What an AI agent is and when it works

An AI agent answers customers from what you have written. Where it works and where it does not is mostly decided before you configure anything.

An AI agent answers customers from what you have written. Where it works and where it does not is mostly decided before you configure anything.

What it does

It reads what the customer says, finds the answer in your published articles, and replies in its own words. Where you have set up actions, it can also look something up or change something in another system.

When it cannot help, it hands over to a person with the conversation so far attached.

Where it works well

• Questions with a documented answer. If an article answers it, the AI agent can.
• Lookups. Order status, appointment times, account details, given an action to fetch them.
• Volume and hours. Fifty identical questions at midnight are exactly the case.

Where it does not

• Anything undocumented. It has no source but your content.
• Judgement. Whether to make an exception, whether a complaint deserves a refund.
• Upset customers. Technically it can reply; in practice this is where handing over quickly matters more than answering.

The prerequisite nobody wants to hear

Your knowledge base is the product here. An AI agent on twelve thin articles will disappoint, and no amount of configuration changes that.

The work that makes an AI agent good is article work, done first.

Start narrow

One channel, a handful of subjects you are confident about, with a fast route to a human. Widen once the transcripts show it is handling those well.

Launching broadly is how an AI agent gets switched off in week two.

See also

• Creating your first AI agent
• How to write an article that AI can actually use
• Setting up escalation to a human