Section 1
What a support bot is genuinely good at
Three categories, and they cover more volume than most teams expect. Questions with a documented answer: hours, policies, how to change a plan, where an order is. Status lookups, where the bot fetches a fact from a system the customer cannot reach. Triage, where the bot collects the details a human would otherwise spend two messages asking for, then routes the conversation to the right person with the context attached. What it is not good at is anything requiring a judgment call about an unhappy customer, an exception to policy, or a promise about the future. On the wider personalisation question, see [How AI Personalizes the Customer Journey](/blog/how-ai-personalizes-the-customer-journey).
Section 2
The handover is the product
Every complaint about chatbots is really a complaint about a handover. The loop that never escalates. The transfer that loses the conversation so the customer retypes everything. The hidden human option. Design the exit first, before the answers. The bot should escalate on an explicit request, on repeated failure to resolve, and on any signal of frustration or a high-value account. It should carry the transcript with it, so the human starts informed. And the route to a person should be visible from the first message rather than earned after four attempts. A bot that resolves half of contacts and hands over the rest cleanly will outperform one that resolves seventy percent and traps the remainder.
Section 3
Building it without irritating your customers
Start from your own history rather than a template. Pull the last five hundred conversations, group them by intent, and count. The top ten intents usually account for most of the volume, and that is your build scope. Write the answers as answers, not as links to a help article. A customer who wanted to read the documentation would have read it. Then run in shadow mode for a week: the bot drafts, a human sends. You will see exactly where it is confidently wrong before a customer does. Related campaign mechanics are in [Building Customer Loyalty with Relatable Stories](/blog/building-customer-loyalty-with-relatable-stories).
Section 4
What it must never decide alone
NIST organises AI risk around trustworthiness, design, evaluation and use. In a support context the practical version is a short list of things the bot may say and a shorter list of things only a person may. Refunds, credits, cancellation exceptions, anything about legal rights, anything about safety, and anything that commits you to a date. Those need a human, every time, regardless of how confident the system sounds. Two more rules. Disclose that it is a machine, plainly and early, because customers work it out anyway and discovering it late reads as deception. And do not push customer records into a prompt that has no use for them: answering a question about delivery times does not require a payment history.
Section 5
Resolution, not deflection
Deflection rate is the metric vendors report and the one most likely to mislead. A conversation the customer abandoned in frustration counts as deflected. So does one that ended with a wrong answer. Measure resolution instead: the share of conversations that ended without a follow-up contact on the same issue within a week. Add satisfaction measured separately for bot-resolved and escalated conversations, since a low score on escalations usually means the handover is broken rather than the answers. Also track how often the bot answered something it should have escalated. That number should be zero, and if nobody is sampling transcripts, nobody knows what it is.
Section 6
Where this stops being worth it
Volume decides. Below roughly a few dozen repetitive contacts a day, a well-written help page and a good email template will outperform a bot for a fraction of the effort, and the bot becomes a maintenance obligation with no payoff. It is also the wrong tool where every contact is unusual, where conversations are long and consultative, or where the relationship is the product. Complex professional services fit that description more often than their vendors admit. Review the intent list quarterly. Products change, and a bot answering last year's questions accurately is still giving customers the wrong information.