The Chat Assistant
Framework
Most website chatbots make things worse.
A chatbot that invents answers erodes trust the moment it's caught. One that only knows a few canned replies sends everyone to "contact us" anyway. And the common "fix" — route every visitor straight to WhatsApp — just moves the problem: now a human answers the same basic questions all day, and the genuinely valuable leads are buried in the noise.
We built assistants for our own sites and brands and refused all three failure modes. This is the framework we use.
Answer from what you control. Qualify. Then hand off.
Greet → answer → qualify → hand off.
| 1 | Greet | A clear opening that sets what the assistant can help with. |
| 2 | Answer | Respond from the knowledge base, and cite the page or product link. If it's not known, say so. |
| 3 | Detect intent | Buyer, browser, support, or partner? The reply adapts to who's asking. |
| 4 | Qualify | For commercial intent, ask up to three short questions — what they want, who they are, how to reach them. |
| 5 | Hand off | Capture the lead and context, notify a human, and tell the visitor what happens next. |
Two brains: behaviour and facts.
| Layer | What it holds | How often it changes |
|---|---|---|
| Config (behaviour) | Role, tone, what it can and can't do, when to escalate, how to hand off. | Set once, rarely. |
| Knowledge base (facts) | Products, prices, policies, FAQs, guides. | Whenever the facts change. |
Behaviour lives in the config; facts live in the knowledge base. Change a price? Edit the KB — you never touch how the assistant behaves. Change the tone? Edit the config — the facts are untouched.
Two models, one auto-switch.
Most questions are simple and should run on a fast, cheap model. The hard ones — a tricky comparison, a long reasoning chain — should escalate to a stronger model automatically, on conditions you define. You get speed and cost control on the 90%, and quality on the 10% that needs it.
The guardrails.
Warm leads, with the context already gathered.
When a conversation qualifies, the assistant collects the essentials — name, best contact, and a one-line summary — and notifies your team. The human picks up already knowing who they are and what they want. No "how can I help you?" from scratch.
Keep customer data where it belongs.
Conversations and leads contain personal data. Handle it under your privacy obligations — in Malaysia, the PDPA. Keep only what you genuinely need (leads you're working, and de-identified data to improve the assistant), and never scatter customer data into places it doesn't belong, like a code repository.
Where the assistant fits in a launch.
The chat assistant is the last step of launching a brand — it only works once the website, shop and knowledge base exist to feed it. The companion playbook covers that whole order.