AI Chatbot Pricing in 2026: Per-Resolution, Credits or Flat Fee?

AI chatbot pricing has never been cheaper per token, yet many businesses are paying more than they planned. This is not a paradox you have to guess at. Gartner put a name on it in August: the “Inference Paradox”. Token prices keep falling, but the cost per AI workflow is rising. Gartner forecasts that inference costs per agentic workflow will increase more than fivefold through 2028.

The market is reacting. On October 22, ThoughtSpot launched AgentSpot with a pricing model based on outcomes: no seats, no tokens, one credit per completed run. A day earlier, OpenAI expanded ChatGPT ads to 26 more countries, bringing the total to 89 markets. Free AI is increasingly paid for with attention and data.

If you are comparing chatbots for your website, the price on the pricing page tells you little. What matters is how you are billed. This guide breaks down the four pricing models you will meet in 2026, what each one really costs as you grow, and the hidden costs most comparisons skip.

Why AI chatbot costs are hard to predict in 2026

Gartner’s explanation is simple. As per-token prices drop, companies route more work to expensive frontier models and to multi-step agents. Analyst Will Sommer described the gap: a simple chatbot reads a question and answers it, while an AI agent “must constantly reason, negotiate, and question itself.” According to Gartner, sending a task to an agentic reasoning model raises inference costs by at least five times compared with a basic chatbot exchange, and often much more as tasks get complex.

For a business website, this matters in two ways:

1. Vendors pass costs on. If your provider runs an agent behind your widget, its costs scale with every reasoning loop, and its pricing will follow.
2. Most website questions do not need an agent. “What are your opening hours?”, “Is this product compatible with X?”, “What does your return policy say?” These are retrieval questions. They need the right paragraph from your documents, not a chain of autonomous actions.

The 4 AI chatbot pricing models compared

1. Per-resolution pricing

How it works: you pay each time the AI “resolves” a conversation. Intercom’s Fin is the best-known example, at $0.99 per resolution. Third-party pricing guides also report a monthly minimum when Fin runs outside Intercom’s own helpdesk, and higher fees for certain outcomes such as lead qualification.

The upside: you only pay when the bot works, in theory.

The catch: your bill grows with your success. The better your chatbot gets and the more traffic you receive, the more you pay, with no ceiling. The definition of “resolved” is set by the vendor, and critics point out that a customer who simply gives up can count as a resolution. Budgeting becomes a forecast exercise.

2. Credits or metered conversations

How it works: you buy a pool of AI conversations or message credits each month. Tidio sells its Lyro AI agent as an add-on billed by conversation, metered separately from the base plan. Chatbase uses message credits, and third-party comparisons note that one AI reply can consume several credits depending on the model you choose.

The upside: lower entry price and a clear starting point.

The catch: two meters running at once (base plan + AI add-on), model choices that quietly multiply consumption, and tier jumps when you exceed a quota. Several 2026 reviews describe self-serve plans capping AI conversations before pushing you to a custom quote.

3. Per-seat pricing

How it works: you pay per human agent using the platform, with AI as a bundled or paid extra. Common among helpdesk and live-chat suites.

The upside: familiar and easy to understand.

The catch: it charges for the people your chatbot is supposed to relieve. If your AI handles more questions, you do not save on seats. You often still pay the AI fees on top.

4. Flat monthly fee with clear quotas

How it works: a fixed monthly subscription covering a set number of chatbots, documents and requests. Upgrade only when you choose to.

The upside: predictable cost. You know your AI chatbot budget for the year on January 1st. No per-resolution surprises, no separate AI meter.

The catch: you need to pick a plan that fits your volume. Good vendors make it easy to start small and move up.

The hidden costs nobody puts on the pricing page

Comparing sticker prices is not enough. Check these four lines before you sign:

• Data cost. A free tier funded by advertising, or a provider that trains on your conversations, is not free. With ChatGPT ads now in 89 markets, ask every vendor plainly: are my visitors’ messages ever used for ads or model training?
• Compliance cost. Under GDPR, you are responsible for where your visitors’ data goes. A chatbot that sends conversations to servers in the US can mean a transfer impact assessment, a Data Processing Agreement review and legal time. Since August 2, 2026, the EU AI Act’s Article 50 also requires your chatbot to disclose that it is an AI.
• Setup and maintenance. An “affordable” chatbot that needs weeks of flow-building, prompt tuning or a developer to install is not affordable.
• Error cost. A chatbot that invents a refund policy or a price costs you far more than its subscription. Grounding answers in your own content is a budget decision, not just a quality one.

How DoxyChat keeps your chatbot budget predictable

DoxyChat was designed for the questions your website visitors actually ask, with a cost profile to match.

• Flat monthly plans, no per-resolution fees. You choose a plan with clear quotas for chatbots, documents and requests. Your bill does not grow because your chatbot is doing a good job. And you can start on the free Discovery plan to test it on your own content.
• A RAG chatbot, not a runaway agent. DoxyChat retrieves the relevant passages from your documents, website or RSS feed, then answers from them only. No open-ended reasoning loops, so no hidden inference spiral, and no invented answers outside your content.
• No ads, no training on your data, hosted in France. Your documents and conversations stay in France, answers are generated by Mistral models via Scaleway, and your data is never used to train public models. Fewer GDPR headaches, less legal overhead.
• Live in 2 minutes. Upload your PDFs or index your site, copy one line of JavaScript, done. No integration project to pay for.

Conclusion: choose a pricing model, not just a price

In 2026, the real question is not “how much does an AI chatbot cost?” but “what happens to my bill when it works?” Per-resolution and credit-based models can be fine for some teams, but they turn success into variable cost. For most small and mid-sized businesses answering questions from their own documents, a flat, predictable plan with a bounded RAG chatbot is the safer bet for both your budget and your data.

Want to see more guides like this? Visit the DoxyChat blog.

Try DoxyChat free: build a chatbot from your own documents in minutes, on a free plan with no per-resolution fees.