May 31, 2026
Everyone Thinks AI Will Replace Software Engineers.
95% of Uber engineers use AI tools monthly. 70% of committed code is AI-generated. Yet the COO still can’t point to a single meaningful…
The Latency Gambler
3 min read
Everyone Thinks AI Will Replace Software Engineers. Then Uber Asked a Simple Question: Where's the ROI?
95% of Uber engineers use AI tools monthly. 70% of committed code is AI-generated. Yet the COO still can't point to a single meaningful customer feature that came from it.
The narrative has been loud and consistent for two years: AI is coming for your coding job. Companies race to adopt AI tools, developers nervously watch their commit counts, and LinkedIn fills with hot takes about the death of software engineering as a career.
Then Uber's COO Andrew Macdonald said something no executive had been willing to say out loud.
"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."
And this wasn't a company that skimped on AI. Uber burned through its entire 2026 AI budget for Claude Code and Cursor in four months. Not the quarter. The year. Their CTO had already disclosed this before the COO went public with the uncomfortable follow-up: we have no idea if it worked.
The numbers tell a confusing story
- 95% of Uber engineers use AI tools monthly
- 70% of committed code is now AI-generated
- $951 million spent on R&D in Q1 2026 alone (up 17% year-over-year)
- 0 clear links between that spend and useful consumer-facing features
High adoption. High cost. Unclear outcome. That's the enterprise AI situation in 2026, and Uber is just the first company to say it out loud.
The measurement problem no one is solving
The core issue isn't that AI tools are bad at writing code. They're genuinely impressive at it. The problem is that writing code faster is not the same as shipping more value.
Here's what the typical enterprise AI scorecard looks like today:
What companies measure:
- Lines of code per day
- AI tool adoption rate
- PR merge frequency
- Token spend per engineer
What actually matters:
- Features shipped per sprint
- Bug rate after release
- User-facing improvements
- Revenue impact
The gap between those two lists is where billions of dollars in AI spend are quietly disappearing.
The hidden cost architecture
Here's something procurement teams miss: AI tools appear almost free to individual engineers. The company absorbs the real bill and those costs now compete directly with headcount budgets.
// What's visible when procurement signs off
seats: "$X/month per developer"
subscriptions: "Claude Code, Cursor, Copilot"
// What actually blows the budget
tokenOverage: "agentic tasks use 10–100x more tokens"
reviewTime: "AI code still needs human review"
debugging: "hallucinated logic is subtle and expensive to catch"
rework: "confidently wrong is worse than obviously wrong"// What's visible when procurement signs off
seats: "$X/month per developer"
subscriptions: "Claude Code, Cursor, Copilot"
// What actually blows the budget
tokenOverage: "agentic tasks use 10–100x more tokens"
reviewTime: "AI code still needs human review"
debugging: "hallucinated logic is subtle and expensive to catch"
rework: "confidently wrong is worse than obviously wrong"Gartner estimates inference costs will fall 90% by 2030 but that won't rescue enterprise budgets, because agentic workflows consume far more tokens per task. Cheaper per token. Many more tokens per job.
What a real ROI framework looks like
Instead of measuring AI activity, teams should measure AI outcomes. The chain looks like this:
AI spend (tokens + seats)
↓
Task output (features, bug fixes, refactors)
↓
User impact (engagement, NPS, support tickets)
↓
Business value (revenue, retention, churn)AI spend (tokens + seats)
↓
Task output (features, bug fixes, refactors)
↓
User impact (engagement, NPS, support tickets)
↓
Business value (revenue, retention, churn)Most companies have instrumented step one AI spend. Some track step two code output. Almost none have a traceable line connecting AI-assisted work to steps three or four.
That missing chain is exactly what Uber's COO was describing.
A practical starting point
You don't need a sophisticated system to start measuring. You need a habit:
// Tag AI-assisted PRs with metadata
prMetadata = {
aiAssisted: true,
toolsUsed: ['claude-code', 'cursor'],
estimatedTokenCost: '$4.20',
humanReviewMinutes: 12
}
// Tie it to feature flags and analytics
featureFlag('checkout-redesign')
.track({ prId: '#4421', aiAssisted: true })
.onConversion(revenue => {
roi = revenue / prMetadata.estimatedTokenCost
// Now you have a number
})// Tag AI-assisted PRs with metadata
prMetadata = {
aiAssisted: true,
toolsUsed: ['claude-code', 'cursor'],
estimatedTokenCost: '$4.20',
humanReviewMinutes: 12
}
// Tie it to feature flags and analytics
featureFlag('checkout-redesign')
.track({ prId: '#4421', aiAssisted: true })
.onConversion(revenue => {
roi = revenue / prMetadata.estimatedTokenCost
// Now you have a number
})It's not perfect. But it's a chain of evidence which is exactly what's missing at companies like Uber right now.
So is AI replacing engineers?
Not in the way the headlines suggest. What's actually happening is more interesting.
AI is shifting what engineers spend time on. Boilerplate is cheaper. First drafts of functions are faster. That's real. But the cognitive work understanding the problem, deciding what to build, judging what's worth keeping hasn't been automated. If anything, it's become more important as the volume of generated code rises and someone has to make the calls.
The security angle compounds this. The npm ecosystem, infrastructure providers, and developer toolchains have all faced major breaches recently. AI-generated code that nobody fully reads introduces a new attack surface. Trust has to be earned with tooling, not assumed.
The engineers who will thrive aren't the ones who resist AI tools they're the ones who learn to measure output, not just activity.
The real story
Uber's admission isn't a failure story. It's the industry finally asking the right question.
Usage metrics were always the easy part. You could track seats, adoption rates, and commit volume from day one. Outcome metrics are harder they require connecting engineering work to user behavior to business results, across teams and tools that were never designed to talk to each other.
That's the actual problem. Not the AI. Not the engineers. The measurement gap.
Macdonald's comments will likely push procurement teams across the industry to demand outcome-linked justification before renewing AI tool contracts. That's a healthy development. Vendors will need to help customers prove ROI, not just adoption.
The shift from measuring AI usage to measuring AI outcomes is coming. Uber just put a name on it.
Sources: Uber COO Andrew Macdonald's public remarks (May 2026), Fortune, AI Weekly, Gartner enterprise AI cost research 2026.