Research
Uber Burned Its AI Budget in Four Months. Microsoft Cut All Claude Code Licenses. The Real Story Is More Complicated.
Uber and Microsoft's AI budget disasters are real. So are GitHub Copilot's 55% productivity gains and Air India's 40,000 daily AI-handled queries. The difference is not the technology. It is the governance.
Two enterprise AI stories broke within days of each other in late May 2026, and both spread fast.
First: Uber burned through its entire 2026 AI budget by April, just four months into the year, after Claude Code adoption across its engineering organization grew faster than anyone anticipated. Uber COO Andrew Macdonald told Fortune that drawing a line between AI spending and actual improvements for consumers was difficult: “that link is not there yet.”
Second: Microsoft quietly began canceling its internal Claude Code licenses, ending access for its Experiences and Devices division by June 30. The reason was not that engineers disliked the tool. It was that they used it so much the token-based billing consumed the full annual AI budget in months.
Both stories are real and the underlying numbers are significant. They are also being interpreted as confirmation of something they do not actually confirm, that AI investment is broadly wasteful, that the technology does not deliver, or that the spending is irrational.
The full picture is considerably more complicated.
What Actually Happened at Uber
Uber introduced Claude Code to its engineering organization in December 2025. By February 2026, about 32% of engineers were using it. By March, that figure had jumped to 84%. Nearly 95% of Uber engineers were touching AI tools monthly. Roughly 70% of committed code had some AI involvement.
Fortune’s reporting puts the average cost at $150 to $250 per engineer per month. Heavy users were hitting $2,000 per month. Uber CTO Praveen Neppalli Naga told The Information he had spent $1,200 in a single two-hour demo session.
The 84% adoption rate is striking. Most enterprise software rollouts struggle to reach 50% adoption after a year. Uber’s engineers adopted an AI coding tool at near-universal rates in under three months. That is not evidence of waste. It is evidence of product-market fit so strong it overwhelmed a budget built on traditional per-seat software assumptions.
The COO’s comment about not being able to connect AI spending to consumer improvements is legitimate, but it is also a management accounting problem, not a technology problem. Attributing engineering output to specific consumer metrics is hard for any tool investment.
What Actually Happened at Microsoft
Microsoft deployed Claude Code to approximately 5,000 engineers. Per-engineer API costs reached between $500 and $2,000 per month at scale. The company is ending the program by June 30 after the pilot consumed its allocated budget.
The NextWeb’s reporting on the cancellation makes an important point about the token billing model: a seat license for a conventional productivity tool costs the same whether you use it for one hour or eight. A token-based AI tool costs more the more useful it becomes. That creates a fundamental paradox for enterprise buyers: high adoption, which is what you want, produces higher bills than the budget assumed.
This is not a story about AI failing. It is a story about procurement models that were designed for flat-rate software running into usage-based billing and getting surprised by the math.
The Token Billing Paradox
The structural issue sitting underneath both stories is one that 98% of FinOps practitioners are now grappling with: AI tools are billed differently from every other enterprise software category.
Traditional SaaS: cost is fixed. A Microsoft 365 seat costs the same for a light user and a heavy user.
AI tools: cost scales with usage. One employee using AI for email summarization might consume 10,000 tokens per day. Another using it for code generation might consume 10 million tokens per day. That is a 1,000x cost difference between two people with the same nominal tool access.
The result: 73% of enterprises reported their AI costs exceeded original projections in 2026. AI costs surged 108% year-over-year, with 78% of IT leaders seeing unexpected charges they had never budgeted for.
This is a governance gap, not a technology gap. The billing model changed, the controls did not.
Goldman Sachs Found No Economy-Wide Productivity Impact
The most credible skeptical data point on AI returns comes from Goldman Sachs research published in early 2026. Goldman found no meaningful relationship between AI adoption and productivity at the economy-wide level, and only 1% of S&P 500 management teams had quantified AI’s effect on earnings.
That is a striking finding and worth taking seriously. Goldman also found that FOMO, the fear of falling behind competitors, is a stronger incentive for enterprise AI investment than demonstrated performance outcomes.
Two caveats matter here. First, Goldman identified a genuine 30% productivity boost in two specific use cases: coding and customer service. Second, economy-wide productivity data lags significantly. Transformative technologies typically take years to show up in aggregate economic statistics. The personal computer showed productivity effects starting in the mid-1990s, roughly a decade after widespread adoption. Expecting AI to move GDP numbers in 2026 may simply be too early a measurement window.
Where the Real Returns Are
The skeptical stories are real. So is the other half of the picture.
GitHub Copilot is the most rigorously studied AI productivity tool. Controlled experiments by GitHub’s own research team found developers completed programming tasks 55% faster with Copilot than without it. Pull request cycle time dropped 75% in Accenture’s large-scale deployment. GitHub Copilot generates 46% of code across its user base and has been adopted by 90% of Fortune 100 companies.
Air India deployed an AI agent that now handles 40,000 customer queries per day, according to PwC’s 2026 AI Performance Study. The carrier reports savings of millions of dollars in customer service costs.
Palo Alto Networks used AI to automate 90% of IT operations, improving automated IT coverage from 12% to 75% of operations and halving IT operations costs.
Tru Cooperative Bank deployed Microsoft 365 Copilot and reached 93% employee adoption with 90% weekly active usage, rates most enterprise software deployments never achieve.
These are not cherry-picked outliers. PwC’s 2026 study found that 66% of organizations are reporting genuine productivity and efficiency gains from enterprise AI. Deloitte’s State of AI in the Enterprise report found that 45% of companies using generative AI report it has at least doubled employee productivity.
The gap is not between AI working and not working. The gap is between companies that have governance frameworks in place and companies that rolled out tools without them.
The 80/20 Split
PwC’s research identified a pattern that matters more than either the success or failure stories in isolation: 74% of AI’s economic value is being captured by just 20% of organizations.
The top 20% are not necessarily using better AI tools. They are using AI tools differently. Specifically:
- They set token usage limits and monitoring before deployment, not after the first invoice
- They measure AI impact against specific business metrics, not general productivity sentiment
- They distinguish between use cases where AI delivers clear returns (coding, customer support, content production, data processing) and use cases where benefits are diffuse or hard to measure
- They treat AI tools like any other variable-cost infrastructure, with consumption dashboards, budget alerts, and department-level attribution
The companies in the 80%, including Uber and Microsoft to some degree, rolled out powerful tools to large groups of people with budget assumptions built on seat-license logic. When usage exploded, the bills did too.
The 42% Abandonment Rate Is Also Real
On the other side: 42% of companies abandoned most AI projects in 2025, up from 17% the year before. Most of those abandonments were not because AI failed technically. They were because projects never made it out of proof-of-concept into production.
The industry term is “pilot purgatory”: technically successful experiments that generate no business value because they never scale. One consulting firm estimated pilot purgatory costs the average enterprise $15 to $25 million annually in wasted development resources and opportunity costs.
The pattern is consistent across both the over-spending stories and the pilot purgatory stories: AI governance is the missing layer, not AI capability.
What This Means for Teams Making AI Tool Decisions
The Uber and Microsoft stories should not make you dismiss AI tools. They should make you budget for them differently.
Before deployment:
- Get a token consumption estimate for light, average, and heavy users, not a single average
- Set department-level spending caps with alerts before you hit them
- Identify the two or three specific use cases where AI impact is measurable, and measure them before expanding access
During rollout:
- High adoption rates are good, but monitor the cost distribution. Uber’s average was $150 to $250 per engineer per month, but the heavy users were hitting $2,000. The budget should account for the distribution, not the mean.
- Distinguish between AI that replaces time (measurable efficiency) and AI that people use because it is interesting or easy (which is valuable but harder to attribute to business outcomes)
On governance:
- Token-based billing requires FinOps treatment, not IT procurement treatment. Put someone in charge of AI spend who understands consumption-based infrastructure costs
- Set up a cost-per-outcome metric for at least one AI use case in the first 90 days. If you cannot measure the output, you cannot defend the budget
The Actual Takeaway
The AI budget crisis at Uber and Microsoft is real. So is the Goldman Sachs finding that most companies cannot quantify AI’s effect on earnings. Those facts should inform how enterprises deploy and govern AI tools.
They do not change the verified productivity data from GitHub, the customer service results at Air India, or the 66% of organizations reporting genuine efficiency gains. AI is delivering returns, but the returns are concentrated in organizations that treated AI deployment like a capability build rather than a software purchase.
The headline version of these stories, the one circulating on LinkedIn, presents the Uber and Microsoft situations as proof that AI spending is irrational. The underlying data suggests something more specific: flat-rate budget assumptions applied to usage-based billing at enterprise scale will always produce surprises. That is a procurement and governance problem with a known solution.
The bubble, if there is one, is not in the technology. It is in the expectation that adopting AI tools is sufficient without building the operational infrastructure to measure and manage what they cost and what they return.