What AI agent development costs, and where the money goes.

The short answer

There are three ways to put an AI agent to work, and each is billed differently.

The vendor agents cost little to set up and the most to run at volume. A custom agent costs more up front and far less per task. Hosting the model yourself adds setup and running work, and it is the only option in which your data never leaves your own servers.

A worked example: 10,000 tasks a month

These are illustrative numbers. The point is the shape of the costs, so swap in your own volumes.

An operations agent handles 10,000 tasks a month: answering a customer, updating an order, checking a record. Each task makes several model calls, and across them it reads about 8,000 tokens (its instructions, the documents it looked up, the conversation and the results of its tool calls) and writes about 1,000.

The same 10,000 tasks bought as vendor agents:

The gap is smaller than it looks. A vendor's price covers the platform, the interface, the integration with its own product and years of maintenance you don't pay engineers for, and the model cost of a custom agent covers none of that. What the arithmetic does show is where the break-even sits: at a few hundred tasks a month a vendor agent is usually the cheaper buy, and at tens of thousands the build usually pays for itself.

Two things lower the model bill further. Prompt caching charges a fraction of the input price for the instructions and documents an agent re-reads on every call: Anthropic lists cached reads at $0.10 per million tokens for Haiku 4.5. Batch processing, for work that can wait, costs half.

Estimating your own usage

The token numbers above are assumptions. Yours will be different, and a short pilot measures them directly, because every model API reports the tokens each call used.

1. Collect 50 to 100 real tasks from last month: emails, tickets, calls or orders, with names removed if needed.
2. Run them through a prototype of the agent and record the input and output tokens for every task, including the extra calls it makes along the way.
3. Multiply the average by your monthly volume, and price it against two or three models.
4. Check the answers against what a person did. The cheapest model that meets your accuracy bar is the one to budget for.

The pilot also tells you something the price lists can't: how often the agent needs a person, which is usually the bigger cost.

What moves the build cost

The model is rarely the expensive part of building an agent. These are:

• The systems it touches. Reading from one well-documented API is a small job. Writing to a CRM, an ERP and a scheduling system, each with its own rules, is most of the project.
• Whether it acts or only answers. An agent that changes records needs permissions, approval steps for anything that matters, and an audit trail.
• How right it has to be. A test set built from real cases, run on every change, is what tells you the agent still works. Building that set takes time, and it is the part cheap projects skip.
• Where the data may go. If data can't be sent to a model vendor, the agent runs on an open-weight model on your own hardware, which adds setup and operations work.
• The channel. A voice agent adds telephony and a latency budget. Our voice agent cost guide covers that case.
• What happens when it isn't sure. A good handoff to a person, with the context attached, is design work, and it decides how much people trust the agent.

Costs after launch

• Model usage, which grows with volume and with how much context each task reads.
• Hosting, logging and monitoring.
• Re-testing when a model changes. Vendors retire models and change prices: Google's own pricing page lists Gemini 3.8 Flash at double its current price from January 1, 2027. An agent with a test set can be moved to another model in days. One without a test set becomes a guess.
• Tuning as your processes, products and policies change.

Running the model yourself

For a dental group that could not send patient audio to outside services, we ran speech recognition and an open-weight language model on a single GPU server inside the practice's network. There is no per-token cost and no third party in the data path. The trade is that you pay for the hardware and someone has to run it. Read the case study.

Why we don't print a build price

Published agent development price lists disagree by an order of magnitude because they price different things. A chatbot answering from a document folder and an agent that writes to your ERP under approval rules are not the same job, and a single number would be a guess.

We fix the price before the build instead. The two-week Automation Audit maps the workflow and the systems, and ends with a fixed scope and a fixed price. If a vendor agent is the better buy, the audit says so, and we will not quote the build. For market rates on software work in general, see what custom software development costs.

Which option wins

• Buy the vendor agent when the work lives in one product you already use, volume is modest, and your data can be processed by that vendor.
• Build a custom agent when the work crosses several systems, the volume makes per-outcome pricing expensive, or you need to control how it decides.
• Host the model yourself when data can't leave your network, or when a regulator or a customer contract says where it may be processed.

Our AI automation page covers how we pick the first workflow to automate.

Sources

Prices as published on 25 September 2026. They change often, so check them before you decide.

• Anthropic: Claude pricing
• OpenAI: API pricing
• Google: Gemini API pricing
• Salesforce: Agentforce pricing
• Fin AI Agent pricing
• Microsoft: Copilot Studio pricing