An AI agent in your helpdesk can answer most of what arrives. The question is which messages it should answer, which it should draft, and which it should never touch.

Answer outright

Questions with a factual answer that exists in your own material: delivery times, how to reset a password, what the returns window is, where an order is. The agent reads the order record and the help centre, answers in your voice and closes the ticket. We measure these at about 70 % of volume for online retailers.

Draft for a person

Anything that involves a decision: a goodwill refund, a change to a contract, a complaint about a person. The agent prepares the reply and the context; a human reads it and presses send. The reading takes 20 seconds. The writing used to take five minutes.

Never touch

• Legal threats and anything mentioning a regulator
• Safety issues
• A customer who has written three times about the same thing

These go to a named person within minutes, with everything attached.

Give it the facts, not the internet

An agent that answers from your help articles and your order data is accurate. One that answers from general knowledge is confident and wrong. The difference is the setup, not the model.

Every answer the agent sends is logged with the source it used. If a customer disputes it, you can see exactly why it said what it said.

Measure three things

Tickets closed without a person, the reopen rate on those tickets, and the satisfaction score on them compared with human replies. If the second number climbs, narrow what the agent is allowed to answer.