A May 2026 reality check on enterprise AI.
Two stories landed this month that anyone selling, buying, or building on AI should sit with for a minute. Microsoft, arguably the company most committed to AI as a business model, started pulling AI tools away from its own engineers. And Uber, which had publicly bragged about its AI coding adoption, conceded it cannot draw a line between any of that activity and a better product for its customers.
If you know Gartner's Hype Cycle, you recognise the shape of these stories. They look like the slope into the Trough of Disillusionment.
Microsoft pulls back
On May 14, Tom Warren's Notepad newsletter at The Verge reported that Microsoft is cancelling most internal Claude Code licences across its Experiences and Devices division, the team responsible for Windows, Microsoft 365, Outlook, Teams, and Surface. The deadline is June 30, 2026, and the internal memo came from Executive Vice President Rajesh Jha. Engineers are being steered toward GitHub Copilot CLI, a cheaper tool Microsoft already owns.
The official reason is cost. The unofficial reason, reading between the lines, is that Claude Code had become, in Warren's words, "perhaps a little too popular" inside Microsoft, with engineers choosing Anthropic's tool over Microsoft's own product. The bills caught up.
The shift in language is at least as interesting as the policy change. In a new Windows 11 e-book, Microsoft now says it has observed a fundamental shift where adding more AI tools does not correlate with higher returns, and that enterprises using fewer AI tools but the right ones tend to see the biggest returns. For a company that put Copilot into every product it ships, that is a material change in posture.
Uber's budget vapour trail
Uber's situation has the same shape, just louder. According to reporting from The Information that the company has since confirmed, Uber burned through its entire 2026 AI coding tools budget on Claude Code and Cursor by April, four months into the fiscal year. Per-engineer monthly API costs ranged from $500 to $2,000. Some 95% of Uber engineers were using AI tools monthly, and roughly 70% of committed code originated from AI. CTO Praveen Neppalli Naga said the company was "back to the drawing board" on AI budgeting.
Adoption was not the problem. Adoption had been engineered. Uber ran internal leaderboards ranking teams by AI tool usage, effectively turning consumption into an office sport.
On May 22, Uber's president and chief operating officer Andrew Macdonald went on the Rapid Response podcast and said the awkward part on the record. Asked whether rising AI usage was translating into better consumer features, he answered: "That link is not there yet." He went on: "Maybe implicitly there's more that is getting shipped, but it's very hard to draw a line between one of those stats and 'Okay now we're actually producing like 25% more useful consumer features.'"
“That is the COO of a company that gamified internal AI adoption telling a podcast audience he cannot prove the spend is producing anything customers can feel.
The pattern under the anecdotes
Two cases would be a story. The data behind them is a trend.
The capital, meanwhile, keeps flowing. Gartner forecasts AI agent software spending will reach nearly $207 billion in 2026, up more than 139% from $86.4 billion in 2025. According to Crunchbase, AI startups captured roughly 80% of global venture funding in the first quarter of 2026.
Spending accelerating while proof of value stalls is, in textbook terms, what the descent toward the trough looks like.
The other half of the picture
It would be misleading to write "we are entering the trough" without acknowledging who is winning, because some companies clearly are.
PwC's 2026 AI Performance study, drawn from interviews with 1,217 senior executives across 25 sectors, found that 74% of AI's economic value is captured by just 20% of organisations.
The leaders are not the heaviest users. They are the ones treating AI as a catalyst for growth and business reinvention, particularly to pursue new revenue from converging industries. Meta, Anthropic, and Nvidia have demonstrated what AI-driven returns look like at their respective layers of the stack. Stema Partners reports that smaller and mid-market firms running focused AI projects in France saw a median ROI of 159%, with payback inside seven months. Deloitte's 2026 enterprise AI report found that 66% of organisations are now seeing real productivity and efficiency gains from AI, even if only about a fifth of them are seeing it bend the top line.
“A minority is generating compounding value. The majority is spending and waiting.
So, the trough?
The Gartner framework is useful here because the Trough of Disillusionment is not the same thing as Failed Technology. It is the standard correction between hype and durable adoption.
Analysts at Talyx had already placed generative AI firmly in the trough at the end of 2025, having cleared the Peak of Inflated Expectations in 2024. What May 2026 looks like is the trough deepening before the climb back out.
Three things tend to happen on that descent, and all three are now visible:
Public expectations get unwound
Microsoft's pivot from "AI everywhere" to "fewer tools, used better," and Uber's COO admitting on a podcast that he cannot prove the spend produced anything users can feel.
Cost reality catches up
Monthly per-engineer bills of $500 to $2,000. Vaporised annual budgets in a single quarter. The CFO's spreadsheet wins the argument.
Attention consolidates on what works
Companies stop running fifty pilots and start running five with named owners, defined KPIs, and a clean line to a business outcome.
Questions worth asking inside your own organisation
If you run an AI programme, the Microsoft and Uber stories should provoke some uncomfortable diligence.
Are you measuring activity or outcomes?
"70% of code is AI-generated" is an activity number. "Customer-facing feature throughput up 25%" is an outcome. Uber had the first and not the second.
Do you know your unit economics per use case?
Token costs no longer behave like fixed line items. They scale with how creative your engineers get. If you cannot state your cost per resolved ticket, generated report, or shipped feature, you cannot really say whether you are winning.
Are you confusing adoption with value?
Leaderboards measure who is trying. They do not measure who is delivering. Uber's experience suggests this is a live failure mode, not a hypothetical one.
Have you identified the narrow set of use cases where AI is genuinely transformative?
And pulled ruthless focus toward them. Microsoft's own internal pivot suggests even the largest AI company in the world is arriving at that conclusion.
Where this leaves us
Are we headed into the Trough of Disillusionment? On the available evidence, yes, and that is probably the healthiest thing that could happen to the AI market right now.
The trough is where the noise gets filtered. Companies stop asking "are we using AI?" and start asking "what is AI specifically doing for our P&L?" Some of the loudest current adopters will go quiet. Some of the quiet ones will surprise everyone on the climb up the Slope of Enlightenment.
“The firms that emerge strongest are likely the ones doing two unfashionable things at once: spending less than their peers on AI, and getting more out of it.
The 20% capturing 74% of the value did not get there by running internal usage leaderboards. They got there by treating AI as a means, not a metric.
If your only tool is an AI
Every wave of technology produces the same failure mode. Someone hands an executive team an exciting new capability and the team starts roaming the business looking for places to apply it. The order is wrong. The capability is the answer pretending to be a question, and the result is a portfolio of pilots that nobody can justify six months later.
AI is in that phase right now. The companies stuck in the trough are largely the ones that bought the tool first and went hunting for problems afterward. The companies doing well are doing the inverse. They started with a business problem they actually cared about, defined what success looked like in numbers their CFO would accept, picked the smallest meaningful slice of that problem they could attack, and only then asked whether AI was the right instrument for the job.
That sequence is not new and it is not glamorous. It is the same discipline that separates good software investment from bad software investment in every category that came before this one. AI did not earn an exemption from it just because the demos are impressive.
A few principles worth committing to as a working operating standard:
Start from the business problem, never from the tool
If you cannot describe the problem in language a sales rep or a support agent would use, you do not yet have a problem worth solving with AI. You have a budget looking for an excuse.
Define success before you spend
What metric moves, by how much, over what period, and who owns the number. If those four answers do not exist on a single page before the procurement conversation begins, the project has already started losing.
Find the thin slice
Pick the narrowest credible version of the use case, prove it works at small scale, and only then commit serious capital. The Uber case is the warning. Mass enablement before validated value is how an annual budget vapourises in a quarter.
Kill failures quickly and visibly
Pilots that are not working get extended for political reasons, and the cost of carrying them eats the budget that should be funding the wins. If a use case has not moved its target metric inside the agreed window, retire it and reallocate, in front of the team.
Champion the wins with the same energy used to launch the pilot
The 20% of companies capturing 74% of the value are not the ones using AI everywhere. They are the ones who concentrated investment on the handful of bets that worked, told everyone about it, and made those wins the template for the next round of work.
And on speed. Moving fast is not a value-add when you are pointed at the wrong target. Every executive deck for the past two years has put "velocity" near the top of its AI principles. Velocity in the wrong direction is just expensive disappointment, delivered sooner. The companies that will look smartest a year from now are the ones that spent an extra month at the front end of each initiative defining the problem properly, then went fast on something that actually moved a number.
AI is a powerful tool. Like every powerful tool before it, it punishes the operator who picks it up before deciding what it is for.