How to evaluate a voice agent before it takes real calls

Testing a voice agent by calling it a few times is how most teams start, and it is how most problems slip through. A handful of friendly test calls cannot show how the agent handles an impatient caller on a noisy line who changes their mind halfway through a booking. Evaluation is what turns a voice agent from a promising demo into something you can put in front of customers.

This article sets out how we evaluate voice agents before launch, and keep evaluating them after. The principles are the same as for any AI system, covered in Evaluating grounded answers in RAG systems, but voice adds two dimensions text does not have: audio and time.

What to measure

A voice agent can fail in several independent ways, so it needs several measures rather than one score.

• Task completion: did the caller get what they called for, such as a booking made, a question answered, or a call routed correctly?
• Correctness: were the details right, including dates, times, names, numbers, and the records written to other systems?
• Tool use: did the agent call the right tools with the right inputs, and avoid actions it should not take?
• Escalation: did it hand off when it should, and only when it should?
• Latency: how long callers waited after speaking, including the slow tail, not only the average.
• Conversation quality: were responses short, clear, and natural to hear, without repetition or talking over the caller?

Build a scenario set from real calls

The evaluation set is a collection of call scenarios, each with a caller goal, the details the caller knows, and what a correct outcome looks like. The best source is real calls: recordings or notes from the people currently answering the phone, covering common requests, awkward edge cases, and calls the agent should decline or escalate.

Include the difficult cases on purpose: callers who give information out of order, correct themselves, ask two things at once, go quiet, have strong accents, or call from somewhere noisy. Keep the set versioned, so every change to the agent is measured against the same scenarios and results can be compared over time.

Simulate callers at scale

Running hundreds of scenarios by hand is impractical, so we simulate them. A second language model plays the caller, following the scenario's goal and personality, and holds a full conversation with the agent. Text-based simulation is fast and cheap, and it catches most reasoning, tool-use, and escalation problems.

It does not catch audio problems. For those, a smaller set of scenarios runs end to end through speech: synthesised caller audio, with background noise and phone-line quality, passed through the real recognition and synthesis pipeline. This is where misheard names, mangled numbers, and turn-taking problems show up. It is also the place to compare text-to-speech voices: we run candidate voices from providers such as ElevenLabs and Fish Audio through the same scenarios over a phone line before choosing one.

Test latency as callers experience it

Latency has to be measured over a real phone connection, from the end of the caller's speech to the first sound of the response, with timings for each stage of the pipeline. Run the measurements at realistic call volumes, because latency under load can be very different from latency on a single test call. The architecture choices behind good numbers are covered in Designing voice agents for latency people can live with.

Roll out gradually

Passing the evaluation set earns a limited launch, not a full one. Common patterns are starting with a single call type, a share of calls, or out-of-hours calls only, with transcripts reviewed daily. Every failure found in production becomes a new scenario in the evaluation set, so the same mistake is caught automatically next time.

Practical recommendations

1. 01Measure task completion, correctness, tool use, escalation, latency, and conversation quality separately.
2. 02Build scenarios from real calls, including difficult callers and calls that should escalate.
3. 03Simulate callers in text for scale, and in audio for the problems only audio reveals.
4. 04Verify model-graded scores with human review.
5. 05Measure latency over real phone lines and under realistic load.
6. 06Launch on a narrow slice of calls, and turn every production failure into a test.

Evaluation is part of our voice AI work, and the core of our production AI work.