How to ship a customer-facing AI chatbot that won't embarrass you

Everyone has seen the screenshots: an airline's chatbot inventing a bereavement policy, a dealership's bot agreeing to sell a car for a dollar. These are not arguments against AI assistants — they are arguments against shipping one without guardrails.

Ground everything

The core discipline is retrieval-grounding: the assistant answers only from an approved content base — your help docs, policies, product data — and cites where each answer came from. If the content base does not contain the answer, the correct behavior is "I don't know, let me connect you with the team," not creative writing.

Constrain the job

An assistant that does everything guarantees surprises. Scope it: answer product and policy questions, capture and qualify leads, book meetings, hand off to humans. Refuse everything else politely. Narrow assistants feel smarter because they are consistently good within their lane.

Test like an adversary

Before launch, we run red-team scripts against every assistant: prompt injection attempts, requests for discounts and refunds, legal questions, provocations. The transcript review with the client is often the most valuable meeting of the project — it converts vague anxiety about AI into a concrete, fixable checklist.

Watch it in production

Launch is the midpoint. Review conversations weekly at first: where did the assistant decline when it knew the answer? Where was it confidently wrong? Feed both back into the content base and guardrails. Deflection rates typically climb for months after launch — but only for teams that keep looking.

Done this way, an assistant becomes what it should be: your fastest, most patient team member — one that never invents a refund policy at 2 a.m.