The cost of building an AI agent depends far more on what surrounds the model than on the model itself: how many of your systems the agent can act on, how much data it answers from, whether it talks or types, and how high the accuracy bar is. An assistant that answers questions from a handful of documents and an agent that books, refunds and updates records are both "AI agents", and they cost very different amounts.
This guide explains what moves the price of custom AI agent development, the common shapes an agent takes, and the running costs people forget, so you can place your idea before anyone quotes.
What drives the cost of an AI agent
| Cost driver | Smaller build | Larger build |
|---|---|---|
| What it can do | Answers questions only | Acts in several systems: books, updates, sends |
| Data it answers from | A few clean documents | Large, messy or fast-changing sources |
| Channel | Text chat | Phone calls, with speech and live handoff |
| Accuracy bar | Helpful most of the time | Must be right, with tests on every change |
| Hosting and privacy | Hosted model APIs | Models and data inside your own environment |
| Volume | Occasional use | Thousands of conversations a day |
The first row matters most. Every tool an agent can call needs permissions, tests and a plan for when it fails, so an agent that only answers is a fraction of the work of one that acts. The accuracy bar is the next biggest lever: an agent that must be right needs an evaluation set of real questions with known answers, run against every change, and that work is a large share of a serious build.
Four common shapes of AI agent
An assistant over your documents
A chat assistant that answers from your help content, policies or product data, with citations. Most of the work is retrieval: how your documents are split and searched so the right passage reaches the model.
An internal agent that acts in your systems
An agent that reads a request and does something about it, such as triaging a ticket, updating a CRM record or drafting a reply for a person to approve. Cost scales with the number of systems and the limits it must respect.
A voice agent
An AI voice agent that answers or makes phone calls, books into a real calendar and hands hard calls to a person. Voice adds speech recognition, natural turn-taking and telephony to everything a text agent needs.
A multi-step agent pipeline
Several specialised agents passing work along, such as one that reads a blog post and others that cut video, write platform copy and schedule it. Each stage is simpler than one large agent, but the pipeline needs orchestration and monitoring of its own.
Running costs that start after launch
An agent has usage costs that a normal app does not, and they are billed by the providers rather than by your developer:
- Model usage, billed by the language model provider per amount of text processed.
- Speech and telephony for voice agents, usually billed per minute of conversation.
- A vector database or search index for retrieval, which grows with your data.
- Monitoring and periodic re-testing, because models and data change and an agent that was right in March can drift by June.
We estimate these with you at scoping from your expected volume, and build a cost ceiling into the agent so a loop cannot run up a bill overnight.
How THWorks prices an AI agent build
THWorks prices the job rather than the hours. After a scoping call you get a fixed price for the agreed scope, settled before the work starts. If a new feature or integration comes up midway, it is priced and agreed with you before it is built, so the total never moves without your say.
The cheapest good agent is one built for a single job first. Once it is doing that job reliably, extending it to the next is a smaller piece of work than building everything at once.
How to get a useful estimate
- Write down the one job the agent should do first, and what a good outcome looks like.
- List the systems it must read from or act in, such as a CRM, a calendar or a help desk.
- Collect twenty real examples of requests it would receive, including awkward ones.
- Decide whether it works in text, on the phone, or both.
- Note any rule about where your data may live.
