And the tools you are buying will not fix it
Uber's engineering organisation reportedly burned through its entire annual AI budget in four months. Five thousand engineers with coding agents, spend running ahead of every projection, and a finance team discovering the scale of the overrun after the fact rather than before it.
The response across the industry has been immediate and uniform: buy visibility. FinOps platforms, token dashboards, anomaly alerts on spend. The cost tooling market is having its moment, and the logic seems sound. If spend is out of control, see the spend.
Visibility was never the problem.
I have spent the past two years watching enterprise AI initiatives cross, or fail to cross, the gap between pilot and production. The cost blowouts making headlines follow the same pattern every time, and the pattern has little to do with the price of tokens. The economics were never modelled before deployment. No one defined what a successful outcome should cost. No kill criteria were set. No one was given the authority to retire an agent that fails to earn its place. The dashboard arrives later, to display in fine detail a problem the operating model created on day one.
The FinOps Foundation reports 73 percent of enterprises running over their AI spend projections. The reflex is to treat that as a pricing problem, and the pricing story does not hold. Blended token costs have fallen sharply year over year. The price of the input is dropping while the bills are exploding, which means the problem cannot be price.
What changed is consumption. Agentic workflows burn 5 to 30 times the tokens of the chatbot assumptions the original business cases were built on. An agent does not answer a question and stop. It plans, calls tools, retries, checks its own work, and hands off to other agents, and every step in that loop is metered. The business case modelled a conversation. Production runs a process.
A cost dashboard will show you that divergence beautifully. It will not tell you that the divergence was knowable before the architecture was locked, because the teams that knew it did the modelling first.
The Uber story, and Microsoft's reported cancellation of internal Claude Code licences, a platform standardisation move toward GitHub Copilot CLI with usage-based billing among the reported reasons, are coding-agent stories. They are the visible, cancellable version of the problem. A coding agent is a productivity tool with a licence attached. When the spend stops making sense, procurement cancels the licences and the burn stops that week.
Production agents are the harder version. An agent embedded in a customer workflow, a claims process, or a supply chain decision cannot be cancelled with a licence change. It has integrations, dependencies, and a business process reshaped around it. Switching it off means switching a piece of the operation off.
And that is exactly where the spend is heading. Gartner expects agentic AI spending to rise more than 140 percent in 2026, to nearly $202 billion. The same firm forecasts the cost per resolution for agentic customer service passing three dollars by 2030, a number moving in the wrong direction for a technology sold on falling unit costs. Forrester puts integration and evaluation at an additional 28 to 44 percent on top of model spend. And analysis from Teneo points to an inversion: simple tasks stay cheap as volume grows, while complex multi-step work gets more expensive per outcome, not less.
The coding-agent overruns are this year's symptom, painful and recoverable. The production-agent version of the same failure compounds for years, and procurement cannot fix it.
Buried in the reporting on the enterprises that did not blow up is a consistent habit: they modelled token volume and cost per workflow type before finalising the architecture. They knew which workflows could carry agentic economics and which could not, and they designed accordingly. The ones reconciling spend after the fact skipped that step, locked the architecture, and are now discovering what it costs to run.
That is a delivery failure, not a pricing failure. And it explains why the tooling response keeps disappointing. A dashboard cannot give you two things.
It cannot model a unit you never defined. Cost per outcome, the cost of a resolved ticket or a completed task rather than total token spend, is a unit that has to be designed into the deployment: instrumented at the workflow level, connected to the business result, compared against the human baseline it replaced. No tool can retrofit a unit of measurement the operating model never specified.
And it cannot retire an agent nobody has the authority to switch off. Retirement is an organisational decision. It requires someone empowered to act on the number the dashboard shows, against the instincts of the team that built the agent and the sunk cost the business has already absorbed.
Four moves, none of which require buying anything.
Model cost per outcome per workflow before architecture lock, including the integration, evaluation, and oversight costs the vendor deck leaves out. Set a kill criterion before deployment, the economic line below which the agent does not survive. Name the one person with the authority to enforce that criterion over the objection of the team that built the agent, a single named owner rather than a committee. Then measure each agent on its own economics rather than in portfolio aggregate, because the aggregate is exactly what hides the three agents silently destroying value behind the four that are not.
The better cost tools do help with that last move. Per-agent measurement is precisely what they enable, and enterprises should use them for it. What no tool can supply is the unit, the criterion, or the owner.
The enterprises making headlines did not get there because their tooling was inadequate. They got there because they treated an AI agent as a feature to ship rather than an asset to run. A feature has no cost discipline, no retirement plan, and no named owner.
The ones who stayed out of the headlines are quieter for a reason. They worked out the economics before they built, and they kept the authority to retire whatever failed to earn its place. They will buy FinOps tools too in time, and the tools will help. But they know what a tool is and what it is not. It is a better gauge on the dashboard. The discipline to read that gauge and act on it was never something an enterprise could buy.
That is the part no dashboard ships with.
Vijayan Seenisamy is the author of The Pilot Trap and The AI Delivery Manager Blueprint, and the creator of the AI Role Operating Framework (AI ROF™).
