Quotes

Most teams budget for an AI agent by looking at the model’s price per million tokens and stopping there. Then the first invoice arrives and it is three or four times higher than expected. Token prices matter, but they are only one of several costs that grow with usage.

This guide breaks a typical agent’s monthly running cost into its parts and shows a simple way to estimate each one before you build.

The five costs that make up the bill

1. Model usage. Input and output tokens for every call the agent makes, including the calls you do not see: planning steps, tool selection, retries and self-checks.
2. Retrieval. If the agent searches your documents, you pay to create embeddings, to store them in a vector database and to query it.
3. Tools and integrations. CRM, helpdesk, payment or telephony APIs often charge per request or per seat.
4. Hosting and observability. The service that runs the agent, its queues, logs, traces and evaluation runs.
5. Human review. Time your team spends checking escalations and improving prompts. It is rarely on the invoice, but it is real.

A worked example

Take a customer support agent that handles 20,000 conversations a month. Each conversation averages six turns. Suppose each turn sends about 3,000 input tokens (instructions, conversation history and retrieved help-centre passages) and returns 300 output tokens.

• Turns per month: 20,000 × 6 = 120,000
• Input tokens: 120,000 × 3,000 = 360 million
• Output tokens: 120,000 × 300 = 36 million

Now multiply by your provider’s current rates. Output tokens are usually priced several times higher than input tokens, so the 36 million output tokens can cost more than their volume suggests. Add 20 to 40 per cent for hidden calls such as tool selection and retries, and you have a realistic model figure.

Rule of thumb: estimate from turns, not conversations. Every turn resends the instructions and history, which is why long conversations get expensive quickly.

Where the money usually leaks

• Bloated context. Sending the whole conversation and ten retrieved documents on every turn. Summarise older turns and retrieve fewer, better passages.
• One model for everything. A small, fast model can classify intent or extract fields; save the larger model for reasoning and replies.
• No caching. Many providers discount repeated prompt prefixes. Stable system instructions placed first can cut input costs noticeably.
• Unbounded loops. An agent allowed to call tools without a step limit can burn through tokens on a single confused request.

Estimate before you build

A one-page spreadsheet is enough. List the expected monthly volume, turns per task, average input and output tokens, the share of tasks that need retrieval or tools, and your escalation rate. Run three scenarios, cautious, expected and busy, and compare the cost per resolved task with what the same task costs today.

If the cost per resolved task is comfortably below your current cost, and the quality bar can be measured, the agent is worth prototyping. If it is close, start with a narrower scope where the numbers clearly work, then expand.

Key takeaways

• Model tokens are usually the largest line, but rarely the only one.
• Measure cost per resolved task, not cost per conversation.
• Route simple steps to smaller models, cap tool loops and cache stable prompts.
• Track the real numbers from the first week of a pilot and update your estimate.