Model prices drop every few months. Enterprise AI bills keep climbing. A July report from McKinsey on the economics of agents explains the paradox: 60% of the cost of an agentic task goes into refining responses before delivery. And 93% of the enterprises surveyed for their FinOps study admit they exceeded their AI budget.
We see the same pattern in the deployments we come across. The pilot consumed a fraction of what production consumes. The agent's architecture explains the difference.
Where the money goes
An agent does not answer one question and stop. It reads documents, forms a plan, executes a step, validates the result, corrects and retries. Each step is a model call. Each call resends the full accumulated context. McKinsey estimates an agentic task can consume close to a thousand times more tokens than a conventional chat query.
The main cost drivers the report names:
- Refinement. Checking, correcting and improving the answer before delivery takes 60% of the cost. Fixing a response costs more than generating it.
- Long-lived context. The agent rereads the whole history on every step. The longer the conversation lives, the more expensive each call gets.
- Oversized models. Routing a classification or a data extraction through a frontier model costs many times what the task needs.
- Retries. Every failed tool call and every hallucination someone fixes by hand adds another round of tokens plus your team's time.
One more cost never shows up on the provider's invoice: human time. When the agent delivers output someone has to review and repair, the real cost per task includes those hours. The report's summary: tokens are the bill, value is something else.
Measure per task before you scale
The metric that matters is full cost per completed, correct task: a claim resolved or an order processed. That number includes agent spend, retries and the reviewer's time.
Four habits keep the bill under control:
Budget per workflow before you build. Estimate reasoning loops and context depth for each task type. Teams that modelled token volume before choosing an architecture avoided the surprises; the ones that skipped it now reconcile invoices.
Match the model to the task. Most of an enterprise agent's work (classifying, extracting, summarizing) does not need a frontier model. A tiered architecture with a cheap model for routine work cuts the bill without touching the outcome.
Track quality next to spend. Token counts say nothing on their own. An agent that spends more per task but delivers work nobody has to redo can be cheaper than a per-token bargain your team reworks.
Review the economics every quarter. Which workflows justify their spend, which get redesigned, which move to a cheaper model. It is the discipline you already apply to cloud, applied to a cost that grows with usage.
At Luxion we build agents connected to real processes and data, and we put numbers on the case before writing code: which tasks, what volume, what cost per completed task. If you want to know whether an agent pays off in your business, we start with a small prototype with costs measured from day one.