TL;DR
- Write yes/no criteria a judge can check. Critical errors fail the call.
- For AI voice agent testing, use an LLM-as-a-judge at temperature 0. Calibrate it against humans.
- Set thresholds per criterion, from measured agreement.
Write a rubric a judge can apply
Start by splitting your criteria into two piles. Hard gates are critical errors that fail the whole call. Weighted criteria are quality items that move a score. QA leaders say a checklist score can hide what matters. A call can be perfect on procedure and still leave the customer angry, and one critical error should fail the call.
Hard gates for a voice agent:
- The agent says it is AI at the start. Article 50 of the EU AI Act has applied since 2 Aug 2026 to any AI that talks to people, phone systems included.
- The agent gives no wrong policy detail.
- The agent hands the call to a human when asked. Gartner found that 87% of 3,566 customers expect a route to a human, so this is a gate.
Weighted criteria: tone, follow-up questions, and resolution.
Write every criterion in the same shape:
- A yes/no question, such as "Did the agent state it was an AI in its first turn?"
- The observable evidence the judge must quote.
- One pass example and one fail example.
- An N/A rule, so the judge never guesses.
Add voice items that a clean transcript can hide: interruptions the agent talked over, long silences, slow replies, and answers in the wrong language. The judge cannot see timing in plain text, so mark these "needs timing data" and pass them in as metadata (the next step covers this).
Allow for wording and order. If a required topic is covered later, or in different words, it counts as covered. A telehealth QA team saw a required topic scored as missed when it was covered later. Put "anywhere in the call, in any wording" into the criterion text.
Configure the judge to score consistently
Set the judge up as a strict, repeatable instrument. Send it one criterion per call, so a long rubric does not blur its attention. Ask for the reasoning first and the verdict second. Return structured JSON with the fields evidence_quote, reasoning, verdict (pass, fail or N/A). Set temperature to 0 and pin the exact model version. A silent model update will otherwise move your scores.
Give the judge two inputs. The first is a transcript with speaker labels. If diarization puts the agent's words under the caller, every criterion about who said what will break, so check label quality before you tune any prompt. The second is timing metadata: turn start and end times, gaps in seconds, overlap flags and language tags. The judge cannot infer silence or latency from text.
Know the judge's biases. The MT-Bench and Chatbot Arena paper by Zheng et al. names position bias, verbosity bias, self-enhancement bias and limited reasoning ability. It also says some biases are minor or can be mitigated. Your mitigations:
- Use a judge from a different model family than the agent under test.
- Add anchored pass and fail examples to each criterion, so long answers do not look better.
- Score each criterion alone, so there is no ordering to favour.
You do not need a blank page. Insight7 has starter rubrics on the Criteria screen. For campaign agents, which are in beta, you can talk to the agent in the browser with no phone number. Campaigns in beta can also place a test call to your own phone. Read each transcript when the call ends. Feed those calls to your judge as your first test set. Alerts can watch keywords and low scores and send a weekly digest of the weakest calls by email or Slack.
Calibrate the judge against human review
Build a golden set of 50 to 100 calls before you trust any judge score. Stratify it on purpose. Include clear passes, clear fails, edge cases and several examples of each critical error type. A random sample of good calls will hide the failures you care about most.
Have two reviewers label every call blind, each working alone with the same rubric. Measure how often they agree before you look at the judge, because human agreement is the ceiling the judge can reach. If your reviewers split on a criterion, the wording is unclear, so fix it first. Then let them resolve each disagreement together. The resolved labels become your gold labels.
Next, compare the judge with the gold labels, one criterion at a time.
- Cohen's kappa measures agreement after removing the agreement expected by chance. A score of 0 is chance and 1 is perfect. Raw percent agreement misleads when 95% of calls pass, because a judge that always says "pass" looks accurate.
- A confusion matrix per criterion shows the two error types. A false pass lets a bad call through. A false fail wastes reviewer time. Track them separately, because for a hard gate a false pass costs far more.
Then run the fix loop:
- Read every disagreement, not a summary.
- Tighten the criterion wording, or add an anchored pass or fail example.
- Rerun the judge on the same set.
- Repeat until kappa stops improving.
Freeze the golden set once you start, so scores stay comparable between runs. Add new calls to a separate set. Recalibrate whenever the judge model, the agent script or the rubric changes.
Our view is that an automated score must match the human auditor, or it is not ready. Calibration finds that gap before launch, not after.
Set pass/fail thresholds per criterion
Write one threshold rule for each criterion type before the first test run. Let your calibration kappa decide what each criterion is allowed to do.
| Criterion type | Threshold rule |
|---|---|
| Critical gate | Zero fails across all trials. Kappa is high and false passes are near zero. |
| Compliance | Zero fails. A human reads every judge fail. |
| Quality (tone, follow-up, resolution) | Pass rate above a target you set, such as 90%. |
| Style | Tracked as a trend. Never a gate. |
If a criterion's kappa is low, do not use it as a gate. Send it to human review until you fix the wording.
Check that your test set is big enough. A pass rate from 100 calls is a range, not a point. A Wilson interval gives that range and stays reliable on small samples. Zero fails in 100 calls still leaves a real chance of failures at the interval's upper bound. Agents are non-deterministic, so run each scenario several times and score every trial.
Your go-live rule is short. Every gate passes, every quality criterion clears its target, and the thresholds were fixed before testing.
Keep scoring after launch
Schedule a monthly human spot-check. Review a random sample of live calls and every call the judge failed. Track agreement per criterion each month, because a falling kappa means the judge is drifting. Gartner analysed 432 AI customer-service use cases and found that 42% could not be measured.
Related reading: AI Agent Regression Testing Checklist for Every Script or Prompt Change, AI Voice Agent Failure Modes Explained With 10 Examples and How to Prevent Them and Simulated Caller Testing vs Live Pilot Testing for AI Voice Agents Compared.
Questions people also ask
What is LLM-as-a-judge?
It is a method where one language model grades another system's output against a written rubric. For voice agents, the judge reads a call transcript plus timing data. It returns a pass, fail or N/A verdict with reasoning.
How reliable is LLM-as-a-judge?
It depends on the criterion. Clear, observable yes/no items often reach high agreement. Vague ones like "sounded empathetic" drift. Measure kappa per criterion against human labels instead of trusting one headline figure.
How do you calibrate an LLM judge against human reviewers?
Have two humans label 50 to 100 calls blind. Resolve their splits into gold labels. Compare judge verdicts per criterion with kappa and a confusion matrix. Fix the wording where they disagree, rerun, and freeze the set.
What is a good pass rate for an AI voice agent?
No universal number exists. Set zero fails on critical gates and a target such as 90% on quality items. Fix the numbers before testing, and judge results by the interval's lower bound.
How many test calls do you need to evaluate a voice agent?
Plan for at least 50 to 100 scored calls. Rare critical errors need targeted scenarios, each run several times. Add cases for other languages, accents and noisy lines, because volume alone will not surface them.
Which model should I use as the judge?
Choose a strong model from a different family than your agent, to limit self-preference. Pin its version. Run two or three candidates on your golden set and keep the one with the highest kappa on your gates.
How do you test voice AI agents before launch?
Run scripted scenarios: normal paths, interruptions, angry callers, wrong-language replies, off-script questions and requests for a human. Score every transcript with the judge. Then have people review all fails before go-live.
Calibrate first, then gate
Insight7 scores every call against a weighted scorecard your team owns. Rule of thumb: no criterion becomes a gate until the judge agrees with your humans on it.

