Guide
What actually drives AI chatbot cost?
Updated:
Chatbot cost is not one number: setup, usage and maintenance are separate line items. Most disappointing projects skipped the third.
The honest answer to “what does a chatbot cost?” starts with “which chatbot?”. An assistant that answers frequent questions and one that can look up an order and start a return differ enormously in both cost and risk.
Three separate line items
Quotes that ignore the third item look cheap. But an unmaintained assistant keeps answering customers from outdated knowledge — it does not solve a problem, it creates a new one.
What drives the usage cost?
Model providers charge by the volume of text processed. In practice four variables decide the bill:
| Variable | Effect | Controllable? |
|---|---|---|
| Questions per month | Direct multiplier | Partly — triage reduces it |
| Context sent with each question | High | Yes — good design cuts it sharply |
| Model tier chosen | High | Yes — not every question needs the strongest model |
| Answer length | Medium | Yes — can be capped |
The most common mistake is the second row: sending every document with every question. The correct design selects the relevant passages and sends only those. Two setups doing the same job can differ several times over in cost because of this alone.
The third row matters too: routing simple frequent questions to a smaller, cheaper model and only complex ones to a stronger model reduces cost noticeably. Sending everything to the strongest model is the most expensive and usually unnecessary route.
Off-the-shelf or custom build?
The honest answer: for most businesses an off-the-shelf product is the better choice. Subscription assistants that answer from your content and embed into your site have matured, and they set up quickly.
A custom build earns its place in three cases:
We measure which case you are in during discovery. If a ready product is enough, we say so — even when it is against our interest.
What failed projects have in common
Chatbot projects that do not work almost always share three gaps: no success metric, no handover path to a human and content that was not ready.
Without a metric the project ends at a demo. Without a handover the customer is lost at the first difficult question. If the content is scattered the answers are scattered too — an assistant cannot be better than what it is fed. On most projects the real work is not configuring the model but organising the content, and that is what sets the timeline.
Frequently asked
What does a chatbot cost?
A single number would mislead, because there are three separate items: setup (one-off), usage (monthly, tied to question volume) and maintenance (monthly, fixed). A narrow setup answering frequent questions from your own content sits far below an assistant that connects to your systems and looks up order status. We define the scope in writing during discovery and show all three items separately — quotes that give one number have usually left maintenance out.
Is the monthly usage cost predictable?
Yes, though not fixed. It varies with question volume and the amount of text processed per question. Predictability comes from design: selecting only the relevant content passages, routing simple questions to a cheaper model, and capping answer length. A monthly ceiling can also be defined — when it is reached the assistant hands over to a human rather than shutting down.
Can we reduce our support team?
That is not the realistic goal, and proposals promising it deserve careful reading. What you can expect is that a share of repeat questions is answered automatically and the rest reaches the right person with context attached. The team does not shrink, but the same team handles more. Setups that remove humans entirely lose the customer at the first hard question.
Who is responsible if it answers wrongly?
You are — which is why the design decisions matter. Three measures reduce the risk: the assistant answers only from your own content rather than inventing, it hands over to a human when unsure, and it redirects rather than answering on binding topics such as pricing or commitments. Every question and answer should also be logged, both for quality and for any later dispute.
How long does it take to build?
A narrow setup is usually 2–5 weeks. What sets the timeline is not the model but the state of your content: if frequent questions and their answers already live in one organised place it moves fast; if the knowledge sits in different people’s heads and scattered documents, organising it is the real work. That is why we ask you to answer “what will the assistant know?” in writing before the project starts.
Which questions should we automate?
Share the questions your support team receives most. We will measure how many can be answered automatically and what that is worth.