Uber encouraged its employees to use AI as much as possible. Uber then burned through its entire 2026 AI coding budget in just four months. These two facts sit side by side, and if you build bots for a living like I do, the tension between them tells you everything about where enterprise AI adoption actually stands right now.
The company has now instituted a cap of $1,500 per month per employee per agentic coding tool. According to reporting from Bloomberg, this came after the previous “use it all” approach produced spending that nobody in finance saw coming. Uber’s CTO reportedly said he’s “back to the drawing board.” And honestly? As someone who watches token costs pile up every single day while building and deploying bots, I have thoughts.
Token Costs Are the New Cloud Bill Nobody Plans For
If you’ve ever deployed an agentic coding assistant across a team — even a small one — you know how fast API calls add up. I run a small operation compared to Uber, and I still wince when I check my monthly usage dashboards. Now multiply that by thousands of engineers, all told to go wild with these tools, and you get Uber’s situation.
What’s fascinating is the specific tool that reportedly triggered the budget implosion costs around $200 per month per seat. That sounds reasonable in isolation. But agentic tools don’t just sit there waiting for a prompt. They run loops, call APIs repeatedly, spin up context windows, and burn through tokens at rates that scale with usage intensity. When you tell an engineering org of Uber’s size to adopt these tools aggressively, you’re essentially writing a blank check to the model providers.
Why This Matters for Anyone Building with AI Agents
Here’s what I think bot builders and architects should take from this story:
- Usage-based pricing is unpredictable by design. We build bots with token budgets baked into their architecture. Uber apparently didn’t set per-employee guardrails until after the damage was done. If you’re deploying agents in production, you need hard spending limits in your code, not just in your procurement policies.
- Encouragement without constraints is a recipe for cost explosions. I’ve seen this in my own projects. You give an agent permission to iterate, and it will iterate. Agentic coding tools are especially hungry because they reason in multi-step loops. Each “thought” costs tokens.
- $1,500 per month per tool is still a lot of money. That cap is not stingy. It signals that Uber still believes in AI-assisted development — they’re just trying to make the spending predictable. For those of us building bots on tighter margins, this number is a useful benchmark for what a large enterprise considers acceptable per-developer AI spend.
What I’d Do Differently
If I were architecting an AI coding tool rollout for an org of any size, here’s the approach I’d take based on what I’ve learned running my own agentic systems:
First, set token budgets at the agent level, not just the account level. Every bot I deploy has a per-session and per-day ceiling. When it hits the limit, it stops gracefully and reports what it couldn’t finish. This gives you visibility into where the real value is being generated versus where tokens are being wasted on circular reasoning.
Second, track cost-per-outcome, not just total spend. How much did it cost the agent to ship a working PR versus how many tokens were burned on failed attempts? This data is gold for tuning your system prompts and deciding which tasks are actually worth delegating to an agent.
Third, stage your rollout. Don’t flip the switch for everyone simultaneously. Start with a cohort, measure their usage patterns for a month, build your forecasting model, then expand. Uber apparently skipped this step.
A Healthy Growing Pain
I don’t read this story as a failure of AI tools. I read it as the predictable collision between enthusiasm and operational reality. Every new technology wave has a moment where the bill arrives and finance departments start asking hard questions. Cloud computing went through this exact cycle a decade ago. Now we have FinOps teams and cost-optimization tooling everywhere.
AI agent spending will follow the same trajectory. We’ll see better monitoring dashboards, smarter token allocation systems, and probably a whole category of startups building “AI FinOps” products. For those of us building bots today, the lesson is simple: plan for costs as carefully as you plan for capabilities. The smartest agent in the world isn’t worth much if it bankrupts your budget before lunch.
🕒 Published: