AI chatbot cost in 2026: typical US ranges

Off the shelf website chat is often discussed at $50 to $500 per month. That band covers products that add a chat window to a site, a handful of scripted answers or a light sync from a help center, and a handoff to email or a person. The vendor hosts the widget. Someone on your team configures it. The monthly bill is the software plan: seats, a chat volume, or a feature tier. Public plan pages in this band rarely include a writer, a custom integration, or a careful pass over which prices and policies the bot is allowed to repeat.

This band is a reasonable budget for a site that needs a faster front door to questions you have already answered in public. Hours, a service area, a link to book, and a way to reach a person are the usual wins. The band is a weak fit when the bot is expected to know exceptions, to quote a job, or to act inside another system. Those expectations belong in the custom launch conversation, and sometimes they belong in a different product entirely.

Read the plan's volume line before you compare two $50 offers. A low monthly price with a tiny chat cap, and overage after that, is a different yearly cost from a plan that covers the traffic you already have. The software fee is still not the model fee if the vendor meters AI answers separately. Ask which bill you are looking at.

A custom setup trained on the business's own content is often discussed at $2,000 to $15,000 to launch, and then usage. Launch work in that public band usually means choosing the pages or documents the bot may use, keeping drafts and stale offers out, writing the questions it must refuse, connecting the widget to the site, and testing with real questions from the front desk. The dollar span is the setup project. It is the work of deciding what the bot knows. It is not a prepaid bundle of model usage for the year.

A serious launch includes a source list you can update, a refusal list, a path to a person, and a log you can read. The source list is the product. If the bot may quote a service, that service needs a current page. If it must not quote a draft price from an old PDF, that file stays out. Navigation chrome, outdated posts, and internal notes are the usual exclusions. The test set matters as much as the model name: twenty questions the owner actually hears, with the answers you are willing to stand behind.

The width of the $2,000 to $15,000 band comes from how much of that material already exists. A business with a clear set of service pages and policies is closer to the low end of public discussion. A business that needs the knowledge written, translated into several brands, or wired to a help desk and a calendar is closer to the high end, and some of that wiring is integration work beyond a read only bot.

After launch, model providers bill for the tokens or the requests the bot actually uses. Those bills vary with how many chats you receive, how long the answers are, and which model is selected. They are not a flat agency fee. Some software vendors resell usage inside a higher monthly plan. Some pass the provider bill through. A quote should say which arrangement you are in, and what happens if volume doubles. This guide does not invent a token price. Provider rate cards change, and a single number would pretend they do not.

Other extras, unless named in the launch: writing the knowledge base when the site does not already state services, prices, and policies; opening tickets in a CRM or help desk; more than one language; and actions that change a record, book a calendar, or take a payment. Those actions have failure cases, and they sit above a bot that only answers from approved pages. A human review of transcripts for the first weeks is also a scope choice. Voice is a different product. A website chatbot and an AI receptionist that answers the phone have different ranges. If the job is calls, use the AI receptionist cost guide.

A monthly widget in the $50 to $500 band, described as if it had been trained on your private policies, is the first flag. Configuration of a generic script and a custom source list are different jobs. Hidden usage is the second. If nobody can point to the model bill, you do not have a full price. A bot allowed to invent prices, appointment slots, or medical and legal answers is the third. The safe version refuses those questions and hands them to a person.

Also look for a missing handoff, for the phrase unlimited AI with no volume note, and for a launch price that includes no test questions. A low setup that goes live on your whole site, including draft pages and old offers, will answer from the wrong material. Cheap, in this case, often means the source list was never edited. Ask who removes a page when a policy changes, and how fast that removal takes effect.

We separate three numbers. The launch project is a fixed price for the sources, the refusal rules, the widget, the handoff, and the test set. The software plan, if we are using a hosted chat product, is its own monthly line in the public $50 to $500 shape or whatever plan actually fits the volume. Model usage is the third number, and we do not flatten it into an agency fee. If the provider bills by use, the quote says so. If a vendor plan includes a pool of answers, the quote says what happens past the pool.

We do not sell a bot that invents prices or policies. Stale prices stay out of the source list. CodeHunterz prices are confirmed in a written quote. They are not chatbot research and they are not our chatbot fee. Related pages: the AI chatbot service, chatbot and LLM apps, and AI integration. The article on choosing a chatbot for a small business website is the buying walkthrough beside this range guide.