The short answer
An AI chatbot trained on your documents costs $5,000 to $20,000 to build with Lumeor, usually over two to four weeks. Two questions decide where you land: how messy your source content is, and whether the bot only answers questions or also looks things up in your systems.
A help-center FAQ bot on a website sits around $5,000 to $9,000. A bot that verifies the customer, checks an order and hands the conversation to a person in your help desk sits near $15,000 to $20,000.
If what you actually need is an assistant that searches thousands of internal files and respects who is allowed to see what, that is a RAG knowledge assistant, which is a larger build.
Where the money goes
| Phase | Share of effort | What happens |
|---|---|---|
| Content audit and cleanup | 15–25% | Pick the pages the bot should use, retire outdated ones, fill gaps in common questions |
| Indexing and retrieval | 10–15% | Split content into passages, embed it, tune search so the right passage comes back |
| Prompting and guardrails | 15–20% | Tone, scope, refusal rules, citations, behavior when unsure |
| Integrations and handoff | 20–30% | Help desk handoff, lead capture, order or account lookups |
| Testing | 15–20% | Run real customer questions, fix wrong or vague answers, retest |
| Widget and launch | 5–10% | Styling, analytics, alerts and a short handoff document |
Content cleanup is the line buyers underestimate. A chatbot is only as accurate as the documents behind it. If your returns policy says 30 days on the website and 14 days in an old PDF, the bot will find both and may quote either. Fixing that is cheaper than writing prompts to work around it.
What it costs to run
There is no single monthly fee. A custom chatbot has a handful of usage-based line items, and each one scales with something you can predict:
- Model usage. Billed per token by the model provider. It grows with the number of conversations, how many turns each one runs, and how much retrieved text is sent with each question. Routing routine questions to a small model such as Claude Haiku 5.5 keeps this line low.
- Embeddings. A one-time cost to index your content, plus small top-ups when pages change.
- Search index and hosting. For a few hundred pages the index can live in an existing Postgres database with pgvector. Larger content sets may justify a managed vector database, which is billed on storage and queries.
- Logging and monitoring. Conversation logs and error alerts, billed by data volume if you use a hosted tool.
Our page on monthly AI running costs breaks these down further.
How to keep the cost down
- Clean the content before the build starts. Deleting stale help articles takes an afternoon and saves days of tuning.
- Launch answer-only. Ship a bot that answers from documents and hands off to a person. Add lookups once the logs show which questions people really ask.
- Start with one channel. Get the website answers right before adding SMS or WhatsApp.
- Trial the built-in option. If your help desk already ships an AI agent, a month of real traffic costs less than any custom build and tells you a lot. Our Intercom Fin vs custom support agent comparison covers where each one wins.
When not to build one
Skip a custom chatbot if you get only a few chat questions a day, if nobody owns the content and it changes weekly, or if your help desk’s bot already answers most questions well. Custom work pays off when the bot needs data from your own systems, when per-resolution pricing gets expensive at your volume, or when you want control over which model runs and where the data sits.
How Lumeor prices chatbot projects
We quote a fixed price after reading your content and a sample of real customer questions. The quote lists the content sources, channels, lookups, handoff path and the test set we will use to sign off. If your help desk’s own AI would cover it, we say so before you spend anything. The build steps are in our guide to building an AI FAQ chatbot from your docs.