Test a bot before you point a channel at it
Fifteen minutes of deliberate testing prevents the transcript you have to apologise for. Here is the checklist worth running every time.
Almost everybody tests a bot by asking it a few things it should be good at, concluding it works, and shipping it. That tests the easiest 20% of what it will actually face.
The checklist below tests the parts that break. It takes about fifteen minutes and it should be run after every instruction change, not just before launch — because a change to the instruction changes the bot on every channel it serves.
Steps
Ask the five questions customers actually ask
Take them from real transcripts, not imagination. The gap between what you think people ask and what they ask is usually the whole problem with a bot.
Ask something it should refuse
A price you told it never to quote, a medical question, a delivery promise. Confirm it refuses in the words you wrote rather than improvising something softer.
Ask something it cannot possibly know
A made-up order number, a branch you do not have, a product you do not sell. This is where invented answers appear, and where you find out whether your "I do not know" sentence is working.
Trigger every handover condition, one at a time
Deliberately. Then confirm a real person received a real conversation on the other side — not just that the bot stopped.
Be rude to it
Customers are, when they are frustrated. Check it de-escalates and fetches a person rather than arguing or apologising in a loop.
Test in every language you enabled
Including switching mid-conversation, and including how it pronounces your brand name and your prices.
For a call bot, use the web widget test card
The web widget is the one chat type that also gets a start-call test card, which lets you exercise a voice agent end to end without dialling anything. It is the fastest feedback loop available.
Check it worked
Write down anything the bot said that you would not want a customer to see. Each item is either an instruction change, a knowledge base gap or a missing handover condition — and each is much cheaper to fix now.
The parts people get wrong
- Read the Used-by list before changing a live bot. You are changing it everywhere at once.
- Test after every instruction change, not just at launch. Small edits have surprisingly large effects on behaviour.