AI chatbots
How much does an AI chatbot cost? Build price, running cost and when not to build one
By Waleed Siddiqui, founder & lead engineer
"How much does an AI chatbot cost?" has three honest answers, because there are three very different things people mean by "a chatbot": a subscription tool you switch on, a platform you configure, and a custom assistant built into your product and systems. They differ in price by two orders of magnitude.
This article explains each, what drives the cost of a custom build, and what it costs to run a chatbot once it is live — with worked arithmetic at current API list prices. We also tell you when you should not pay us, or anyone, to build one.
Three ways to get an AI chatbot
| Option | What it is | Typical cost | Good for |
|---|---|---|---|
| Off-the-shelf tool | A helpdesk or website-chat product with an AI assistant feature that reads your help centre | A monthly subscription, often priced per seat or per resolved conversation | Standard support questions answered from public help articles |
| Configured platform | A no-code bot builder or helpdesk AI, set up by you or a consultant with your content, tone and hand-off rules | Subscription plus days to a few weeks of setup | Support and lead capture with modest integration needs |
| Custom build | An assistant built for your product, data and systems, with its own retrieval, actions and evaluation | A project budget — our RAG knowledge-base builds start at $20,900 | Private or messy data, actions in your systems, in-product assistants, strict accuracy or privacy requirements |
What drives the cost of a custom chatbot
A chatbot that answers from one tidy help centre is a modest project. These are the factors that move the price up:
| Driver | Cheaper end | Expensive end |
|---|---|---|
| Content sources | One help centre or a folder of documents | Several systems: wiki, PDFs, tickets, product database, CRM |
| Content quality | Current, consistent articles | Scanned PDFs, contradicting versions, knowledge that lives only in people’s heads |
| Answers or actions | Only answers questions | Looks up orders, changes bookings, creates tickets — with the user’s own permissions |
| Who can see what | Everyone sees the same content | Per-customer or per-role access, so one client’s documents never reach another |
| Channels | One website widget | Website, in-app, Slack or Teams, and helpdesk drafts |
| Accuracy bar | Helpful most of the time, with hand-off | Measured against a large test set; regulated or high-stakes answers |
Notice what is missing: the number of screens. A chatbot's cost is almost entirely in retrieval (finding the right passage), guardrails (what it may say and do) and evaluation (knowing how often it is right). That is also where cheap builds cut corners.
For reference, these are our published starting prices: a RAG knowledge base with up to three content sources, hybrid search, permission-aware retrieval, citations and accuracy measured on your real questions starts at $20,900 (6–8 weeks). Adding one AI feature to an existing SaaS — such as an in-app assistant over data you already have — starts at $13,900.
Running cost: worked arithmetic per 1,000 conversations
Model usage is billed per token, separately for input (instructions, retrieved passages, the conversation so far) and output (the reply). A grounded support conversation is input-heavy, because every turn sends the retrieved passages and history again.
Take a typical support conversation: 4 turns, adding up to 12,000 input tokens and 1,200 output tokens in total. At list prices published on the providers' pricing pages on 20 October 2026 (standard tier, before caching):
| Model (per 1M tokens, in / out) | Per conversation | Per 1,000 conversations |
|---|---|---|
| Claude Sonnet 5.5 ($2 / $10) | 12,000 × $2 ÷ 1M + 1,200 × $10 ÷ 1M = $0.036 | $36 |
| OpenAI gpt-6.1-sol ($2 / $10) | $0.036 | $36 |
| Gemini 3.6 Flash ($0.75 / $3.75, through 31 Dec 2026) | $0.009 + $0.0045 = $0.0135 | $13.50 |
| Claude Haiku 5.5 ($0.10 / $0.50) | $0.0012 + $0.0006 = $0.0018 | $1.80 |
So the model bill for 1,000 conversations a month ranges from a couple of dollars to a few tens of dollars. For most businesses the token cost is small next to the build and maintenance — if the bot is designed for it. What changes the number:
- Model choice, decided by your test set. Many support questions can be answered well by a small model. Running your real questions through two or three models before launch tells you which is the cheapest that is still accurate enough.
- Prompt caching. System instructions repeat on every turn and are billed at a fraction of the normal rate when cached.
- How much context you send. Retrieving eight long passages for every question multiplies cost; good retrieval sends fewer, better ones.
Prices change often — re-run this with current prices and your own token counts from a prototype. We measure tokens per conversation during testing so the monthly cost is known before launch.
The other running costs
- Search infrastructure. A vector index — often just the
pgvectorextension in the Postgres database you already pay for — plus re-indexing when content changes. - Monitoring and evaluation. Logging conversations, reviewing failures, and re-running the test set when you change prompts, content or models.
- Content upkeep. The most underestimated cost. A chatbot is only as good as the content it answers from; someone has to fix the gaps it reveals.
- Maintenance. Providers retire models and change APIs. Budget a small amount of engineering each month.
Why very cheap custom chatbots get expensive
It is possible to wire a document folder to a model API in an afternoon, and quotes based on that demo can look very attractive. The costs arrive later:
- Confident wrong answers. Without a rule that the bot must say "I don't know" when retrieval finds nothing, it improvises — including refund policies you do not have.
- No way to know if it is getting worse. Without a test set, every prompt or model change is a guess, and customers find the regressions.
- Data leaking between customers. Without permission-aware retrieval, one client's documents can appear in another's answers.
- No hand-off. Customers get stuck in a loop instead of reaching a person with the conversation attached.
If you already have a chatbot that gives wrong or vague answers, the fix is usually retrieval and evaluation, not a rebuild — our chatbot development service includes fixing existing bots.
How to get an estimate you can trust
Bring three things to any vendor: a sample of the content the bot should answer from, 20–50 real questions customers or staff ask (with the right answers), and a list of systems it must read from or act in. A quote given without seeing these is a guess. And ask every vendor the same question: "How will we know how often it is right?" The answer separates a production chatbot from a demo.
