoft and Uber discovered it often costs more than the human workers it augments. Compute spending now surpasses payroll in some cases."
A similar observation lands harder coming from inside Nvidia — the company that sells the GPUs powering much of the AI revolution. Bryan Catanzaro, vice president of applied deep learning, told Sasha Rogelberg at Fortune (echoing an earlier interview with Axios):
"For my team, the cost of compute is far beyond the costs of the employees."
Microsoft's reversal — at least inside one product group — builds on the same theme. Remy reports that Satya Nadella now claims 30% of Microsoft code is AI-generated. Microsoft has invested $13 billion in OpenAI. It built GitHub Copilot. It has made AI central to every product in its portfolio. And, from Remy's piece:
"Claude Code launched inside Microsoft's Experiences and Devices group in December 2025. Six months later, the experiment is ending. The tool was not canceled because engineers disliked it. It was canceled because they used it too much."
The tool worked. The engineers loved it. The math did not.
Even Altman is dialing it back
Sam Altman, who did more than any other executive to plant the jobs-apocalypse story (and before that, the "AI is dangerous" tale), has started distancing himself from it. From Rebecca Schneid in TIME:
"I don't think we're going to have the kind of jobs apocalypse that some of the companies in our space advocate or talk about... I thought there would have been more impact on entry-level white-collar jobs being eliminated by now than has actually happened."
The CEO most responsible for the rhetoric is now walking it back. Read together with the cost numbers, this is the early shape of the hype unwind: the productivity claims were oversold, the cost claims were undersold, and the executives writing the checks have begun to notice.
Looking beyond the smoke
None of this says AI doesn't work; just as layoffs blamed on AI productivity didn't prove that the companies got value from the tools. It means AI deployed without diagnosis does not work, and "let everyone in the company use it for everything" is the deployment most large organizations default to.
Worse, some of these organizations pushed "token maxxing"; they encouraged their teams to use AI as much as possible. Did it matter how or why they used AI? No. The goal was to spend as much as possible.
When AI has no defined job, two things happen. Engineers run it wide open because it is interesting — and because the company is rewarding them for running it. And the spend dominates the line item before anyone has connected a single dollar of compute to a single dollar of output. That is the trap Uber and Microsoft just found themselves stuck in.
The companies actually getting returns from AI right now do not look like that. They look small, narrow, and boring on the outside. Each AI deployment starts with a hypothesis, and the results are evaluated against business goals rather than some disconnected metric.
For more on why most companies stall at surface-level AI usage, see The AI Adoption Arbitrage: Why Most Companies Are Stuck.
The same logic, outside business: imagine you want to improve your health. Where do you start? You could exercise, you could improve your diet, but this line of thought starts with the solution rather than a diagnosis. Maybe you have food allergies or poor balance. Those insights could drastically change how you improve your health.
These big companies aren't sure what their problems are. Sometimes they don't even have a hypothesis. In the absence of a diagnosis, an AI solution might worsen the undiagnosed problem.
What works instead
A few patterns from the past year of advisory work and weekly AI Roundtable conversations:
Narrow, defined automations. A SaaS company replaced its cold email SDRs' first-touch appointment setting with an agent that handles inbound replies, proposes meeting times, and escalates to a human when prospects sound angry. One job. Clear scope. The agent stays in its lane because the lane is drawn.
Another company embedded a small agent in their cancellation flow: when a customer hits "cancel," a chat opens whose only task is to surface the root cause and offer a single counteroffer. It does not perform the cancellation. It captures the reason, makes one offer, and hands it off to a human. The cost is negligible compared to the revenue it supports. The retention signal is real.
Foundation-first deployments. A team modernizing a 20-year-old .NET codebase built a domain-vocabulary file first — every service named with the team's own aliases, each one's responsibilities documented — and fed it to Claude as project context before letting the agent touch a single line of code. AI refactors and bug fixes now speak the team's language. The alternative is an agent inventing names and patterns that no human on the team recognizes, which is how most enterprise AI deployments quietly accumulate debt.
Compounding skill chains, not chat windows. Years of recorded sales call transcripts are ingested as a single input. A pipeline extracts the actual phrasing of each prospect's pain point for each market segment. Those phrases become segment-specific landing pages and long-tail keywords inside Google and Meta ad campaigns. Five separate small skills, each doing one thing, chained together, with a human reviewing before launch. The total compute cost is a rounding error. The output is ad copy a marketing team would never write by hand because no individual segment is big enough to justify the work, but a pipeline that targets all of them at once is.
Defensive guardrails before access. A team running
multiple Claude Code sessions added a pre-commit hook that blocks the
agent from reading standard secrets files — .env, AWS
credentials, anything sensitive — even when otherwise permitted. Another
operator runs his agent environment on a separate, blank laptop with no
personal data, so the agent can have broad permissions inside its own
sandbox without ever touching his primary machine. Cheap to set up.
Saves you from the news story you do not want to be in.
A coaching habit, not a tool. One founder starts every 1:1 with his developers by asking what they could not use AI for that week. The blocker is invariably a mental-model gap, not a real limitation. The 1:1 becomes a working session on how AI could have done it — write the Postman requests as a script, ingest the support tickets, and produce a categorized list. The cost of the habit is zero. The compounding effect on adoption is enormous.
The same prompt shows up in Claude Team vs Pro vs Enterprise: How to Pick a Tier as the conversation-not-dashboard test for whether AI is actually creating value.
None of these is flashy. None requires a $3.4 billion budget. None justifies a press release. They share a structure: a specific job AI is doing for a specific reason, with a human staying close to the judgment that matters.
AI Workshop for CEOs
The hard part is the diagnosis — identifying which gap is worth closing first. The tool choice follows. Three hours live with eight CEOs, plus a 1-on-1 follow-up — we work through the diagnosis together before talking about deployment. Money-back guarantee.
Reserve Your Seat →Diagnose before you prescribe
Most companies do not have an AI spending problem. They have a diagnosis problem. The hype cycle convinced leadership that AI was a horizontal capability — like cloud, or like the internet — that you adopted across the organization and let people sort out. Cloud and the internet are horizontal. AI in 2026 is not. There are dozens of vertical applications for AI, each of which has to be matched to a specific opportunity for a specific team. You're looking for gaps where AI can accelerate a process that's already creating value, or where it enables a new opportunity.
For the mapping exercise that has to come before deployment, see Map Your Operating System Before You Apply AI.
If you do not know which opportunity you're targeting, no amount of AI spend will create value. You will get the Uber result: large numbers, no signal, and a CFO asking when the productivity shows up. You've just used a new tool to expand the hidden gap between productivity and value.
AI is the most useful new tool for creating value in a generation. The companies winning right now treat it as exactly that — a precision tool, applied thoughtfully, with humans staying in the judgment loop.
You do not need more AI. You need better questions about what specific value it can create.