Microsoft has started canceling most of its internal Claude Code licenses, with engineers in the Experiences + Devices division — Windows, Microsoft 365, Teams, Outlook, and Surface — directed to switch to GitHub Copilot CLI by June 30. The same week, Uber’s COO disclosed that the company burned through its entire 2026 AI coding budget in four months and cannot draw a clear line from token consumption to consumer features. Neither company failed at AI. Both ran into the same problem: agentic tools don’t run on chatbot economics.

Key takeaways:

• Microsoft is canceling most Claude Code licenses by June 30; Experiences + Devices engineers move to GitHub Copilot CLI.
• Uber burned its entire 2026 AI coding budget in four months; COO says the ROI link is “not there yet.”
• Goldman Sachs forecasts 24x token demand by 2030; Gartner warns a 90% per-token cost drop still won’t reduce enterprise AI bills.
• Most operator AI budgets were sized for chatbot math; agentic tools operate on fundamentally different economics.

Why Do Agentic Tools Cost So Much More?

A chatbot exchange uses hundreds to low-thousands of tokens. An agentic coding tool works in multi-step loops — plan, write, execute, read output, retry, verify — turning a single developer task into tens of thousands of tokens. Multiply that across hundreds of engineers daily and the math compounds fast. Goldman Sachs Research forecasts 24x token demand by 2030 driven by enterprise agentic adoption. Gartner projects inference costs will fall 90% by 2030 — but warns the savings won’t hold: agentic workflows consume far more tokens per task, and providers are unlikely to pass all reductions through. The unit price drops. The volume explodes. The total bill rises.

What Should Operators Do Now?

Microsoft’s cancellation is partly a first-party decision — Microsoft owns GitHub Copilot and has a business reason to consolidate. Uber’s COO Andrew Macdonald attributed the overrun to an internal leaderboard that rewarded total AI tool usage, not outcomes. But both cases expose the same structural gap: enterprise AI budgets were sized for chatbot-era consumption, and agentic tools cross that ceiling in months. The Salesforce/Anthropic token-spend picture — $300M in coding tokens projected for 2026 alone — confirms the pattern. Three moves this quarter:

1. Decouple your AI cost model from seat-license math. Token-based tools don’t behave like SaaS subscriptions; per-token costs scale with usage intensity, not headcount.
2. Set per-team spending caps before deployment, not after the bill arrives. Uber gamified usage and paid for it. Build governance in from day one.
3. Negotiate consumption controls now. Enterprise agreements should include hard spend caps or alerts, model routing options, and volume tiers sized to actual agent-loop consumption — not flat per-seat pricing.

Ask sharper vendor questions before your next AI coding tool renewal: What are monthly consumption limits? Can we set spending governors by team? Are there model routing options for routine agentic tasks? What happens when we exceed the expected usage envelope?

Watch next: Whether Microsoft and Uber signal a Q3 2026 enterprise pullback in AI coding licenses — or whether this reflects early over-deployment with misaligned incentives that more disciplined operators can avoid.

---

Frequently asked questions

Why is the per-token cost falling while my AI bill is rising?

Per-token pricing has fallen sharply as models become more efficient. But agentic tools — AI that plans, executes, and self-corrects across multi-step tasks — consume orders of magnitude more tokens per interaction than a simple chatbot prompt. The math inverts: cheaper per token, but vastly more tokens consumed per developer per day. Gartner estimates even a 90% cost reduction won’t offset volume growth in enterprise agentic deployments.

Should we stop deploying AI coding tools after the Microsoft and Uber cases?

Not necessarily. Both situations involved specific failure modes — first-party vendor consolidation at Microsoft and a misaligned usage-reward leaderboard at Uber — not inherent tool failures. The right response is to build spending visibility and governance into deployment from day one: track tokens per team, set budget caps, and tie AI tool use to measurable output metrics rather than usage volume.

---

By Advanced AI · May 31, 2026 · Briefings