Contents
- The paradox: the cheapest token in history, the most expensive bill in history
- Where the money actually goes: it isn't the first answer
- The numbers, all from primary sources
- An illustrative example: the pilot that tripled its budget
- What to do before scaling an AI pilot
- FAQ
- Cost doesn't control itself — it's designed
The price of an AI token has kept falling for two years straight. And yet the average company’s monthly AI bill keeps climbing. That’s not a contradiction: almost nobody measures the real cost of an AI agent per completed task. They measure it against the model’s list price, which turns out to be the smallest part of the problem. 67% of enterprises overran their AI agent spend budget by more than 10% in the past 12 months, according to IDC’s Future Enterprise Resiliency and Spending (FERS) Survey, Wave 4, July 2026. McKinsey, in its own survey of 120 enterprises, found an even higher number: 93% went over their planned AI budget.
The paradox: the cheapest token in history, the most expensive bill in history
Comparing an AI agent to an ordinary chatbot on token price alone is the first mistake. According to Gartner’s analysis, completing a task with an agent can require 5 to 30 times more tokens than a comparable chatbot exchange. The reason isn’t that the agent “talks” more — a single reasoning step or tool call usually produces only a few hundred tokens of output. What drives the cost is that, at every single step of a multi-step task, the agent re-reads a context window that can be ten or a hundred times longer than a simple chat prompt. On a 20-step task, the company pays for that full context 20 times over. Gartner forecasts that even as the price per token keeps falling, inference cost per agentic workflow will rise more than fivefold through 2028.
Where the money actually goes: it isn’t the first answer
McKinsey, in its report “Cost versus value: Managing agentic AI system performance” (QuantumBlack, 2026), isolates the exact driver: roughly 60% of the total cost of an agentic task isn’t the first answer at all — it’s the cycles of checking, correcting, and re-verifying the agent runs before it produces something actually usable. An agent that reviews its own output, retries a query that failed, or escalates to a pricier model because it isn’t confident enough in its answer can multiply the task’s cost several times over without showing up anywhere in the pilot’s spreadsheet. 95% of AI agent pilots fail for reasons close to this one: they’re designed around the model’s price, not around what finishing the task actually costs.
The numbers, all from primary sources
- IDC, FERS Survey, Wave 4, July 2026: 67% of enterprises overran their AI agent spend budget by more than 10% in the past 12 months. Breakdown: 43.1% ran over by 10% to 25%, 18.5% by 26% to 50%, and 5.4% by more than 50%.
- McKinsey (QuantumBlack), own survey of 120 enterprises, 2026: 93% exceeded their planned AI budget. Roughly 60% of an agentic task’s cost goes to reviewing and correcting the output, not generating it the first time.
- Gartner, August 2026 forecasts: an agentic task needs 5 to 30 times more tokens than a standard chatbot exchange for comparable work; inference cost per agentic workflow will rise more than fivefold through 2028.
- Ramp AI Index, September 2026 data, 85,000+ U.S. companies: the top 1% of AI spenders pay a median of $9,567.53 per employee per month. The median company among those that actually pay for AI spends $15.70 per employee per month. That gap is so large partly because most companies don’t track what they really spend, by agent, by task, or by department.
An illustrative example: the pilot that tripled its budget
(Illustrative example to explain the mechanism, not a real case from one of our clients.) A company estimates the cost of deploying a customer-support agent by multiplying the model’s token price by the projected volume of conversations, and presents a fixed monthly budget to the board. Two months in, the real bill is three times higher. The token price hasn’t changed; what changed is that the agent, whenever it doesn’t understand a request on the first try, retries the search, asks a pricier model for a second opinion before answering, and in edge cases repeats the whole reasoning loop two or three times before producing an answer the customer accepts without escalating to a human. None of that showed up in the pilot’s spreadsheet, which only counted the first call to the model.
What to do before scaling an AI pilot
1. Set the budget per completed task, not per token or per session. The model’s list price doesn’t predict real spend; the number of retries does. 2. Measure how many refinement cycles each agent runs before it delivers a usable answer. If that number doesn’t exist yet, that’s the first sign nobody is watching the spend. 3. Cap the number of steps and tools per task before letting the agent escalate to a pricier model on its own. 4. Put one named person in charge of weekly spend review, not a quarterly one: at this burn rate, a single misconfiguration can exhaust a month’s budget in a day. 5. Measure the pilot against real traffic for 4-6 weeks before multiplying users or agents, with containment and permissions already tested, not bolted on at the last minute.
A properly sequenced AI implementation guide catches most of these surprises before they happen: cost gets designed up front, not discovered on the first invoice.
FAQ
Why is the AI bill going up if the token price is going down? Because the token price is only part of the cost. An agentic task consumes far more tokens than a normal query — 5 to 30 times more, per Gartner — due to the retries, checks, and escalations to pricier models the agent runs on its own before delivering a usable answer.
What is the “response refinement” McKinsey talks about? It’s the set of checking, correcting, and re-verifying cycles an agent runs after producing a first answer and before considering it good enough to deliver. McKinsey puts this at roughly 60% of an agentic task’s total cost.
How much does a typical company spend on AI per month? Per the Ramp AI Index (September 2026, 85,000+ U.S. companies), the median company that pays for AI spends $15.70 per employee per month, versus $9,567.53 per employee per month for the top 1% of spenders. Part of that enormous gap comes from most companies not tracking real spend by agent or by task.
Is this only a big-enterprise problem with many deployed agents? No — the mechanism (retries, answer review, escalation to pricier models) shows up from the very first single-agent pilot. A small company that doesn’t set a per-task budget from day one can burn through a year’s AI budget in weeks.
How do we avoid blowing the budget before scaling an AI pilot? By measuring cost per completed task (not per token), capping retries and escalation before deployment, and auditing the pilot against real traffic for several weeks before expanding it. That’s exactly what OptimaQuantum’s AI implementation audits do.
Cost doesn’t control itself — it’s designed
The cheapest token in history has convinced a lot of companies that agentic AI is cheap. The numbers from IDC, McKinsey, Gartner, and Ramp say otherwise: without a per-task budget, a retry cap, and someone reviewing spend every week, the pilot that looked cheap turns into the invoice nobody modeled. At OptimaQuantum we audit the real cost of an AI agent deployment before the first surprise bill arrives, not after. Contact us and we’ll review it before you scale.
Sources consulted: IDC, Future Enterprise Resiliency and Spending (FERS) Survey, Wave 4, July 2026 · McKinsey & Company / QuantumBlack, “Cost versus value: Managing agentic AI system performance” (2026) · Gartner, agentic workflow inference cost forecasts (August 2026) · Ramp AI Index, September 2026 data (ramp.com/data/ai-index).
Get our new articles by email
One email per article. Unsubscribe anytime.