Prices for AI services change often and vary by provider, so any figure I printed here would soon be wrong. The structure of the costs is stable and is what to ask about.
The build
- Defining what it should and should not answer.
- Preparing your content so it can be used reliably.
- Connecting it to the website and any systems it needs, such as bookings or a CRM.
- Handling failures, handover to people and limits on abuse.
- Testing with real questions and difficult cases.
The running costs
- Model usage. Most providers charge by the amount of text processed, so cost scales with conversations and their length.
- Hosting of the parts you run yourself.
- Monitoring and improvement. Reading conversations and fixing weak answers is ongoing work.
- Any search or database service used to look up your content.
How to keep it predictable
- Set limits on conversation length and on requests per visitor.
- Estimate from a realistic number of conversations, and add a safety margin.
- Ask the provider about price changes and the exact billing unit.
- Choose the smallest model that answers well; a larger one costs more.
The most helpful number is usually not a total but a cost per conversation, which lets you decide whether the bot saves more than it costs.