How much does AI agent development cost? What you actually pay for

"How much will an AI agent cost?" is usually the second question business owners ask us, right after "what can it do?". The honest answer is that an agent has two price tags: the one-off cost of building it, and the ongoing cost of running it. The first depends mostly on how many of your systems it has to work with. The second depends on how much it is used, which AI model it runs on and how carefully it is engineered.

This guide breaks both down in plain language, using the AI providers' own published pricing, so you can budget sensibly and ask a developer the right questions. It is general information; your own costs will depend on your systems, volumes and requirements.

What counts as an "AI agent"?

An AI chatbot answers questions. An AI agent can also take actions, such as looking up an order, booking an appointment or updating a ticket, by calling tools that your developers connect to it. Anthropic describes tool use as the model deciding when to call a function you define and returning a structured request that your application then runs. If you are not sure which one you need, start with our guide to AI chatbots vs AI agents.

That difference is the main reason agents cost more: every tool is a connection to build, secure and test.

What are you paying for when you build an agent?

Most AI agent projects follow the same lifecycle. Each stage takes time, and time is what you pay a developer for.

1. Discovery and design. Choosing one or two tasks worth automating, mapping how they are done today, and deciding what the agent may do on its own and what needs a person's approval.
2. Knowledge and data preparation. Gathering the policies, FAQs, product data and past tickets the agent will rely on, and cleaning them up. Messy or out-of-date content is one of the most common hidden costs.
3. Integrations (tools). Connecting the agent to your store, CRM, calendar, help desk or database through their APIs. Each tool needs a clear description, narrow permissions, error handling and logging.
4. Guardrails and approvals. Rules for what the agent must never do, approval steps for anything that moves money or changes an account, and a clean hand-over to a person.
5. Testing and evaluation. Running the agent through real past conversations and edge cases before launch. Anthropic warns that the autonomy of agents means higher costs and the potential for compounding errors, and recommends extensive testing in sandboxed environments.
6. Launch, monitoring and improvement. Dashboards, review of conversations, fixing gaps in knowledge and adjusting prompts and tools as your business changes.

What makes a build cheaper or more expensive?

• Scope: one task on one channel (for example website chat) costs less than several tasks across chat, email, phone and WhatsApp.
• Tools: read-only tools (looking up an order) cost less than tools that change things (refunds, account changes).
• Systems: modern systems with documented APIs cost less than older systems with no API, or data spread across spreadsheets.
• Knowledge: content that is already written and current costs less than knowledge that lives in people's heads.
• Autonomy: having a person approve every action at first costs less than fully automatic actions that need extra safeguards.
• Data: standard data protection needs cost less than health, financial or children's data, or strict data residency.

A sensible first project automates one well-defined task, keeps a person in the loop for actions, and is extended once the results are proven.

What does it cost to run an agent each month?

Running costs come from three places:

• Model usage. AI providers charge per token (roughly, pieces of words) for what goes into the model and what comes out. An agent sends more than the customer's message: it also sends your instructions, the tool descriptions, the conversation so far and any documents it has looked up. Anthropic's pricing documentation confirms that tool definitions count as input tokens, and that tool use adds a small system prompt of its own.
• Extra tools. Some built-in tools cost extra. For example, Anthropic charges $10 per 1,000 web searches on its API, on top of token costs.
• Hosting and other services. Servers or serverless functions, a database or search index for your documents, logging and monitoring, plus any software your agent connects to.

A worked example

Here is an illustrative calculation using Anthropic's published API prices (checked 27 October 2026; prices change, so always check the current page). It is a rough planning aid, not a quote.

Imagine a customer-support agent that handles 1,000 conversations a month. Each conversation takes about 6 model calls (understanding the question, looking up an order, checking a policy, writing replies). Each call sends about 8,000 input tokens (instructions, tool descriptions, history and retrieved text) and gets back about 400 output tokens.

That is 48 million input tokens and 2.4 million output tokens a month.

• Claude Haiku 5.5 (small, fast): $0.10 input / $0.50 output per million tokens, so about $6 a month
• Claude Sonnet 5.5 (mid-range): $2 / $10 per million tokens, so about $120 a month
• Claude Opus 5.5 (most capable of the three): $4 / $20 per million tokens, so about $240 a month

Three things stand out:

• The model you choose changes the bill more than anything else. Many agents use a small model for simple steps and a larger one only where it is needed.
• Input usually costs more than output in agents, because instructions, tools and history are re-sent on every call.
• Model fees are often smaller than people expect, while the people-time to monitor and improve the agent is often larger.

How can you keep running costs down?

• Use prompt caching. When the same instructions and tool descriptions are sent again and again, providers can cache them. On Anthropic's API, cache reads cost a fraction of normal input (0.1x on most models, and less on some newer ones), so the repeated part of each call becomes much cheaper.
• Batch work that is not urgent. Anthropic's Batch API gives a 50% discount on input and output tokens for requests that do not need an instant answer, such as overnight summaries or tagging old tickets.
• Keep context lean. Send only the documents and history the current step needs. Shorter tool descriptions and trimmed conversation history cut every call.
• Route by difficulty. Let a small model handle routine questions and pass only hard cases to a larger model.
• Set budgets and alerts. Put spending limits and usage alerts on your AI provider account from day one.
• Measure outcomes, not just cost. Track how many conversations were resolved without a person. A cheaper agent that resolves fewer issues can cost you more overall.

What hidden costs should you budget for?

• Staff time to review conversations, correct the knowledge base and approve actions, especially in the first months.
• Changes in connected systems. When your store, CRM or help desk changes its API, the agent's tools need updating.
• Security and data protection work. Limited-permission accounts for the agent, logging, and checks under the privacy laws that apply to you (for UK businesses, see our guide to GDPR-ready AI chatbots).
• Risk management. The US National Institute of Standards and Technology's voluntary AI Risk Management Framework, organised around govern, map, measure and manage, is a useful checklist for deciding how much oversight a customer-facing agent needs.

What should you ask a developer before you agree a price?

• Which tasks will the agent handle in the first release, and which are left for later?
• Which of our systems will it connect to, and with what permissions?
• Which actions need a person's approval?
• Which model or models will it use, and what monthly usage cost do you estimate at our volumes?
• How will you test it before launch, and what results will we see?
• Who monitors it after launch, and what does ongoing support cost?
• Who owns the prompts, code and data, and can we move to a different AI provider later?

A good developer will answer these clearly and will usually suggest starting smaller than you planned. If you would like help scoping an agent, our AI chatbot and agent development team builds agents for businesses in the USA, UK, Canada and Australia, starting with a short discovery call.

Sources

• Anthropic: Pricing (Claude API documentation) — model prices, prompt caching multipliers, Batch API discount, tool-use tokens, web search pricing
• Anthropic: Building effective agents
• Anthropic: Tool use overview
• NIST: AI Risk Management Framework