How to evaluate voice agents
Evaluate a voice agent on three questions: did it follow its instructions, did the caller get what they came for, and did the call work well for the caller. Use deterministic checks for explicit rules and an LLM judge for judgment calls, then run the same checks on real calls.
By Aishwarya Birla, Arun Prasad · Published · Last updated
On this page
Why voice agent evaluation matters
A voice agent can finish every step, with every tool returning success, and still leave the caller without what they needed. AI agents don't fail like normal software. A tool can return a 200 and the task can still fail. Standard monitoring shows the request went through, not whether the outcome was right.
Voice adds problems that a text transcript hides. Voice agents depend on speech recognition, real-time response and conversational flow, so a caller can hit a slow or awkward exchange even when the words on the page look fine. Evaluation is how you find these failures before callers do.
Step 1: Decide what to measure
Measure in three groups: execution (did the agent follow its instructions), outcome (did the caller reach their goal) and experience (responsiveness, clarity and friction).
Speech recognition, latency and real-time performance, and conversation flow as what sets voice agents apart. Intent classification and entity extraction, response quality and task completion are also worth tracking; that part is this page's own guidance. A rough mapping:
- Execution: intent and entity extraction, and whether required steps and rules were followed.
- Outcome: task completion and whether the caller's goal was met.
- Experience: latency, response quality, conversation flow and speech recognition accuracy (a transcription error the caller has to repeat shows up as friction).
Pick a small set per group to start. You can add more once you see where calls fail.
Step 2: Choose deterministic checks or LLM judges
Use deterministic checks for explicit requirements and an LLM judge for semantic questions.
A deterministic check gives the same verdict for the same input and uses no model. Good uses:
- A required disclosure was read.
- A tool was called with the right parameter.
- A forbidden phrase was not spoken.
An LLM judge is a model that reads the exchange. Good uses:
- Was the tone appropriate?
- Was the reply clear?
- Did the caller get what they came for?
Match the evaluator to the signal. If a simple check can enforce the rule, use the check. It is cheaper, repeatable and easier to explain when it fails. Keep the judge for questions where reasonable people could disagree.
Step 3: Measure latency
Record timing for each turn, from when the caller stops speaking to when the agent starts responding, and look at the spread, not only the average. A few very slow turns can ruin calls even when the average looks fine.
Also break the time down by component: speech recognition, the model, tool calls and speech output. That shows where the time goes. Latency and real-time performance are evaluation areas and do not ignore them until production.
This guide doesn't give a target number. We didn't find a threshold we can back with a source, and the right limit depends on your callers and use case.
Step 4: Test whole conversations, not only single replies
Score both single messages and the full call. This split is evaluating single messages versus the conversation level. A reply can be fine alone and still sit inside a call that loops, repeats a question or ends without the goal met.
Build test calls that cover more than the happy path:
- Callers who give everything you need, in order.
- Callers who interrupt or change their mind.
- Callers who leave out details.
- Questions outside what the agent should answer.
Evaluate components in isolation only as a common pitfall. Test the pieces and the end-to-end call.
Step 5: Keep scoring real calls in production
Run the same checks on live calls, review the failures, and turn repeated failures into new test cases. Test calls alone won't show how real callers behave.
Whatever tool you use, you need three things: the transcript and timing of each call, the scores for each check, and a way to see which failures repeat. A failure that appears once is an anecdote. One that appears across many calls is a fix to prioritize.
How this looks in production
PRISM by Block Convey connects to ElevenLabs voice agents in one of two ways: a post-call webhook, or the live PRISMtraceVoiceTracer. Enable only one. Audio is never sent to or stored by PRISM, and transcripts are PII-scrubbed server-side. The PRISM docs are the place to check setup details.
PRISM's Agent Observability has four connected views: Traces (request, response, model, cost, timing and scores for each run), Sessions (the multi-turn journey and drift), Agent Runs (steps, tool calls, retries, handoffs and state) and Scores & Alerts. It tracks the outcome against the user's goal, not just span status. It also groups repeated failures into patterns and shows how many sessions share each one. Agent steps appear when your integration sends them.
How the two kinds of evaluator run:
- Deterministic rules run on 100% of sessions, with no sampling.
- The LLM judge is not sampled. It scores every incoming trace until the plan's allowance runs out (Free: 100 a day, resetting at 00:00 UTC). After that, only the deterministic rules run.
- Custom evaluators are on Builder and above. Core scores are on Free.
Limits to know about:
- ElevenLabs is the only voice connector PRISM lists today.
- Because audio never reaches PRISM, it works from transcripts. We don't claim PRISM measures audio quality, so you'll need another way to judge how the voice sounds.
- Every supported connector is available on every plan, including Free.
If you use ElevenLabs, you can connect your voice agent to PRISM with a post-call webhook or the live tracer.
Common mistakes
Most voice evaluation mistakes come from testing too narrowly or too late.
- Testing only happy paths. Fix: add interruptions, missing details and out-of-scope questions to your test calls.
- Ignoring latency until production. Fix: record timing per turn from the first test call, and look at slow turns, not only the average.
- Judging components in isolation only. Fix: also score the full conversation, because each piece can pass while the call fails.
- Using an LLM judge for rules a simple check can enforce. Fix: write a deterministic check for explicit rules. It gives the same verdict for the same input and, on PRISM, runs on 100% of sessions with no sampling.
Connect your ElevenLabs agent without storing audio
Use a post-call webhook or the live tracer. Audio is never sent to or stored by PRISM.
Free to start, 100 credits each month.