AI chatbot pricing: what you really pay for

AI chatbot prices look simple until you try to compare them. One vendor charges per seat, another per “resolution”, a third per message, and a fourth sells credits. The sticker price says little about what you will actually pay once customers start asking questions. This guide explains the pricing models you will meet, what really drives the cost of each AI answer, and how to estimate your monthly bill before you sign up for anything.

The five ways AI chatbots are priced

Almost every AI customer support chatbot uses one of these models, or a mix of them:

1. Per seat. You pay for each person on your team who uses the tool. Predictable, but the AI itself does the work, so you end up paying for logins rather than for answers. It is common in help desks that added AI to a product built for human agents.
2. Per resolution or outcome. You pay each time the AI “resolves” a conversation. It sounds fair, but read the definition: a resolution is often counted when the customer simply stops replying. A busy month can produce a surprisingly large bill, and the cost per resolution is usually far higher than the AI actually costs to run.
3. Per message or conversation. A bundle of messages per month, with overage charges beyond it. Simple to understand, but every message is priced the same, whether the answer took a cheap, fast model or a large one.
4. Credits (usage-based). Each plan includes a number of credits that AI usage draws from. Done well, credits track what answers really cost, so a short answer costs less than a long one. Done badly, the exchange rate is hidden. Ask what one credit is worth.
5. Flat tiers with caps. A fixed price per month with limits on agents, knowledge or traffic. Easy to budget, as long as you know which limit you will hit first.

None of these is automatically good or bad. What matters is whether the price grows in line with the value you get, and whether you can predict it.

What drives the cost of one AI answer

Behind every answer, the chatbot calls a large language model. The AI provider charges by the token, a piece of a word, for both what goes into the model and what comes out. So the cost of an answer depends on:

• The model. Prices between models differ by a factor of ten or more. A fast, efficient model answers most support questions as well as a premium one, at a fraction of the cost.
• The context. A support chatbot sends the model the passages from your documentation that match the question, plus the recent conversation. More context means more input tokens.
• The length of the answer. Output tokens usually cost several times more than input tokens. Concise answers are cheaper, and customers prefer them.
• Extra reasoning. Some models can “think” before they answer. That helps with hard problems, but adds tokens that most support questions do not need.

Indexing your content (reading your website and files and making them searchable) also uses AI, but it is a small, mostly one-off cost compared with answering questions every day.

How to estimate your monthly cost

You only need three numbers:

1. Conversations a month. Start from your support inbox and your website traffic. If you do not know, count last month's support emails and chat messages.
2. Questions per conversation. Most visitors ask one to three questions. Use 2 to 3 if you are unsure.
3. Cost per answer. Your vendor should be able to tell you this for its default model. If it can't, that is a warning sign.

A worked example. A small software company has 1,200 conversations a month, with 2.5 questions in each. That is 3,000 answers. On Aidfolk, an answer with the default models uses about half a credit, so the month needs roughly 1,500 credits. That fits in the Starter plan's 2,000 monthly credits, at $29 a month.

How Aidfolk prices: plans with AI credits

Aidfolk combines flat plans with credits, and tells you exactly what a credit is: 1 credit is $0.001 of AI usage at the provider's list price. Each answer uses credits for what it really costs, so a short answer uses less than a long one. With the default answer models, a typical answer uses about half a credit.

A few details that matter when you compare:

• No charge per seat or per resolution. Starter includes 3 team members and Pro 10, and a conversation costs the same whether or not the customer comes back.
• Extra credits when you need them. Credit packs (2,000 for $15, 10,000 for $59) are valid for 1 year and are used after the month's included credits, so a busy month doesn't have to take your chatbot offline. The smaller pack covers about 4,000 extra answers.
• Bring your own AI key on Pro. Answers that run on your own provider key don't use Aidfolk credits: you pay your AI provider directly.
• Yearly billing gives 2 months free on paid plans.
• A real free plan. Free includes 50 credits a month (about 100 answers): enough to try Aidfolk on your own content, with no card.

See every limit side by side on the pricing page.

Questions to ask any vendor before you buy

• What does one answer cost with the default model, and can I choose a cheaper or better one?
• What happens when I run out: does the bot stop, slow down, or bill overage automatically?
• Is removing the vendor's branding from the widget an extra charge?
• How much content can I add, and how many web pages does it read each month?
• Are there setup fees, minimum contracts or annual-only plans?
• Can I see every conversation, and export my data if I leave?

Is an AI chatbot worth the money?

Compare it with what answering the same questions costs today. Take the number of repetitive questions your team answers in a month, multiply by the minutes each takes, and price that time. Even a few hundred answered questions a month usually costs far more in staff time than an AI plan, and the chatbot answers at night and on weekends too. The real return comes from customers getting an answer in seconds instead of waiting for a reply.

One caution: a chatbot is only as good as the content behind it. Budget an hour or two to fill the gaps it finds. Our guide to training a chatbot on your own data shows how.