The agent bill nobody budgeted for
Token prices keep falling and AI bills keep rising. Agents do far more work per task than a chatbot, and most budgets were built for the chatbot. Here is how to price an agent before it ships.
The per-token price of a language model has fallen steadily for three years. Many teams took that to mean AI would get cheaper to run. For agents, it has not. An agent plans, retrieves, calls tools, checks its own work and tries again, and every one of those steps is billed.
Gartner put numbers to this in August 2026. It predicts that inference costs per agentic workflow will increase more than fivefold through 2028, and that routing a task to an agentic reasoning model costs at least five times as much as a basic chatbot interaction. Lower unit prices are being outrun by higher volume.
Most organisations are finding this out after the fact. McKinsey's Enterprise AI FinOps survey, as reported in July 2026, found that 93 percent of respondents had already exceeded their AI budgets, and that about 60 percent of agentic AI spend goes on the checking and correcting an agent does before it produces a usable answer.
Price the action, not the token
A token price tells a finance team nothing it can use. The useful number is cost per completed action: one invoice matched, one ticket resolved, one report drafted. Measure it in the pilot on real work, including retries and failures, and compare it with what that action costs today in staff time.
Include the human cost. Every exception the agent hands back and every output someone has to check is staff time, and it belongs in the same calculation.
If you cannot state the cost of one completed action, you are not ready to scale.
Four levers that move the bill
Use the smallest model that passes. Most steps in a workflow are classification, extraction or formatting, and they do not need the most capable model. Route those to a cheaper model, keep the expensive one for the step that needs judgement, and confirm the swap against your evaluation set.
Take steps out of the model. Anything with a fixed rule (a date check, a lookup, a total) should be ordinary code. It is cheaper, faster and right every time.
Cap the loops. Set a ceiling on retries and steps per task. An agent that has failed five times is unlikely to succeed on the sixth; it should hand the case to a person.
Cache and reuse. Policy documents, product data and instructions that do not change between requests should not be paid for on every request.
Put a budget in the design
Treat cost like latency: a requirement with a threshold, agreed before release and monitored after. Set a cost-per-action ceiling, alert when a week's average moves above it, and review the most expensive one percent of tasks each month. Those outliers usually point to a broken step rather than a hard problem.
- Gartner, Gartner Predicts AI Inference Costs Per Agentic Workflow Will Increase More Than Fivefold Through 2028, press release, 17 August 2026. gartner.com
- MarketScale, 60% of agentic AI costs go to response refinement, and most enterprises are already over budget, reporting McKinsey research, 25 July 2026. marketscale.com
- —Falling token prices do not mean falling agent costs. Agents use many more steps per task.
- —Measure cost per completed action in the pilot, including retries and human review.
- —Route routine steps to smaller models and fixed rules to plain code.
- —Set a cost ceiling per action before release and review the most expensive tasks monthly.