Simple Chatbot vs AI Chatbot: What They Actually Cost to Run

This page used to tell you that AI chatbots resolve 78% of queries against 52% for rule-based ones, generate 3.4 times more leads, and return $8 for every $1 spent. We went looking for where those numbers came from and could not find an origin for any of them. They appear on hundreds of agency blogs, always without a study attached. So we deleted them.

What follows is the version we can stand behind: the cost of running an AI chatbot worked out from prices the vendor publishes, and an honest account of which of the two options is the better buy.

What an AI chatbot actually costs per conversation

You can calculate this yourself, which is the point. Anthropic publishes per-token prices for the Claude API. At the time of writing, Claude Haiku 4.5 is $1 per million input tokens and $5 per million output tokens. Claude Sonnet 5 is $2 and $10 — a price Anthropic confirmed in August 2026 would become permanent rather than rising to $3 and $15 as originally scheduled.

Our method: assume a support conversation runs six back-and-forth turns. Each turn sends the system prompt plus whatever business information the bot needs to answer, which we budget at roughly 3,000 input tokens, and comes back with about 200 output tokens. That is 18,000 input tokens and 1,200 output tokens across the conversation.

Two things follow. The first is that a thousand answered conversations a month costs somewhere between roughly $24 and $48 in model tokens, not the several hundred dollars usually quoted. The second is that if you see a much higher number, it is not the tokens. It is a per-seat SaaS subscription, and you are paying for the vendor's dashboard rather than for compute.

Prompt caching cuts the input side further when your system prompt and business context stay the same between turns, which for a support bot they almost always do. We have not put a figure on that because how much you save depends entirely on how your prompt is structured.

Where the money actually goes

If model tokens are cheap, the cost has to be somewhere else, and it is. It is in making the bot answer correctly.

A scripted bot cannot be wrong about your business, because it can only repeat sentences you wrote. An AI bot generates its answer, so it needs to be grounded in your real information and then tested to confirm it stays grounded. That testing is the line item nobody quotes you for.

There is published work on how hard this is. In Seven Failure Points When Engineering a Retrieval Augmented Generation System (Barnett et al., 2024), the authors built and studied three retrieval-augmented systems across research, education, and biomedical domains, and catalogued where they broke. Their conclusion is the useful part for a business owner: they argue that validating a system like this is only feasible during operation, and that its robustness evolves rather than being designed in at the start.

Read that as a budget warning rather than a technical footnote. You cannot fully test an AI chatbot before launch. Some of the work happens after it goes live, watching real questions and correcting what it gets wrong. Budget for the weeks after launch, not just the build.

The case for a scripted bot

A simple chatbot maps questions to answers you wrote in advance, using buttons, keywords, or a decision tree. It is unfashionable and frequently the correct choice.

• It cannot invent anything. For a business where a wrong answer is expensive, that guarantee is worth more than fluency.
• The cost is knowable up front. No per-conversation charge, no bill that scales with a traffic spike.
• It is fast to change. Editing an answer takes a minute and does not require re-testing whether the bot still behaves.

The trade-off is real and it shows up quickly. A scripted bot is stumped by anything you did not anticipate, and the maintenance never ends, because every change to your business is a change you make by hand.

The case for an AI chatbot

An AI chatbot reads what someone meant rather than matching what they typed. It handles a question phrased three different ways, a question with two parts, and a follow-up that only makes sense in the context of the previous message.

That is worth paying for when the range of questions is genuinely wide. If you sell one product and take bookings, it is not. If you have a large catalogue, detailed policies, or customers who ask about combinations of things, a script becomes a losing battle long before it becomes a good experience.

How we decide with clients

We start with the questions, not the technology. If you can export a month of customer emails or chat logs, we read them and count how many distinct things people actually asked. Below roughly twenty recurring questions, a scripted bot usually covers it and the AI is a solution to a problem you do not have. Above that, the script starts costing more to maintain than the AI costs to run.

We build both as part of our AI and automation work, and we have talked clients out of the more expensive option often enough that it is worth saying plainly: if a scripted bot answers your questions, buy the scripted bot. If you want to go a level deeper on the AI side, our guide on AI chatbots versus RAG covers what changes when the bot has to read your documents.

If you would rather have someone work through the numbers with you against your own volumes, we do that in a free consultation, and we will tell you if the answer is the cheaper one.