How to evaluate an AI voice agent

A good AI voice agent sounds natural, but that's only the beginning. What matters is whether it can handle real conversations, take the right actions, and hand off smoothly when needed. Test it on your own call types, define what "done" means before you listen to anything, and score what the agent actually did in your systems rather than how it sounded.

Choose one or two call types that are:

• Frequent enough to matter. A task that makes up a meaningful share of volume.
• Repeatable. The steps are mostly the same each time.
• Checkable. Afterwards, you can confirm in a system whether the right thing happened.

Delivery rescheduling, appointment booking and lead callbacks are good candidates. Disputes, cancellations with retention offers, or anything that needs an exception decision are poor first choices.

For each call type, write a short definition of a successful call: the outcome that must exist in a system afterwards, the steps that must happen along the way (disclosure, identity check, confirmation with the caller), and the situations where the correct outcome is a handoff. Then build a test set from real calls (redacted) or synthetic scripts modeled closely on them:

After each test call, check the systems, not the transcript alone. Ask each vendor for a record of every action the agent attempted (what it looked up, what it changed, and whether each action succeeded) and compare that record with what the caller agreed to.

• Completed correctly: the system shows the right outcome.
• Completed incorrectly: something changed, but not what the caller asked for. Treat this as worse than no change.
• Correct handoff: the call was out of scope and went to a person.
• Failure: the call ended without the task done and without a handoff.

Watch for agents that tell the caller something was done when the system shows it wasn't. That gap is invisible if you only read transcripts.

Handoff is where many callers either get helped or give up. Test it with the real destination:

• Does the transfer reach the right queue or number for that call type and time of day?
• Does the person who picks up receive a summary of who the caller is, what they need and what was already done?
• Does the caller have to repeat themselves?
• What happens out of hours: a callback task, a voicemail, or a dead end?

Check disclosure and consent as part of the same review. The agent should say early and plainly that the caller is speaking with an automated assistant. For outbound calls, the FCC has confirmed that AI-generated voices are "artificial" voices under the Telephone Consumer Protection Act, so calls using them require the called party's prior express consent where the TCPA applies (FCC Declaratory Ruling FCC 24-17). Ask how consent is recorded and checked before dialing, and review your campaigns with counsel. This guide is not legal advice.

Run the same scenarios against every agent you compare, and judge audio quality by its effect on callers: long silences after the caller stops talking, whether callers can interrupt, whether numbers and names are read back correctly, and how the agent responds to frustration. Have the people who would supervise these calls listen to a sample.

Before any live calls, agree the pilot gates in writing: the pass rate required to start, the share of calls the agent may handle at first, which calls are reviewed in QA and by whom, the errors that pause the pilot, and the outcomes that decide whether to expand.

Evaluation checklist

Use these checkpoints to assess an AI voice agent in a realistic test.

Evaluation criteria

Look for clear evidence in each area, not just a good-sounding conversation.

Use the same scorecard for AI calls that you use for people, with a few AI-specific criteria added. Our companion guide on a scorecard for human and AI calls covers how, and the AI voice agents page shows how TrackAgent approaches actions, handoff and review.