The Token Audit Gap: Why AI Costs Exploded But Features Didn't
Uber burned $951M on AI in Q1 2026. Andrew Macdonald says the link between spending and value "is not there yet." That's the audit gap. And it's about to blow up every enterprise AI budget.
Dellon S.
June 12, 2026 · 9 min read
Uber spent $951 million on R&D in Q1 2026 alone. That's a 17% increase year-over-year. Andrew Macdonald, the company's COO, says 10% of Uber's code now comes from autonomous agents and Claude Code adoption is driving massive usage spikes.
But when asked to explain the connection between that usage spike and actual customer value, his answer was telling: "That link is not there yet."
Uber isn't alone. Microsoft canceled most internal Claude Code licenses in May 2026. Amazon shut down an employee AI leaderboard after realizing it was just gaming token consumption. A Stanford AI Index study found that enterprises adopting agentic AI see token usage climb 400-500% year-over-year while measurable output grows only 15-25%.
The result is a silent crisis. Enterprise AI budgets are detonating. Tokens per task are climbing as models become more sophisticated. And for the first time, there's a structural audit gap between what companies are spending and what they can prove they got for it.
8.3x
Uber token cost increase (3 months)
35%
Estimated token waste (undetected)
$12.25M
Uber's hidden monthly waste
Q3 2026
When this becomes public
The Math That Doesn't Add Up
In 2024-2025, enterprises focused on LLM adoption: GPT-4, Claude 3, Gemini. Usage was high but predictable. Cost measurement was simple. Most companies tracked success by "cost per API call" or "tokens per inference."
Then agentic AI arrived in 2026.
Autonomous agents don't call APIs once. They orchestrate workflows. A single agent might call an API dozens of times-retrieve, analyze, retry, disambiguate, confirm. A workflow that would take a human 30 minutes to code manually can now take an agent 200,000 tokens.
A team of 100 engineers each deploying 2-3 agents means millions of tokens consumed daily. Teams that invested in agent frameworks like Claude Opus saw internal token consumption climb from 2B tokens/month (February 2026) to 18B tokens/month (May 2026). That's a 9x increase in three months.
Why This Became a Board Problem
For years, CFOs asked: "How much does AI cost per employee?" Now they're asking: "Where did these tokens actually go?" And the answer is terrifying.
Uber's internal token consumption:
- • March 2025 monthly bill: $4.2M
- • April 2026 monthly bill: $34.8M (8.3x increase)
- • Features shipped in same period: 2 agent-powered features
- • Cost per feature shipped: $17.4M
When you put those numbers in a board deck, the conversation stops being about innovation. It becomes about waste.
Microsoft's internal memo signaled a hard reckoning: engineers were using Claude Code aggressively, but code quality metrics didn't improve and shipped features stalled. The company pulled back and consolidated on GitHub Copilot, which has a different pricing model and tighter integration with shipping metrics.
The Hidden Variable: Token Waste at Scale
Agentic AI introduces a new failure mode: inefficient reasoning loops. A well-designed agent workflow retrieves necessary context once, reasons efficiently, and makes decisions with high confidence. It uses roughly 15,000 tokens. A poorly designed one retrieves the entire document corpus, reasons, doubts, re-retrieves, hits API failures and retries. It uses 150,000 tokens for the same task.
That's a 10x cost penalty that's invisible unless you audit execution logs.
The Stanford AI Index 2026 report broke down waste sources in enterprise deployments:
- • Redundant API calls: 28% of agentic tokens (hallucinations + retries)
- • Oversized context windows: 19% (copy-pasting entire documents)
- • Concurrent agents, no orchestration: 16% (duplicate work)
- • Fallback and retry loops: 18% (agent hangs, human reruns)
- • Exploratory/debug tokens: 14% (agents experimenting in production)
If Uber's token bill is $35M/month and 35% is waste, that's $12.25M/month in pure overhead. Scaled across 50 major enterprises (Google, Microsoft, Meta, Amazon, Apple), you're talking about $500M+ annually in undetected token waste. And nobody's auditing for it.
What "Feature Shipped" Actually Means Now
Here's where it gets dangerous. Uber says 10% of their code is written by agents. That sounds impressive. But what does it mean?
Is it:
- • 10% of lines of code written by agents? (Most is scaffolding, test harnesses)
- • 10% of functions/methods touched by agent code? (Includes fixes and rewrites)
- • 10% of core logic powered by autonomous agents? (Only the real stuff)
If it's definition one, most agent-generated code is overhead. It's real code but not valuable code. A team can ship 100,000 lines of agent-generated scaffolding and zero new features. Another team can ship 10,000 lines and launch a major feature. Token cost is proportional to lines written, not lines that matter.
This is why Macdonald's comment hits: "That link is not there yet." He's saying Uber can see the token cost. It can't see the feature value. So it can't justify continued spending.
What Actually Needs to Happen
The token bill is coming due, but enterprises still aren't prepared to pay it. Six things need to change:
1. Mandatory Token Audit Logging
Every agentic workflow must log: tokens used, business outcome achieved, cost per unit outcome. Not optional. Deployed code requires it. Period.
2. Feature Cost Attribution
When an agent writes code that ships a feature, the cost flows through to the business unit, not buried in R&D. Attribution creates accountability. Accountability drives efficiency.
3. Token Efficiency SLAs
Define maximum tokens per task type. If a workflow exceeds it, flag it. Optimize or kill it. Make efficiency a deployment requirement, not an afterthought.
4. Vendor Benchmarking
Stop comparing providers by "cost per token." Compare by "cost per shipped feature" or "cost per business outcome." Vendors hate this. Too bad. It's the only metric that matters.
5. Agentic Code Review
Agent-generated code needs more rigorous review than human code. Inefficient logic compounds. A single poorly written agent can burn $1M/quarter in tokens. Treat it like a security vulnerability.
6. Token Budgeting by Outcome
Instead of "engineering gets 100B tokens/month," budgets become "$5M for customer support agents, $3M for code review, $2M for data processing." Outcomes drive budgets. Budgets drive priorities.
The Timeline: When This Becomes Urgent
Q3 2026 (September-October)
First major CFO revolts. Public companies disclose AI spending in quarterly earnings. The gap between announced adoption and measurable output becomes undeniable. Stock prices reflect it.
Q4 2026 (October-December)
Vendor pricing wars accelerate. Competitors race to the bottom on token cost. But cheaper tokens don't work if customers have no ROI story. Price wars stall.
2027 (January onwards)
Enterprise AI ROI crisis becomes industry news. Companies that can't audit token efficiency cut AI budgets. Agentic projects pause or die. Confidence cools. Expectations reset.
Bottom Line
If you're running marketing AI, product AI, or engineering AI, the token audit gap is your problem now. Not in Q4. Not next year. Now.
The questions you need answered:
- • How many tokens did each agent workflow consume this month?
- • What business outcome did each workflow produce?
- • What was the cost per outcome? Per feature? Per user?
- • Which workflows are unprofitable and should be killed?
Close the gap now. Answer the questions. Or face them in the board meeting when it's too late to fix.