How to test a voice agent before it takes real calls

A voice agent that sounds great in a demo can still stumble on a real caller: someone who interrupts, gives details out of order, asks for something it doesn't know, or just wants a person. The fix is the same as for any new hire: practice the hard calls before the first shift.

This is the checklist we'd use to test any AI phone agent before it answers a real number, with how each step works in Apollo.

What you're testing for

Every test call is checking one of four things. Apollo's readiness tests use the same groups:

• Core tasks: The jobs callers ring for.
• Knowledge: Answers from your info, nothing invented.
• Guardrails: Callers pushing it off its rules.
• Escalations: Emergencies, urgency, and handoffs.

If you only test the first group, you'll ship an agent that works when callers behave. The other three are where real calls go wrong.

Step 1: Call it yourself in the browser

Start with your own voice. In Apollo you can call any agent from the dashboard in your browser, with no phone number needed and no minutes used. Make a short list of the calls you get most, and make each one:

• The easy version: a calm caller with everything ready.
• The messy version: you interrupt, change your mind halfway, give the phone number before your name.
• The question it can't answer: something not in its information. It should say it doesn't know and offer a person or a message, not guess.
• The handoff: "Can I talk to someone?" It should offer the transfer.

Browser test calls are capped per hour and per day, and on the Free plan each one runs up to 90 seconds, so plan a few focused calls rather than one long one. On paid plans a call can run up to 30 minutes.

Listen for these

• Does it read numbers, times and prices back clearly?
• Does it ask one thing at a time, or pile on questions?
• Does it confirm details before it acts, like a booking's date and party size?
• Does it end the call politely once you're done?

Step 2: Run the readiness tests

You can't play every caller yourself. Readiness tests have an AI caller ring your agent with a goal and a pass condition, and grade the conversation. Every agent gets the universal set (asking what it can do, leaving a message, insisting on a person, going off topic and probing for its instructions), and each industry starter adds its own: a clinic agent is tested on requests for medical advice and on emergency symptoms, a home-services agent on a caller who smells gas, a real-estate agent on fair-housing probes.

Each scenario runs 3 times, and only passes if every run does. That's deliberate: a voice agent that handles a case two times out of three will fail a real caller sooner or later.

Vary the caller, not just the goal

The same request sounds very different from different people. Scenarios can use different caller styles (calm caller, interrupter, hesitant, non-native speaker, upset caller, elderly caller), which tests turn-taking and comprehension, not only the happy path.

Write your own scenarios

The built-in tests can't know your business. Add your own (up to 20 per agent): a caller with a goal, and what passing means. Apollo can also suggest tests for your agent. Good ones to write:

• The question you get most, phrased the way customers actually phrase it.
• A request just outside your rules: a time after closing, an area you don't serve, a party bigger than your largest table.
• A caller who wants a price you don't publish.
• The scenario you're most nervous about.

Readiness tests are simulated conversations with the agent's tools mocked, so they don't book real tables or create real records, and they don't count toward your minutes.

Step 3: Let other people try it with a talk link

You know what the agent is supposed to say, which makes you a bad tester. Share a talk link: a page where anyone with the link can talk to your agent in their browser, without signing in. Send it to your staff, a friend who's never called you, or a regular who'll tell you the truth.

Talk links are real conversations, so the agent's actions run for real, and each conversation is logged like a call, with its recording and transcript. On a paid plan, their talk time counts toward your minutes like a call does. You can turn a link off or reset it at any time, and conversations are limited per hour and per day.

Step 4: Read the transcripts

Listening back is slow; reading is fast. Every call in the log has its transcript and a short summary, including whether the call reached its goal and why it ended. Look for:

• Places the agent said it didn't know something. Each one is a missing FAQ.
• Questions it asked twice, or in an odd order.
• Handoffs that could have been avoided with one more fact, and ones that came too late.
• Details captured in the Inbox that were wrong or incomplete.

Fix what you find by chatting with Apollo or editing the prompt, then run the readiness tests again. Every save is live on the next call.

Step 5: Go live gradually

When the tests pass and the transcripts look right, put it on a number, but don't hand it every call on day one. Forward only when your line is busy or nobody answers for the first week, so your team still picks up when they can. Here's how to set that up.

The checklist

1. 1Make your most common calls yourself in the browser: easy, messy, unknown and handoff.
2. 2Run the readiness tests until every scenario passes all its runs.
3. 3Add your own scenarios for your riskiest calls.
4. 4Share a talk link with people who don't know the script.
5. 5Read the transcripts and fix the gaps.
6. 6Go live on busy and no-answer forwarding first.

All of this is free until the agent answers a real number. Start building free, see vibe coding a voice agent for how agents get built, or check the FAQ.