Stage 4 of 7 · Testing
Testing: we try to break it before customers can
We put your AI team through hundreds of real-world scenarios, including the awkward ones. It goes live only after it passes our gate and you have accepted the results.
What this stage is
This stage is user acceptance testing, or UAT. We build scenarios from your approved plan and your real past conversations: public holidays, 2am messages, language switching, photos, upset customers and people who change their mind.
A tester who did not build your AI signs it off. Builders never mark their own work.
Every failure is fixed and tested again. Then you see the results and three live scenarios, one of them your own worst case.
See it in action
Example chats and screens from this stage. Chats are illustrative, for a fictional business.
We test in a safe panel first
The respond.io test panel: we message the AI like a customer, with nothing reaching real chats.
An awkward scenario we test
Late-night messages, holidays, slang and two people at once: all in the test plan.
Your scorecard at the walkthrough
| Scenarios run | 240 |
| Passed first time | 212 |
| Fixed and tested again | 28 |
| Serious issues open | ✓ 0 |
| Minor quirks, fixed in hypercare | 2 |
Example figures. Yours come from your own test run.
Every reply can be traced
Run logs show why the AI answered the way it did, so a fix targets the cause.
What we do
- Write the test plan: at least 200 scenarios, built from the acceptance points in your plan.
- Replay real conversations from your history with personal details removed.
- Score every reply for the right answer, tone, hand-over and action.
- Keep a set of chats the AI has never seen, so a pass shows it learned rather than memorised.
- Check the safety switches: the off switch, limits and alerts.
- Message your real number like a customer would, after telling your team.
What we need from you
Each item has a due point, so you can plan your team's time.
Ticks are saved in this browser only, so you can work through the list over several days.
How you know it is done
Two gates, both required: no serious issues open, and the safety switches verified. Then your acceptance. Without all three, there is no launch.
What you will hear from us
- When testing starts
The dates, what we will test and the walkthrough week. - At the walkthrough
A scorecard in plain words: what was tested, what failed and got fixed, and any minor quirks. - When the stage moves
Your cutover plan and launch date.
Common questions
What counts as a serious issue?
Anything that gives a customer wrong information, makes a false booking or misses a hand-over a person should take. These must be at zero.
Will my customers see test messages?
No. Testing runs in a test workspace. The few live tests on your real number come from us, and your team is told first.
What is a "minor quirk"?
An issue that does not harm a customer, such as a slightly long reply. We list them openly and fix them during hypercare.
Can we test it ourselves?
Yes, and we encourage it. We will show you how to send test chats and how to report what you find.
Questions about your stage?
Message us in your project WhatsApp group, or start a chat here. Your account owner replies.