Skip to content
Fast Company Executive Board

The Fast Company Executive Board is a private, fee-based network of influential leaders, experts, executives, and entrepreneurs who share their insights with our audience.

The hidden meter running on your AI

A framework for avoiding cost surprises

[Photo: Getty Images]

Profile photo of Sasi Kiran Malladi

Principal Technical Account Manager · Amazon

Most companies can tell you whether their artificial intelligence is online. Far fewer can tell you what it spent in the last hour, or why. They are content to find out at month’s end, when the invoice lands. It’s a mistake too many make.

Treating AI cost as something you reconcile after the fact is how good products become unprofitable ones. The gap between a system that looks healthy and a system that is quietly expensive is the part of the AI era that has caught many leaders off guard.

For decades, software costs were boring in the best way. You bought software, the price was roughly fixed, and you could plan a year ahead. AI broke that arrangement. Every interaction with a model consumes tokens, the small units of text it reads and writes, and you pay for each. With usage-based pricing, spend now moves with behavior rather than with a contract. The more your customers use the AI agent, and the more steps AI takes to answer them, the more it costs your organization. The meter is always running, and most organizations cannot see it move until it is too late to do anything about it.

CHEAPER BY THE UNIT, BIGGER BY THE BILL

The confusing part is that AI keeps getting cheaper to use while the bills keep climbing. Gartner forecasts that by 2030, running inference on a one-trillion-parameter model will cost providers more than 90% less than it did in 2025. Yet, enterprise AI spending is rising anyway, because consumption is growing faster than prices are dropping. When something becomes cheaper and more useful, people use far more of it.

Gartner also predicts that by 2028, AI coding costs will exceed the average developer’s salary, driven by rising token consumption and the shift to consumption-based pricing. Their analysts make the point plainly: Many organizations are underestimating the financial impact.

The hidden cost of AI is not a rare problem for the unlucky. In one survey of 218 technology leaders, 78% reported unexpected charges tied to AI or consumption-based pricing in the last year. As a result, 61% said they were forced to cut projects. Companies are not just overspending. They are canceling good ideas to pay for surprises they never saw coming.

WHY AI AGENTS MAKE THE METER SPIN FASTER

The newer wave of AI raises the stakes. A simple chatbot answers a question once. An AI agent works toward a goal. To get there it can call the model many times for a single request, retrieving documents, invoking tools, and retrying when it is uncertain. Gartner’s March 2026 forecast puts agentic tasks at roughly five to 30 times the token consumption of an equivalent chatbot exchange.

Several things compound that cost, and most are invisible to standard finance processes. Input tokens, not just the AI’s answers, often drive the bulk of the bill. Every document and instruction you feed the model counts. Long context windows and retrieval-augmented workflows multiply the charge. An agent caught in a retry loop can quietly run up dozens of paid calls trying to recover from a single failure. None of this shows up on a traditional operational dashboard.

OBSERVABILITY MUST INCLUDE COST

Engineering teams already have a discipline for this. They watch the metrics, logs, and traces that software gives off—what is broadly called observability—to understand not just whether a system is running but how it is behaving and why. AI discipline must expand to include consumption. Watching latency and uptime while ignoring token usage is like monitoring a vehicle’s speed but never its fuel gauge.

The practical version of this is straightforward. Treat usage as a first-class signal, captured in near real time and attributed to the workflow, feature, or customer that generated it. Tracing matters here, too. When an agent fans out into a chain of model calls, you want to see the whole path and find the step burning most of the budget rather than guessing. The goal is to turn a single opaque number at month’s end into a stream of signals you can act on while the spending is still happening.

MAKE THE NUMBERS SPEAK IN BUSINESS TERMS

Visibility alone is not enough. A raw token count means little to anyone outside the engineering team. The metric that earns a place on an executive dashboard is cost tied to an outcome: cost per customer, cost per resolved support ticket, and cost per completed transaction. Those numbers tell you whether the AI is paying for itself.

Reframing changes the conversation. Instead of asking whether AI is expensive in the abstract, you can ask whether a specific workflow earns its keep and catch the ones that don’t before they scale. Cost architecture becomes a design choice, not an afterthought: routing simple tasks to smaller models, caching repeated requests, and declining to use a model where ordinary code would do the job more cheaply.

Any leader rolling out an AI feature should apply the following:

  • Put usage on a dashboard the same week you launch, not after the first invoice. The bill is a lagging signal, and by the time finance sees it, the money is already gone.
  • Express cost as cost-per-business outcome, not raw tokens. A number your board understands is a number you can manage.
  • Set alerts on spending spikes the way you set them on downtime. A runaway agent should alert someone before it drains a budget.
  • Assign clear ownership of AI unit economics, the way someone already owns the cloud bill.

AI is a signal to be managed in the moment. Successful companies will see what it costs them in time to act, while the feature is still a success and not yet a surprise.

Sasi Kiran Malladi is a principal technical account manager at Amazon.