Parag K Adarkar

Insights AI that pays off

The AI bill is coming: FinOps for the age of agents

Parag K Adarkar · · 5 min read

FinOpsGenerative AIAgentic AI

In May 2026, Axios reported a story that spread quickly through technology circles: an AI consultant said one of their enterprise clients had run up a bill of around $500 million on Anthropic’s Claude in a single month, after rolling it out to staff without usage limits. The company was not named and the figure has not been independently confirmed, so it is best treated as a warning rather than a fact.

The warnings that are on the record are just as instructive.

What has actually happened

Uber. Uber’s CTO, Praveen Neppalli Naga, disclosed that the company had used up its entire 2026 budget for AI coding tools within the first four months of the year, as adoption of Claude Code among its engineers surged. Part of the push had been an internal leaderboard that ranked engineers by AI usage. Uber’s fix was not to switch the tools off. It increased prompt caching, changed which models were used by default, evaluated more efficient models and gave engineers visibility of their own usage and cost. By August, Naga was saying publicly that the industry was “coming to the end of the so-called tokenmaxxing era.”

Microsoft. According to reports by Windows Central and The Verge, Microsoft cancelled most Claude Code licences in one of its largest engineering divisions in mid-2026 and moved engineers to its own GitHub Copilot CLI. The official reason given was toolchain unification; commentators widely linked the decision to cost.

The common thread is not a particular vendor. It is that AI was adopted at the speed of enthusiasm, while cost controls moved at the speed of the annual budget cycle.

Why agents change the economics

A chatbot answers a question and stops. An AI agent plans, calls tools, reads results, revises its approach and tries again, and every step consumes tokens.

Research from Stanford’s Digital Economy Lab with Microsoft Research, published in May 2026, measured this for coding agents. It found that agentic tasks consumed around 1,000 times more tokens than code chat or code reasoning. Even more important for anyone who manages a budget:

  • The same agent on the same task varied in cost by up to 30 times from one run to the next.
  • Agents could not reliably predict their own costs in advance.

Unpredictable consumption, multiplied across thousands of users and many workflows, is exactly how a reasonable pilot budget turns into an unexplained overrun.

AI integration without financial governance, usage caps and monitoring is a recipe for disaster. Build your guardrails before you scale.

The guardrails I would put in place first

The FinOps Foundation, which sets practice standards for cloud financial management, now publishes specific guidance for AI. Drawing on that guidance and on the cases above, these are the controls I would expect any enterprise to have before it scales AI beyond a pilot:

1. Budgets and caps at every level. Set monthly limits per business unit, per application and per user, with automatic alerts at 50%, 80% and 100%. Hard caps for experiments; softer, approved limits for production workloads.

2. Sensible defaults. Route routine work, such as summaries, classification and simple drafting, to smaller, cheaper models by default. Make the most capable models available on request, for the tasks that genuinely need them.

3. Caching and context discipline. Reuse responses and shared context wherever possible. Long, repeated context is one of the largest hidden drivers of cost.

4. Visibility for the people spending. Show users and teams what their usage costs, in rupees, close to real time. Uber found that visibility changed behaviour faster than restrictions did.

5. Measure value, never volume. Retire any metric or incentive that rewards raw usage. Leaderboards of tokens consumed measure activity, not outcomes.

6. Unit economics tied to the business. Track cost per outcome rather than cost per token: cost per resolved complaint, per verified bill, per closed service ticket. The FinOps Foundation highlights measures such as cost per inference and anomaly detection rate; the business needs to see the next step, cost per result.

7. Anomaly detection. Agents can loop. Alert on unusual spikes by user, application and hour, and give someone the authority to pause a runaway process immediately.

Why this matters even more in utilities

In a regulated utility, budgets are planned, approved and scrutinised in advance. An unplanned overrun is not just a finance problem; it is a governance question that has to be explained to the board and, ultimately, to the regulator.

That makes the discipline above essential rather than optional. Treat AI spend like any other controllable operating cost: forecast it, allocate it to the business owner, review it monthly and link it to measurable benefits.

The bottom line

AI is not too expensive to use. Uber’s experience shows the opposite: usage can grow substantially while unit costs fall, if the right engineering and financial controls are in place.

What is expensive is AI without governance. The organisations that get this right will scale AI confidently; the ones that do not will either overspend or, worse, pull back from AI altogether at exactly the wrong moment.

Sources