AI chatbot for website cost: what you pay to build and run one
The cost of an AI chatbot for a website has two parts that behave very differently: a one-off cost to build or configure it, and a running cost that grows with every conversation. A ready-made SaaS widget keeps the first part small and charges per seat or per resolved conversation. A custom chatbot built on your own data costs more to build, but you pay model providers directly for usage and own the result.
Neither is cheaper in every case. The answer depends on conversation volume, how much of your content is clean and current, which systems the bot must query live, and how much accuracy matters in your sector.
This guide breaks down each cost line, shows how usage pricing works using vendors' published rates as checked in November 2026, and explains how to judge whether the bot is paying for itself.
What you are actually paying for
A modern website chatbot is not a script of canned replies. It retrieves passages from your content, passes them to a large language model with the visitor's question, and returns an answer, sometimes calling your systems for live data such as order status or appointment slots. Each of those steps has a cost, and a quote that does not mention them is incomplete.
- Content preparation: auditing, cleaning and structuring the documents the bot will answer from.
- Ingestion and indexing: splitting content into passages and creating embeddings stored in a vector database.
- Model usage: tokens sent to and generated by the language model for every message.
- Integrations: live lookups into your CRM, booking, order or ticketing systems, and handover to human agents.
- Interface: the chat widget on your site, plus any other channels such as WhatsApp or your app.
- Evaluation and monitoring: test sets, accuracy checks, logging and alerting.
- Ongoing upkeep: keeping content current, reviewing failed answers and adjusting prompts.
SaaS tools bundle most of these into their price. Custom builds itemise them. Either way, content preparation and upkeep fall on your team or your partner, and they are the lines most often missing from early budgets. Our AI chatbots on your data work starts with a content audit for exactly that reason.
Three ways to buy, and how each is priced
Most businesses choose between a SaaS chatbot inside a support platform, a custom chatbot built on their own data and cloud accounts, or a hybrid in which a custom retrieval layer feeds a SaaS front end. Each has a different pricing shape, which matters more than the headline number.
| Option | Typical pricing shape | Published example | Best suited to |
|---|---|---|---|
| SaaS AI agent in a support platform | Per seat for human agents plus a fee per AI-resolved conversation | Intercom lists its Fin AI agent at $0.99 per outcome, with a minimum monthly commitment | Teams already on that helpdesk with moderate volume |
| Custom chatbot on your data | One-off build, then direct model, hosting and maintenance costs | Model providers publish per-million-token rates; see the next sections | Higher volume, sensitive data, live system lookups, ownership |
| Hybrid | Build for retrieval and integrations, SaaS for the inbox and widget | Varies by components | Businesses wanting a ready agent inbox with custom data handling |
| Bot on WhatsApp via Meta's agent platform | Per-token charge by Meta for each AI reply | Meta lists its Business Agent at $2.00 per million tokens, estimating about 4 to 5 US cents per message | Businesses where most conversations happen on WhatsApp |
Intercom's pricing page says an outcome is counted when a customer confirms the issue is resolved, does not ask for more help after the AI responds, or the AI completes a workflow, and that it charges once per conversation. Per-outcome pricing is predictable per success but scales directly with volume, so model it at your expected monthly conversations, not your launch numbers.
What drives the build cost
Build cost for a custom chatbot depends far more on scope than on the model chosen. A bot answering from twenty clean help articles is a small project. One that answers from thousands of PDFs, respects user permissions and books appointments in your practice software is a much larger one. The table shows the main drivers, with our own ranges left as placeholders until scoped.
| Cost driver | Smaller scope | Larger scope |
|---|---|---|
| Content volume and quality | A few dozen current, consistent pages | Thousands of documents with duplicates and outdated versions |
| Data sources | One website or help centre | Drive, SharePoint, Notion, PDFs and product databases |
| Live integrations | None, answers from content only | Order status, bookings, CRM writes, ticket creation |
| Access control | Public content only | Logged-in users with document-level permissions |
| Languages and channels | English on the website | Hindi and English, website plus WhatsApp and app |
| Accuracy requirements | General product questions | Healthcare, finance or legal content needing graded evaluation |
Timeline follows the same drivers. A focused chatbot over one or two sources needs a data audit, a prototype, evaluation rounds and an internal pilot before customers see it. Every additional integration or permission layer adds testing time, and accuracy-sensitive sectors add evaluation rounds.
If the bot must also take actions, such as creating tickets, changing bookings or updating CRM records on its own, the scope moves from chatbot towards AI agent, with more testing and guardrails. Our guide to AI agent development cost covers what that extra scope involves and how it changes the estimate.
Running costs: tokens, embeddings, hosting and people
Language models are billed per token, roughly a fragment of a word, with separate rates for input sent to the model and output it generates. OpenAI's API pricing page, for example, lists per-million-token rates for each model, with output priced higher than input. Each chat message sends the question, retrieved passages and instructions as input, so input tokens usually dominate a retrieval chatbot's bill.
Indexing is a smaller cost. OpenAI's pricing page lists its text-embedding-3-small model at $0.02 per million tokens and text-embedding-3-large at $0.13 per million tokens, so embedding a million tokens of content costs between 2 and 13 US cents at those rates. Re-indexing is needed whenever content changes or you switch embedding models.
- Estimate monthly conversations and average messages per conversation.
- Estimate input tokens per message: the question, retrieved passages and system instructions.
- Estimate output tokens per message from the typical answer length.
- Multiply by the chosen model's input and output rates from the provider's current pricing page.
- Add vector database and application hosting, logging and monitoring.
- Add people time: weekly review of failed answers, content updates and prompt changes.
Measuring whether the chatbot pays for itself
The only cost that matters in the end is cost per resolved conversation compared with what that conversation costs today, whether by phone, email or live chat. Measure the current baseline before launch: ticket or enquiry volume by topic, handling time and staff cost. Without it, any claim of savings is guesswork.
| Metric | How to calculate it | Why it matters |
|---|---|---|
| Resolution rate | Conversations closed without a human and without a repeat contact on the same topic within a set window | Stops deflection figures being inflated by people who simply gave up |
| Accuracy on test set | Correct answers divided by questions in a graded set of real past queries | Catches regressions after content or prompt changes |
| Cost per resolved conversation | Model, hosting and SaaS fees divided by resolved conversations | Compares directly with the cost of a human-handled enquiry |
| Escalation reasons | Share of handovers by cause: missing content, policy topic, user request | Shows which content to add next |
| Leads or bookings from chat | Conversions attributed to chatbot sessions | Shows revenue impact, not just savings |
Set targets from your own baseline rather than borrowed benchmarks, and review weekly for the first months. We can share resolution rates from comparable projects at 90 days after launch on a call, but your own pilot data is the better guide.
Questions to ask any chatbot vendor or agency
Quotes for AI chatbots vary widely because vendors scope them differently: one includes content clean-up and evaluation, another assumes your documents are ready and prices only the widget. These questions make proposals comparable and expose where costs will appear later. Ask for the answers in writing and keep them with the contract.
- Which costs are one-off and which recur monthly, and how does the monthly figure change at double the volume?
- Who pays the model provider, and are API keys and billing in our company's name?
- Where do the code, prompts, vector index and evaluation set live, and can we take them if we leave?
- How do you measure accuracy before launch, and what test set will you use?
- How is access control enforced for logged-in users?
- How are chat logs stored, masked and deleted?
- Can we switch the underlying model later, and what would that involve?
- What does ongoing support cover, and who reviews failed answers each week?
If WhatsApp is a major channel for you, ask the same questions of any WhatsApp chatbot proposal too, and check how Meta's per-message charges for service replies add to the model cost. A bot that answers well on your website will usually need shorter replies and different handover rules on WhatsApp.
Getting a realistic estimate for your site
A realistic estimate needs a look at your content, your conversation volumes and the systems the bot must reach. We can start with a short scoping session and a walkthrough demo of a retrieval chatbot answering from a sample of your own published content, so you can judge answer quality before discussing budget. A knowledge assistant for internal teams can follow the same pattern.
To go further, see our AI chatbot development service, or book the Growth Audit: a two-week paid diagnostic, with the fee credited to your next engagement, that includes a content readiness check and a build-and-run cost estimate for your chatbot alongside the rest of your funnel.