How much does it cost to run an AI agent?

Expect four cost lines, and expect the model bill to be one of the smaller ones. The others are the platform that governs the agent, the upkeep of its connections, and the people who review, approve and correct its work. The total turns on volume, how much the agent reads per case, and how much work comes back to people.

What are the four cost lines?

The people line is the one that moves most as an agent matures. At depth 2 a person handles every draft. At depth 3, only held items, escalations and samples come back. That is why a cost estimate without a depth is meaningless. The argument is laid out fully in the cost of an agent is not the model bill.

What changes the model bill?

• Context per case. An agent that reads a full customer history for every query costs far more than one that reads only what the task needs.
• Steps per case. Each tool call and each round of reasoning is another model call.
• Retries. Failed connections and malformed outputs repeat work.
• Model size. Larger models cost more per call; smaller ones may need more steps or produce more escalations.
• Volume shape. Some pricing rewards steady use and penalises peaks.

How do you estimate cost for your own agent?

1. Analyst: measures per-case usage in shadow. Average and worst-case model usage per case type. Evidence: usage per filed proposal.
2. Analyst: counts work coming back to people from shadow and certification: drafts, holds, escalations. Evidence: rates per case type.
3. Finance: prices each line at expected volume and at peak. Evidence: a cost model with assumptions stated.
4. Process owner: models two depths, the current one and the next, since cost shifts with depth. Evidence: two scenarios.
5. Finance: tracks actuals monthly per agent, using a shared AI agent cost tracking template. Evidence: a monthly cost line per capability.

What does a cost estimate look like?

Say a mid-sized insurer runs an agent that reads incoming claims correspondence, attaches it to the right claim and drafts a short summary at depth 2. It handles 10,000 items a month.

In shadow, model usage averages a small amount per item, with long medical reports costing about ten times the average. Total model cost for the month comes to the equivalent of a few days of one handler's time. The platform licence is a fixed monthly fee. Connections to the claims and document systems need a few days of engineering a quarter. Reviewers spend about a minute per summary checking and filing, around 170 hours a month, which is the largest line by far.

The cost per completed item is the total divided by 10,000. When the owner later certifies filing at depth 3 for routine correspondence, review time halves and so does the largest line. Model cost barely moves. Tracking how much reviewers change the agent's drafts shows whether that review minute is real work or a quick glance.

What should a running-cost plan include?

• Cost per completed case, not per call.
• A depth label on every figure, since the same agent costs differently at each depth.
• Recertification every 90 days as a recurring line.
• Fallback staffing for the days the agent is down or reduced.
• A peak-month estimate alongside the average.

Banking and finance agents have their own cost profiles, set out in what AI agents cost to run in a bank and what an AI finance agent costs. The full deployment sequence is in deploying an AI agent.