In short
- Scripted bots follow buttons and set answers: predictable, cheap to run, limited.
- AI bots answer in their own words from your content: flexible, but need guardrails and testing.
- A mix suits most businesses: AI for questions, a script for money, bookings and data.
- Whatever you choose, your business is responsible for what the bot says.
The difference
A scripted bot recognises a question (or offers buttons) and gives an answer you wrote. It never says anything you did not write. An AI bot uses a large language model to understand the question and write an answer, ideally only from content you provide, such as your help articles. A hybrid uses AI to understand and answer general questions, and switches to a script for anything that must be exact.
| What matters | Scripted | AI from your content | Hybrid |
|---|---|---|---|
| Understands varied wording | Partly | Well | Well |
| Says only what you wrote | Yes | No: it rewords | Yes, for scripted parts |
| Set-up effort | Writing every path | Preparing content and guardrails | Both, in smaller amounts |
| Running cost | Platform fee only | Plus AI usage | Plus some AI usage |
| Best for | Qualifying leads, bookings, set forms | Large help centres, staff handbooks | Most small business bots |
When a script is better
Choose a script when every customer should get the same questions or the same exact answer: qualifying a lead, taking a booking, explaining a refund policy, or collecting details for a form. Scripts are easy to test because the paths are finite, and they cannot make things up.
When AI is better
Choose AI when there are too many questions to script, the content already exists (a help centre, a handbook), and a small risk of imperfect wording is acceptable. AI handles typos, long messages and questions asked in unexpected ways far better than keyword matching.
Accuracy is the main concern, for good reason. 56% of UK businesses using AI are concerned about bias and accuracy (opens in a new tab). Limit the bot to your own content, tell it to say when it does not know, keep exact facts in scripted answers, and keep a person in the loop to catch mistakes, as the CMA advises (opens in a new tab).
What AI costs to run
AI is charged by the amount of text processed, in tokens. List prices vary widely: Claude Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens (opens in a new tab), and Gemini 3.1 Flash-Lite $0.25 and $1.50 (opens in a new tab); larger models cost many times more. By our own estimate, a six-turn support conversation with your content included uses around 20,000 input and 1,500 output tokens, which is well under a penny on the smallest models.
Helpdesk platforms often charge per resolved conversation instead, which can cost far more per chat. The running cost calculator compares both for your volumes.
You are responsible either way
Whether the bot follows a script or writes its own answers, the business answers for it. In Moffatt v Air Canada, a tribunal rejected the airline’s argument that it was not responsible for what its chatbot said (opens in a new tab). Test every approach with real questions before launch, and after every change.
The bot designer asks how your bot should answer in its channels step, and its test chat lets you see how a plan holds up before you choose a platform.