Custom AI Chatbot Cost: Build Price vs Monthly Running Cost
Develop It Today, , 2 min read
A custom AI chatbot has two costs: the one-off build and the monthly model usage. Here is what each looks like and what changes the total.
When someone asks what a custom AI chatbot costs, they are usually asking two questions at once: what does it cost to build, and what does it cost to run. They have different answers and different owners, so it helps to separate them.
The build cost
We price by what the system has to do:
- AI chatbot: $1,997, 10 days. A support chatbot trained on your own material that answers questions.
- AI agent: $4,997, 14 days. A system that takes actions, not just answers. It can check an order, decide whether it qualifies, and update the CRM.
- RAG system: $5,997, 14 days. Retrieval over a large document set. It starts with a paid pilot on your own documents, so you see how it performs before committing to the full build.
Multi-agent orchestration is quoted individually.
The running cost
Model usage is billed by OpenAI or Anthropic straight to your account. For support chatbots we typically see $20 to $200 a month, depending mainly on how many conversations it handles and how long they are. During scoping you get a projected monthly spend at your expected volume, and we use caching and model routing to keep it down.
Billing goes to you directly because you should be able to see exactly what you use and switch providers whenever you like. There is no markup on tokens.
Chatbot or agent?
A chatbot answers. An agent acts. A chatbot tells a customer your refund policy. An agent checks the order, decides whether it qualifies, issues the refund and updates the CRM.
Agents cost more because they need tool access and guardrails, and getting those right takes longer than getting answers right. If all you need is fewer repeat questions in your inbox, start with the chatbot.
What makes a chatbot expensive to get wrong
The build price is the small number. The costly failure is a bot that confidently invents answers. These are the design choices that prevent it:
- Grounding. Answers come from your documents and policies, with the source cited, instead of from whatever the model remembers.
- Evaluation before launch. We build a test set from your real questions and tickets, measure accuracy against it, and give you the harness to re-run it after any change.
- Escalation. Below a confidence threshold, it hands off to a person by inbox or CRM rather than guessing.
- Logging. You can see what it got wrong and fix it.
Questions to ask any vendor
- Is model usage billed to my account, or marked up?
- How will you measure accuracy before it talks to a customer?
- What does it do when it does not know?
- Which of my tools does it connect to, and is that in the price?
- Is my data used to train anyone's model? On the business API terms of OpenAI and Anthropic it is excluded by default, and your documents should sit in your own vector store.
Next step
Details and scope for each option are on the AI chatbot, AI agent and RAG system pages. If you already know what you want it to do, send a brief and we will come back with a written scope and fixed price.