Testing an AI agent before you publish it

Testing means trying to break it, not confirming it works. Four ways to do that before customers do.

Testing means trying to break it, not confirming it works. Four ways to do that before customers do.

Test it

1. Open the AI agent and start a test conversation.
2. Work through your list of questions.
3. Note every answer that is wrong, vague, or arrives after too much back and forth.
4. Fix the cause, usually in an article, and test again.

One: ask badly

Short, misspelled, without context. "order late", "cant login", "wheres my stuff". That is what real messages look like, and an agent that only handles well-formed questions handles almost nothing.

Two: ask what it should refuse

Something you have not documented. Something outside your business entirely. It should say it cannot help and offer a person, not improvise.

An agent that answers everything is an agent that will eventually answer wrongly with confidence.

Three: change subject mid-conversation

Ask about an order, then about returns, then back. Real conversations wander, and this is where an agent either keeps up or loops.

Four: ask for a human

Directly, and also indirectly by getting frustrated. Handover must work from any point, and it should not require the customer to phrase it a particular way.

Get somebody else to try

You know what the agent can do, so you unconsciously ask in ways that work. A colleague from another team will find in ten minutes what you would not find in an hour.

Keep the list

The questions you test with become your regression set. Run them again after every substantive change, because a fix in one place moves behaviour in another.

See also

• Publishing an AI agent
• Why my AI agent answers incorrectly
• Setting up escalation to a human