How much does an AI chatbot for customer service cost?

The cost of an AI chatbot for customer service depends far more on how the vendor bills AI than on the AI itself. Some tools charge a fee for every conversation the bot resolves, some sell bundles of AI conversations, and some let you connect your own API key so you pay the AI provider's token price directly. For the same number of chats, those models can end up an order of magnitude apart.

This guide explains the three pricing models, shows how to estimate token costs with real, checked list prices, and gives you a formula to compare offers for your own volume.

The three ways AI chatbots are priced

1. Per resolution (or per outcome)

You pay a fixed amount every time the bot closes a conversation without a human. Examples from the vendors' own pricing pages (checked 6 November 2026):

Vendor	AI pricing	Source
Intercom (Fin)	$0.99 per outcome	intercom.com/pricing
Zendesk	$1.50 per automated resolution (committed), $2.00 pay-as-you-go	zendesk.com/pricing

Upside: you only pay when the bot actually solved something. Downside: the price per answer is fixed no matter how simple the question was. "What are your opening hours?" costs the same as a complex return. And the fee comes on top of the seat price.

2. Bundles of AI conversations

You buy a monthly package of AI conversations, for example Tidio's Lyro add-on, which starts at $32.50 per month for 50 conversations (tidio.com/pricing, checked 6 November 2026). Predictable, but you pay for the bundle whether you use it or not, and the next tier is a step up once you outgrow it.

3. Your own API key (BYOK)

The chatbot runs on your account with an AI provider such as Anthropic or OpenAI. The vendor charges for the software; the provider bills you for the tokens your chatbot uses, at their list price. There's no AI margin in between and no cap on conversations – you pay for what's actually used.

This is how asktrix's AI agent Trix works: you connect a key from Anthropic, OpenAI, OpenRouter, Gemini, Mistral and others, and asktrix adds nothing on top.

What does one AI answer cost in tokens?

AI providers charge per token – roughly a short word or part of a word – with separate prices for what goes in (input) and what comes out (output). For a support answer, the input is the bigger part: the instructions, the help articles the bot found for this question, and the conversation so far. The output is the reply itself.

Here are Anthropic's list prices per million tokens, from claude.com/pricing (checked 6 November 2026):

Model	Input	Output
Claude Haiku 5.5 (prompts up to 100K tokens)	$0.10	$0.50
Claude Sonnet 5.5	$2	$10
Claude Opus 5.5	$4	$20

A worked example

Let's assume a typical support conversation where the bot answers three times. These are illustrative assumptions, not measurements – your numbers depend on how long your articles and conversations are:

• Each answer sends about 5,000 input tokens (instructions, a few relevant article passages, the chat so far).
• Each reply is about 300 output tokens.
• Three answers per conversation → 15,000 input and 900 output tokens.

Model	Input cost	Output cost	Per conversation
Haiku 5.5	15,000 × $0.10 / 1M = $0.0015	900 × $0.50 / 1M = $0.00045	≈ $0.002
Sonnet 5.5	15,000 × $2 / 1M = $0.03	900 × $10 / 1M = $0.009	≈ $0.04

Even if your conversations are twice as long as this example, a mid-range model stays in the range of cents per conversation. A small, fast model costs fractions of a cent.

The formula to compare offers

To compare tools fairly, put everything on one line per month:

Per resolution: seats × seat price + resolutions × fee per resolution

Bundle: seats × seat price + bundle price (+ next tier if you exceed it)

Own key: software price + conversations × tokens per conversation × token price

Two things make the comparison honest:

1. Count all conversations for BYOK, not only resolved ones. With your own key you pay tokens for every conversation the bot touches – including the ones it hands over to a human. Per-resolution pricing only charges for successful ones.
2. Count seats. A per-resolution fee usually sits on top of a per-seat plan. A team of six on a $19-per-seat plan already pays $114 a month before the first AI answer.

Example: 500 AI conversations a month

Say the bot handles 500 conversations a month and resolves 300 of them alone.

• Per resolution at $0.99: 300 × $0.99 = $297 per month in AI fees, plus seats.
• Own key, Sonnet 5.5, using the example above: 500 × $0.04 = about $20 per month in tokens, plus the software.
• Own key, Haiku 5.5: 500 × $0.002 = about $1 per month in tokens.

The point isn't the exact figure – your token count will differ – but the order of magnitude. Per-resolution fees are priced per answer; tokens are priced per word.

Hidden costs to look out for

Whatever model you choose, check these before signing up:

• Seat minimums. Some plans require three or more seats.
• AI only on higher tiers. The bot may be locked to an expensive plan.
• Knowledge limits. How many articles, pages or files can the bot read?
• Answer quality controls. Can you test answers before going live? Does the bot show sources? A bot that answers wrongly costs you more than tokens.
• Your time. Every AI chatbot is only as good as the help articles behind it. Budget an afternoon to write the five to ten answers you give every day – our guide on chatbot vs. live chat explains why the handover to a human matters just as much.

What you pay with asktrix

asktrix sells the software and never resells AI:

Plan	Price (net, per month)	AI agent
Free	$0, 2 seats	– (live chat and inbox only)
Start	$19.99, 2 seats	Trix answers from your help center articles
Team	$44.99, 6 seats	Plus web pages, files, short answers, learning notes and quiet hours
Business	on request	Several keys with automatic failover

On every plan with Trix, the token costs go straight to your provider at their price. If you connect a key, your team also gets reply suggestions, summaries and live translation in the shared inbox – on the same key. Full details on the pricing page.

How to keep token costs low

• Pick the right model. Simple FAQ questions rarely need the largest model. Start small, test the answers, move up only if quality demands it.
• Write focused articles. Short articles that answer one question each mean less text sent per answer – and better answers.
• Set a spending limit at your provider. Anthropic and OpenAI both let you cap or monitor usage in their consoles.
• Hand over early. If the bot isn't sure, a quick handover to a human is cheaper than three rounds of guessing.

FAQ

Is a BYOK chatbot always cheaper? For most small and mid-sized support teams, yes, because token costs per conversation are small compared with per-resolution fees. It can be different if you use a very large model with very long prompts – so test with your own content and look at your provider's usage dashboard after a week.

Do I need technical skills to use my own key? No. You create a key in your provider's console, paste it into the chatbot settings and you're done.

What happens if my key runs out of credit? The bot can't answer until you top up. In asktrix, a conversation the AI can't answer stays in your team's inbox, so your live chat keeps working and your team replies as usual.

Want to try it with your own numbers? Create a free workspace, start a 14-day trial of Start or Team, connect your key and watch the usage in your provider's console.