If Your AI Subscription Isn't a Budget Line, It's a Blind Spot.
Managing an AI token budget on a project means giving agent spend its own line in the budget baseline, not lumping it into a flat subscription fee and hoping it stays flat. That line needs an estimate sized to the work, a contingency reserve sized to how volatile that work is, and controls that catch an overrun before it happens rather than explaining it after. Uber found out what happens without those three things in April 2026, when it burned through its entire year’s AI coding budget in four months and had to build the controls that should have been there from the start.
AI Budget Overruns Are Already the Norm
Uber’s CTO confirmed this year that heavy adoption of AI coding tools had blown through the company’s entire 2026 AI budget by April, with some engineers running up monthly costs of $2,000 each on Claude Code alone. Four months of runway on a twelve-month plan.
It’s also not unusual, just unusually visible. CloudZero and Benchmarkit’s FinOps in the AI Era survey of 475 organisations, published in February 2026, found only one in five companies landed within 10% of their AI spend forecast: 54% missed by 11-25%, another 20% missed by 25-50%. A WitnessAI survey covered by CFO Dive in July 2026 found nearly seven in ten US companies had at least some AI initiatives run over budget in the past year. That’s consistent enough to stop treating it as a one-off.
A Subscription Fee Isn’t a Budget for What Actually Varies
The instinct to treat AI spend as “handled” because there’s a monthly subscription is understandable. It’s also the assumption that produces a four-month runway on a twelve-month budget.
A subscription covers access. It says nothing about consumption, and consumption is where the volatility lives. A chatbot is one prompt, one response, a roughly predictable cost. An agent working through a task makes tool calls, re-reads context, retries failed steps, and iterates, and every one of those burns tokens that never show up as “output” you’d notice.
Analysts at Exponential View call this the “ghost token” problem: on their analysis, well under half of an agent’s token consumption is the actual work you asked for. The rest is overhead invisible from the output but very much on the invoice. Whatever the exact split turns out to be for your workflows, the direction is the same: cost forecasts based on task count or a flat fee are structurally unreliable.
Put AI Token Spend in the Budget Baseline, Not Off to the Side
This is fixable. PMs already know how to handle a cost category that’s genuinely volatile, the same discipline applied to FX-exposed procurement or weather-dependent construction costs: estimate it, attach a contingency reserve sized to how much it actually moves, and build it into the baseline instead of treating it as a surprise at month-end.
AI token spend deserves the same treatment. Estimate consumption per workstream or phase at planning. Size a reserve that reflects how volatile that workflow actually is, not a flat 10% because that’s what you did for everything else. And when actual burn drifts from baseline, treat it as a budget variance requiring the same escalation any other overrun would get.
The Controls That Catch an Overrun Before It Happens
A budget line without controls just tells you how badly you missed. The controls are what stop the miss from happening.
Uber’s own fix, once the problem was visible, was concrete: prompt caching to cut redundant context spend, smarter default model selection so routine tasks weren’t burning frontier-model tokens by default, and usage dashboards the team actually looked at. Usage went up. Cost per unit of usage went down.
Three of those translate directly to project-level controls. Spend thresholds per workflow or agent, with an alert that fires before the cap is hit, not in a monthly report after it’s gone. Model-tier discipline, matching task complexity to model cost instead of defaulting every agent to the most expensive option available. A change-control trigger when a workflow’s token profile shifts (a new agent, expanded scope, more autonomy), the same way you’d trigger change control for any other scope change that affects the budget.
None of this requires new infrastructure. It requires deciding the thresholds before the sprint runs long, not after.
The Budget Conversation Changes When You Have a Line and a Reserve
Avoiding that outcome doesn’t mean spending less on AI. It means pricing the volatility in from the start, and building controls that catch drift while there’s still time to act.
That changes the sponsor conversation. Instead of explaining after the fact where the money went, the PM with a budget line and a reserve can say: here’s what we planned for, here’s what we’ve used, here’s what happens if we’re trending over. That’s a stronger position, and it’s available to any project team willing to treat AI spend like the resource it already is.
Frequently Asked Questions
How do I budget for AI agent token costs on a project? Give AI token spend its own line in the budget baseline: estimate consumption per workstream at planning, attach a contingency reserve sized to how volatile that workflow is, and review actual burn against it like any other cost line.
What controls should I put around AI agent spend? Set spend thresholds with alerts that fire before the cap is hit, apply model-tier discipline so routine tasks aren’t defaulting to the most expensive model, and trigger change control when a workflow’s token profile shifts materially, such as a new agent or expanded scope.
Why do AI costs blow through budgets even with a flat subscription? A subscription covers access, not consumption. Agentic workflows burn tokens on tool calls, context re-reads, and retries that don’t show up in the visible output, so consumption doesn’t track subscription tier the way a fixed cost would.
Does your project have a budget line for AI token spend, or is it still riding along inside a subscription fee nobody’s stress-tested?
Yes - AI helped me to write this :)
Unsubscribe
