In April 2026, Meta engineers had an internal dashboard called Claudeonomics that ranked all 85,000 of them by Claude token consumption. The top user was on pace for 281 billion tokens that month. At Anthropic's API rates, that one engineer's consumption was running close to $1.4 million for the month. The leaderboard handed out titles like Token Legend and Cache Wizard.
The Information broke the story. Meta took the dashboard down inside 2 days.
This was not a Meta problem.
This was the rollout pattern that every large enterprise has been running for the last 2 years, finally on display in the open. Measure tokens. Reward consumption. Call that adoption.
Skip the part where someone in finance has to ask what the business actually got in return.
The bill for that usage pattern is starting to arrive, and a growing number of companies are pulling back on AI spend. Uber's COO told a podcast in May 2026 that the company had burned its full 2026 AI budget in 4 months, mostly on Claude Code, as adoption climbed from 32% to 84% of its 5,000 engineers between February and March 2026. Andrew Macdonald's quote was the one to circle:
If you're not actually able to draw a direct line to how many useful features and functionality you're shipping to your users, that trade becomes harder to justify. That link is not there yet.
Microsoft, in the same month, canceled its own internal Claude Code pilot effective June 30, 2026, after the pilot exhausted its annual AI budget. The official direction was to migrate developers back to GitHub Copilot CLI.
Forrester's 2026 predictions said enterprises will defer 25% of AI spend to 2027 and that only 15% of AI decision-makers reported any EBITDA lift in the prior 12 months. Gartner's standing forecast still has more than 30% of generative AI proofs of concept getting abandoned. S&P Global's 2025 enterprise survey put the share of companies abandoning the majority of their AI initiatives at 42%, up from 17% the year before.
The dominant explanation, in print and at the board table, is that AI did not deliver. The bills got too big. The productivity was a mirage. The technology was overhyped.
That explanation is too easy.
Companies measured the wrong thing
The metrics most enterprises stood up for AI in 2024 and 2025 were never outcome metrics. They were activity metrics. Active user rate, prompts submitted, tokens consumed, weekly active AI users.
Microsoft's 365 Copilot Usage Report ships with Active User Rate, Total Prompts Submitted, and Average Prompts Per User as headline KPIs in the admin center. The Viva Insights Copilot Dashboard exposes the same. GitHub Copilot's metrics API surfaces per-seat acceptance counts. Worklytics, the HR analytics vendor selling adoption scorecards, benchmarks weekly active AI user rates and tells customers that anyone below 45% adoption will "struggle to justify continued licensing."
This is what got measured because this is what was easy to measure. None of it tells anyone whether the work got better.
Shopify made it formal. The leaked Tobi Lütke memo from April 2025 told the company that reflexive AI usage was a baseline expectation, and that AI usage would show up in performance and peer reviews. Salesforce, per reporting in early 2026, told staff to use at least $170 a month in tokens or be flagged, then management walked the policy back. Salesforce's own leadership later conceded that "tokens are very clearly not the right estimation of value. You can burn a million tokens and create no positive outcome for your company."
Those metrics came with consequences at the user level. Most large Microsoft 365 environments now automatically flag Copilot seats inactive for 30 or 60 days and reclaim them. Third-party IT guides treat this as standard hygiene. GitHub Copilot seat dashboards surface "inactive" seats for the same purpose.
The message to the employee is blunt. Use it enough, or lose it.
What did companies think was going to happen?
Leaders told 85,000 engineers their continued access depended on activity. They gave those engineers a dashboard that showed each one's token rank against their peers. And then they acted surprised when one person showed up at the top of the list with roughly $1.5 million in consumption in a single month.
What the latest round of AI rollouts skipped
Set aside the leaderboards for a minute. The deeper failure was an absence, not a presence.
Almost no rollout I have watched, talked to peers about, or read a public writeup of, started with the question that should have been first. What is the work that needs to get better? What does better look like as a number, with a baseline, on a date? Who is accountable for moving that number?
In place of those questions, the rollouts I have seen substituted three others. How many seats can the budget cover? How fast can the tool get in front of people? How many of those people are using it?
All three measure rollout progress. None measure business impact.
That absence is what the high-credibility surveys are documenting. Forrester's October 2025 work found that fewer than one in three AI decision-makers can tie AI to a P&L line. McKinsey's State of AI puts the share of organizations seeing enterprise-level EBIT impact from generative AI at 39%, with around 5% getting meaningful margin. IBM's 2025 CEO study has 25% of AI initiatives delivering expected ROI.
MIT NANDA's GenAI Divide report, the one that produced the viral 95% number, blames the gap on what its authors call the learning gap, not on the models.
The MIT 95% has earned some methodology criticism, and the criticism is fair on the edges. The part of the report nobody seems to read carefully is the inside of the number. Vendor-partnered builds with clear use cases succeed at roughly 67%. Internal builds with no clear use case succeed at one-third that rate.
The pattern in the data is not AI failing. The pattern is AI rollouts failing for the same reasons IT projects have been failing since the 1990s. No outcome owner, no outcome metric, no baseline, no date.
Companies did this to themselves.
Tools without a blueprint
Stand 500 random adults on an empty concrete pad. Hand each of them a top-shelf set of power tools. Tell them to build something. Do not tell them what exactly to build.
Come back in a year. Very few finished structures would be standing. The ones that did go up would be poorly built. Most of the tools would look about the same as the day they were handed out.
A small number of adults would have built something solid, almost always because they had already decided what they were trying to build before the tools showed up.
This is the part of the AI rollout story almost no one wants to discuss. The tools are real. They work.
What was missing was the blueprint. What was missing was the answer to what was being built.
A senior engineer at Uber with Claude Code in their environment can ship features faster than a senior engineer without it. That much is defensible from any reasonable read of the productivity data. What is not defensible is the leap from "the tool is faster" to "the business is better off."
That leap requires the second half of the rollout. The half where someone wrote down what the team was trying to do that it could not do before, and where the improvement was going to show up.
Without that, the tool produces motion. Motion produces tokens. Tokens produce a bill. The bill produces a board meeting where someone asks why the company spent that much money on a tool they cannot connect to a P&L number, and the answer is some version of "the tokens went up."
The providers are not blameless
I would feel better about the rollout discipline conversation if the AI providers themselves had carried more weight on the enablement side.
Anthropic launched the Claude Partner Network in March 2026 with a $100 million commitment, plus alliances with KPMG and PwC, plus a Coursera course and a Skilljar curriculum aimed at enterprise enablement. OpenAI stood up a consulting arm and took a stake in Thrive Holdings to do something similar. Microsoft has an official Microsoft 365 Copilot adoption guide for IT admins.
That is something. It is not nothing. But the support infrastructure still does not match the sales pitch for a class of tools their own providers spent two years describing as the most consequential technology shift since the internet.
An honest enablement push from a frontier AI provider would not stop at a partner network and a course. It would include a published outcome-metric framework by industry, reference baselines, cost-model assumptions a CFO could argue with, and a strong view on which activity metrics companies should not track. None of the major providers has shipped that. The closest substitute has come from HBR articles written by people who are not on any of their payrolls.
The incentive is part of why. Token consumption is the providers' revenue line. A customer who learns to extract the same value from one-tenth the tokens is a customer who shrinks an account. That is the part of this story I would like more of the industry to say out loud.
If you sell a tool you call transformative, the obligation to teach the customer what transformation looks like is part of the sale. It is the difference between a customer who can defend the renewal and a customer who kills the project at the first quarterly review.
What the next round of AI rollouts should do differently
Start with the outcome. Pick the work you want to be better. Write down what better looks like with a number, a baseline, and a date.
Name a single human accountable for that number. Then choose the tool.
Stop treating activity metrics as evidence of value. Active user rate, prompts submitted, tokens consumed, and weekly active AI users are administrative metrics. They tell the IT organization that the tool is working. They tell the business nothing.
Do not put them on a leadership dashboard. Do not threaten employees' access against them. Do not pay a vendor a bonus based on them.
Communicate the vision the rollout is supposed to serve before the first seat is provisioned. Tell the team what the business is trying to do that it could not do before. Tell them what they are expected to produce that they could not produce before. The tool comes after the picture of the work.
Train people on the work, not the tool. The product demo is not the curriculum. The curriculum is the redesigned workflow the tool makes possible, and that takes more than a 30-minute lunch and learn.
Hold the line on outcome accountability when the bill arrives. The temptation when the bill is large will be to either declare victory based on activity metrics or kill the project because the outcome was never defined. Both are failures.
The correct response is to do the work that should have been done at the start. Define the outcome, measure what matters, and decide on the basis of the measurement.
The companies that get the next round right will not be the loudest about it. They will be the ones whose AI line item shrinks while the work it touches gets measurably better.
Their usage bills will be smaller than their peers’. Their outcomes will be larger. They will not be on a leaderboard.
That is the work the last round skipped. The next one does not have to.
