How much does an AI chatbot cost to build and run in 2026?
A custom AI chatbot on your own content costs US$8,000 to 40,000 to build at a small studio and US$50 to 500 a month to run. What drives both figures.
By Sohail Hussain29 October 20267 min read
As of October 2026, a custom AI chatbot that answers questions from your own content on your website or WhatsApp costs US$8,000 to 40,000 to build at a small studio, US$2,000 to 10,000 from a freelancer and US$40,000 to 150,000 at a larger agency, and then US$50 to 500 a month to run at typical small business volumes. An agent that also takes actions, such as booking appointments or looking up orders, costs two to four times more to build. An off-the-shelf bot from a help desk platform costs nothing to build and a monthly fee that scales with conversations.
The build price depends mostly on what the bot is allowed to do and how many systems it touches. The running price depends on volume and which model you use. This guide works through both, with real token arithmetic, and ends with the cases where a chatbot is not worth it.
What does each type of chatbot cost to build?
| Type | What it does | Freelancer | Small studio | Larger agency |
|---|---|---|---|---|
| Off-the-shelf bot | A help desk's built-in AI, trained on your help centre | US$0 build; monthly fee | Same | Same |
| FAQ bot on your content | Answers from your website, documents and policies; hands over to a human | US$2,000 to 10,000 | US$8,000 to 40,000 | US$40,000 to 150,000 |
| WhatsApp support bot | The above on WhatsApp Business, with handover to your team | US$4,000 to 15,000 | US$12,000 to 50,000 | US$50,000 to 180,000 |
| Action-taking agent | Also books, looks up orders, creates tickets, updates your CRM | US$10,000 to 40,000 | US$25,000 to 100,000 | US$100,000 to 300,000+ |
| Internal knowledge assistant | Staff ask questions across policies, manuals and documents, with permissions | US$5,000 to 20,000 | US$15,000 to 60,000 | US$60,000 to 200,000 |
The off-the-shelf row deserves a serious look first. Intercom, Zendesk, Tidio, Crisp and similar platforms now include an AI assistant that reads your help centre and answers tickets, priced per seat plus per resolved conversation. If you already use one of these, have a decent help centre, and only need answers rather than actions, it may be all you need. The comparison in chatbot on your website: build versus Intercom and Tidio goes through when each wins.
Custom becomes worth it when the bot must use content that is not in a help centre, take actions in your systems, run on WhatsApp with your own flows, or when the per-resolution fees exceed what a custom bot would cost to run. Our AI agents and chatbots work covers all of those: support agents on websites and WhatsApp that answer from your content, internal knowledge assistants, action-taking agents, human handover rules, guardrails and testing, and usage reporting.
For how this compares with other digital spending, the main guide to how much a website costs puts chatbots next to sites, stores and apps.
What drives the build cost?
Where the knowledge comes from. A bot that answers from 30 clean web pages is a short job. A bot that answers from 400 PDFs, a product database and three years of support tickets needs the content cleaned, chunked, embedded and kept in sync. Content preparation is often a third of the build. The guide to RAG explained covers how a bot answers from your documents and why the quality of those documents sets the quality of the answers.
What it is allowed to do. Answering is one thing. Each action, such as checking an order status, booking a slot or creating a ticket, is an integration with one of your systems, plus rules about when the bot may act and when it must ask. Budget US$2,000 to 10,000 per action at studio rates, as of October 2026, with more for systems that have poor APIs.
Channels. Website chat is the baseline. WhatsApp adds Meta's Business API setup, template approval and per-message pricing. Voice adds speech recognition and synthesis and a different set of failure modes; see voice AI agents.
Handover and guardrails. The difference between a demo and a product is what happens when the bot does not know. Detection of uncertainty, handover to a human with the transcript, limits on what it will discuss, and refusal of off-topic or risky requests. This is unglamorous and it is where a cheap bot fails in public. The post on human handover sets out the rules we use.
Testing. A set of real questions with expected answers, run before launch and after every content change. Studios that skip this ship bots that invent policies. How to test an AI agent before launch describes the minimum.
What does a chatbot cost to run each month?
Four running costs, as of October 2026.
Model tokens. Every message the bot handles sends your instructions, the retrieved content and the conversation so far to a model, and gets a reply back. Take a support bot with 3,000 conversations a month, five exchanges each, around 4,000 input tokens and 250 output tokens per exchange. That is 60 million input tokens and about 3.75 million output tokens a month. At Anthropic's published rates on its pricing page:
| Model tier | Input per million tokens | Output per million tokens | Monthly cost at this volume |
|---|---|---|---|
| Small (Claude Haiku 5.5) | US$0.10 | US$0.50 | about US$8 |
| Mid (Claude Sonnet 5.5) | US$2.00 | US$10.00 | about US$160 |
| Large (Claude Opus 5.5) | US$4.00 | US$20.00 | about US$315 |
OpenAI and Google price their equivalent tiers in similar brackets; check their pricing pages for current figures. Prompt caching, which stores the repeated part of each request, typically cuts the input figure by half or more. Most support bots run on the small or mid tier with the large tier reserved for hard cases. Put your own numbers into the AI API cost calculator to see the bill at your volume.
WhatsApp messages. Meta charges per message for template messages your business initiates, with rates that vary by country and category; since 1 October 2026, replies inside the 24-hour customer service window are also charged once a number has used its 1,000 free service messages for the month (eligible government and non-profit organisations keep them free). The current rates are on Meta's WhatsApp pricing page. For a support bot where customers start the conversation, this is usually small.
Hosting and the vector store. The bot's server, the database that holds your embedded content, and monitoring. US$20 to 150 a month at small business volumes.
Content upkeep and monitoring. Someone reads the conversations the bot handled badly, updates the content, and re-runs the tests. Two to eight hours a month is realistic; on a retainer that is US$100 to 800 a month, as of October 2026. Our post on what a month of AI agent logs taught us shows what you find when you look.
Put together, a small business support bot runs US$50 to 500 a month. The off-the-shelf alternative is priced per resolution and tends to cross that figure somewhere in the low hundreds of resolved conversations a month.
When is an AI chatbot not worth it?
- Fewer than a couple of hundred conversations a month. A clear FAQ page, a WhatsApp button and a human who replies within an hour will cost less and convert better. The chatbot ROI calculator takes your ticket volume, handling time and staff cost and shows the monthly saving; if it is under a few hundred dollars, wait.
- Your content is wrong or missing. A bot answering from an out-of-date help centre gives out-of-date answers with confidence. Fix the content first; it is cheaper and helps humans too.
- Every conversation is different and high-stakes. Legal advice, medical questions, complaints that need judgement. The bot should route these to a person, and if that is most of your volume there is little left for it to do. See which conversations an AI should handle, and which it should not.
- You want it to sell. Bots that push products into every conversation annoy people and are rarely trusted. Answering well is what builds the trust that sells.
The case for a chatbot is strong when the same questions arrive all day, when enquiries come in outside working hours and go cold by morning, or when your team spends hours looking things up that a bot could answer in seconds.
How do you get a quote you can compare?
Describe the job, not the technology: which questions it should answer, from which content, on which channels, what it may do on its own, when it must hand over, and the volumes you expect. Ask each supplier how they will test it, what happens when it does not know, and what the monthly running cost will be at your volume. A quote without a running cost estimate and a test plan is a demo, not a product.
For a rough number before you talk to anyone, run your volumes through the AI API cost calculator for the model bill and the chatbot ROI calculator for the saving, and compare the two. If the gap is wide, the build pays for itself quickly and the quotes above are worth getting.
Was this guide useful?