AI Chatbot Pricing: Hidden Costs to Avoid

Understanding per-resolution fees, seat licenses, and AI token costs.

The three ways vendors actually charge, and why the sticker price lies

Almost every chatbot vendor uses one of three billing models, and the headline number on the pricing page rarely tells you what you will pay. Per-seat pricing charges for each human agent with a login. It is predictable, which finance teams like, but it punishes you for growing the team and says nothing about how much work the AI is doing. Per-resolution pricing, which Intercom Fin popularized at roughly a dollar per resolved conversation, flips that logic: you pay when the bot succeeds. Per-conversation or monthly-active-user pricing sits in between and charges for volume whether or not anything got resolved.

On top of the model, most vendors add usage costs. If the bot runs on GPT-4o or Claude, someone is paying the model provider per token, and that cost is usually passed through to you. A large knowledge base makes this worse, because every answer stuffs more context into the prompt, and more context means more tokens per reply. Two vendors can quote the same monthly base and bill very differently once your traffic and your document count are real. The sticker price is the floor, not the bill.

"The line that got us was the definition of a resolution. Our contract counted one every time the bot sent a closing message, including the conversations where the customer immediately asked for a human."

How a resolution-based bill can quietly double

Say you run 5,000 conversations a month and the bot resolves 60 percent of them. At a dollar per resolution, that is 3,000 resolutions, or about 3,000 dollars, before any seats or token overages. Now your knowledge base improves and resolution climbs to 80 percent. Your bill goes up, not down, because you are paying for success. That is not a scam, it is the model working as designed, but plenty of teams budget for the base fee and get blindsided by the usage line.

The worse version of this is a bad knowledge base that sends conversations into loops. If the bot keeps re-engaging because it cannot answer cleanly, your conversation count inflates and so does your bill, and you are paying more precisely because the thing is working badly. Read the fine print on what counts as a resolution or a conversation, ask for a sample invoice at your projected volume, and watch for onboarding fees, annual prepayment requirements, and overage rates that run two to four times the base per-unit price.

None of this makes usage pricing wrong. Below a few hundred conversations a month, a per-resolution plan is often cheaper than a flat subscription, and you should not overpay for capacity you will not use. The trap is signing an annual contract at a volume you have not measured yet. Run one month on a plan you can walk away from, count your real conversations and your real resolution rate, then price the annual deal against actual numbers rather than a projection someone made up in a sales call.

Questions to ask before you sign an annual contract:

• What exactly counts as a resolution? If a conversation that escalates to a human still counts, your effective rate is higher than the quoted one.
• Are model tokens included or passed through? A big knowledge base can quietly make token cost the largest line on the invoice.
• What is the overage rate? Base per-unit pricing is often two to four times higher once you exceed the bundled volume.
• Can I start monthly? Anyone confident in their value will let you measure for a month before locking in a year.