Modern AI chatbots use language models and retrieval to answer questions in natural language from a business's own knowledge: policies, product documentation, order data. Done well they remove friction from support and sales. Done badly they invent answers and damage trust. The difference is almost entirely in scoping and grounding.
Rule-based bots vs LLM chatbots
Older chatbots followed decision trees. They were predictable but frustrating. LLM chatbots understand free-form questions and can respond naturally, but must be grounded — given the relevant documents at answer time — to stay accurate.
What a trustworthy business chatbot needs
- A defined scope and a graceful "I can't help with that, here's a human".
- Retrieval over curated, up-to-date content with citations.
- Access to account or order data only through authenticated, permissioned tools.
- Conversation logs and a review process to fix gaps in content.
- Clear disclosure that the user is talking to an AI assistant.
Where chatbots create value
Pre-sales questions on websites, first-line support for known issues, internal HR and IT help desks, onboarding guidance and appointment or order status queries.
My perspective
A chatbot is a content project as much as an AI project. If your documentation is thin or outdated, fix that first — the chatbot will only expose it faster.
Business use cases
- Website assistants answering pre-sales questions and capturing leads
- First-line customer support for known issues with escalation
- Internal HR, IT and policy help desks
- Order status and appointment queries through authenticated tools
- Guided onboarding and product education
Limitations to be honest about
- Without grounding, chatbots hallucinate; with poor content they surface gaps.
- Sensitive actions (refunds, account changes) should not be delegated to a chatbot without verification.
- Users quickly abandon bots that cannot hand off to a human.
- Measuring resolution rate honestly requires effort.
How I approach it
- I start by auditing the questions people actually ask (support tickets, sales emails) and the content that answers them. The bot is scoped to what the content supports, with retrieval, citations, human handoff and weekly review of unanswered questions.
Frequently asked questions
Are AI chatbots accurate enough for customer support?
For known, documented questions — yes, when grounded in your content with citations and a handoff path. For novel or sensitive issues, a human should remain in the loop.
How is an AI chatbot different from an AI agent?
A chatbot answers questions in conversation. An agent pursues tasks across steps using tools. Many products combine both.
What content do I need before building a chatbot?
Accurate FAQs, product documentation, policies and any structured data (orders, bookings) exposed via secure APIs.
Last updated 11 September 2026