You changed one line in the agent's script so it would stop over-explaining the EMI options. Now it answers that well. Did the refund answer get worse? Without a repeatable check, the first person to find out is a customer.
AI Call Studio has two ways to check an agent before anyone real hears it, and neither dials a number. People confuse them, so here is the short version.
Agent tests ask set questions
Agent tests are saved cases in a suite, kept in the agent's Evals tab. You type what the caller says, and how to check the reply. There are three checks: it must contain something, it must not contain something, or it is judged against what you describe, for answers that can be worded many ways.
A suite holds up to 50 cases and 120 caller messages a run. It runs only when you press Run, and every case shows pass or fail with the reply it got. A team's suite runs through its coordinator, so routing and handovers are tested too.
Use this for the things that must stay true. The fee for a course. The fact that it never promises approval. The way it answers “how did you get my number?” If any of these has gone wrong before, it belongs here.
Turn a bad call into a test
The most useful cases come from real calls. On a call in Call Logs, press Add to tests. The caller's words play again in order while the agent answers live, and the check reads the whole replayed conversation. When a call goes wrong on Tuesday, you fix it and add it, and that mistake now has a guard on it. A bad call that does not become a check can come back next month.
The Simulation Lab holds a conversation
The Simulation Lab plays whole text conversations between your agent or team and AI customers who stall, argue and wander off topic. Pick from a library by industry, such as angry, stalling or off topic, or write your own: a scenario, how they behave and what they want.
Each conversation is graded on your QA scorecard. Pass rate by run shows whether your change helped, and the latest run names each customer that scored lower than before. A team rehearsal runs the greeter, the routing and each specialist.
Use this when you want to know how the agent holds up over several turns. People who built an agent test it politely. Real callers interrupt, refuse, and ask about something else. Write customers based on the people you actually meet: the borrower who says the money has already gone, or the parent who wants the principal's number.
What neither of them tests
Both are text. Neither uses the agent's knowledge base or tools, so a question that needs a lookup can fail here and pass on a call. Neither tests the voice, interruptions or how numbers sound when spoken. A pass in the Lab is the AI's judgement of an AI conversation. That is a useful signal and nothing stronger.
So there is always a last step: a real test call in the Live Call Console, in your browser, using the voice and knowledge a customer would get. It is billed at the web-call rate until you press End. Listen for names, prices and long numbers.
An order that works
- After any script change, run the agent tests. They are quick and they catch what you broke.
- After a bigger change, such as a new offer or a new opening, run the Lab too.
- Fix what failed. Treat a fail as a real problem, and a pass as ready to try on a test call.
- Make a test call yourself, with the difficult questions as well as the easy ones.
- Go live on a small list and listen to a few calls.
If you have Agent Coach, it turns failed rehearsals and graded calls into suggested changes. You apply the ones you agree with, and a version is saved first, so you can go back.
Money and plans
Agent tests are on every plan. They come at no charge on the free trial, with a daily limit on runs. Otherwise each run of a suite carries its own price, and a case counts for nothing on its own. The Lab is not on the free trial, and each conversation with an AI customer is charged, except for a run that did not finish, which says why on its row. Current figures are on the pricing page.
The guide to making your calls better shows how these fit with QA. If your concern is people and not agents, the same customers can ring your telecallers: see training on practice calls.