Test your assistant
On the Tests tab of an assistant, the AI makes up customers who talk to your assistant, and checks every answer. You choose what to test and see in one sentence how it went, with a suggested fix for each problem. Nothing is sent to anyone: bookings, emails and other actions are only simulated, the same as in Try it out.
Open it under Assistants → Tests.
1. Choose what to test
Choose Suggest what to test. The AI reads what we already know about your business (the business profile, your website pages and the questions and answers in the assistant's knowledge collection) and what this assistant can do: book appointments, save requests, hand over to a person and its actions. It suggests six to ten topics, such as Book an appointment, Prices, Something we don't offer, Sales call or Unhappy customer.
- Tick the topics you want to test, and delete the ones you don't need.
- To add a topic yourself, type one line, for example "people asking for gift cards", and choose Add.
- Per topic you choose how many new tests the AI writes next, for example Write 3 more. The number next to it is how many tests the topic has now.
- Suggest again replaces the topics the AI came up with earlier that have no tests yet. Topics with tests, and topics you typed yourself, stay.
No website read yet? Then the AI goes by the assistant's instructions. Scan your website first under Knowledge → Sources for better topics.
2. Let the AI write the tests
Choose Create tests for … topics, with the number of topics you ticked. Each test is a short card: who calls, what they want, their mood and language, and what should happen. That last part can be how the conversation should end, what the assistant must or must not do (for example: must not book when someone only asks a price), and a fact it must mention, such as a price from your knowledge.
On each card you can:
- Edit the card. Your changes save automatically.
- Write again to have the AI write a different test in the same topic.
- Delete a test you don't need.
With more than six tests, each topic starts folded: open it to see its tests. Run all tests is at the top of this step too, and Run this topic sits next to each topic's name.
Want more of something? Under More options, ask for more tests in one line, such as "more about cancellations" or "make the callers impatient". The AI adds them to the topic it is about.
From a real conversation
Did a real conversation go wrong? Open it in the inbox and choose Turn into test. The AI adds it as a test of that assistant, in the closest topic, without names or numbers. Once you've changed something, run it again to see whether it's fixed.
3. Run and read the results
Choose Run all tests, or run one topic or one test. The page then moves to the results, where you follow the progress. The tests take a few minutes; you can leave the page in the meantime. While a run is going, the run buttons wait until it is done. Before a run starts, your latest change to the assistant is saved, so you always test what you see.
The results start with a plain sentence, such as "14 of 18 passed. 2 problems with "Prices".", followed by a pass rate per topic. Each problem opens with:
- the conversation as it went,
- which check failed,
- a one-line reason,
- a Suggested fix:
- When the assistant lacked a fact, choose Add this Q&A. The question goes to the proposals under Knowledge → Questions & answers, where you fill in the answer.
- When the assistant had what it needed but did the wrong thing, choose Fix with AI. You see the change to the instructions first, the same as Change with AI on the Instructions tab, and choose Apply change to save it.
Then choose Run again. Each topic shows ▲ or ▼ with the earlier pass rate, so you see straight away whether your change helped.
Tested by chat
The tests are typed, not spoken. They check what matters most: knowledge, tasks, tools, how the conversation ends and the tone. Speech-specific things, such as mishearing or interrupting, are not part of it: call the assistant yourself with Try it out for those. Transfers to a phone number work differently by chat: a test counts handing the chat over to your team as a transfer. A callback request is its own ending.
Flows in a test
The tests are chats, so the flows this assistant uses for chats run in them too: those that start with Call or chat starts or Chat starts. A test can check a flow like any other topic, for example "reports damage to their car". To try a flow for calls, use Call in Try it out.
Compare two assistants
Want to try a change without touching the live assistant? Choose Duplicate at the top of the assistant. The copy is not live and gets the same tests. Change the copy, then open the original and, under Compare with another assistant, choose the copy and Compare. The same tests run against the copy, and the results show ▲ or ▼ against the original's latest run, with the original's name and pass rate next to each topic.
Compare is part of Enterprise.
Who can do what
- Editors and admins suggest, write, change and run tests.
- Viewers see the results, but not the buttons. Viewers don't have Assistants in the menu, so send them the link to the Tests tab.
Limits
- Up to 12 topics and 60 tests per assistant, and up to 6 tests per topic in one go.
- A test stops after 12 messages from the customer.
- The AI can suggest or write tests 60 times a day per business. This is separate from writing instructions: suggesting and writing tests does not use up Change with AI. Fix with AI does, because it changes the instructions.
- Each day you can run 400 tests on a paid plan, 2,000 on Enterprise, and 40 without a subscription. Every test in a run counts: running all 30 tests uses 30, running one topic of 3 tests uses 3. Too few left for everything? Run one topic or one test.
- Only one run per assistant at a time.
- Results are kept for 90 days.
- Trying out costs AI time, so each business has an hourly budget: 300 replies in the test chat and tests together, 30 test calls in the browser, 120 previews of a voice or a pronunciation, and 300 searches by meaning under Knowledge. Past it you see "You've used the AI for this often enough this hour"; wait a while and try again.