What an AI agent can (and can't) do for customer support

"AI agent" gets used loosely enough that it's worth being precise about what one actually does well. In customer support specifically, the useful version is narrower — and more useful — than the hype suggests.

Where an agent genuinely helps

Agents are strong at anything repetitive and well-defined: answering FAQ-style questions from your own documentation, qualifying an inbound lead against a fixed set of criteria before it reaches a salesperson, routing a WhatsApp or email enquiry to the right person, or reading an inbound document and filing the data it contains. These are tasks with a clear right answer, drawn from information you already have.

Where a human still needs to be involved

An agent shouldn't be the final word on anything genuinely ambiguous — a pricing exception, a complaint that needs judgment, a customer who's clearly upset and wants to be heard by a person. The useful design pattern is an agent that handles the routine 80% and hands off the remaining 20% cleanly, with full context, rather than one that tries to handle everything and quietly gets the hard cases wrong.

A simple way to decide if a process fits

If you can write the steps a competent new hire would follow — check the order status, look up the account, reply with the standard answer — an agent can very likely do it today. If the steps require weighing something that isn't written down anywhere, it's a job for a person, at least until there's enough real data to define the rule properly.