Founders
How to build an evaluation dataset for AI voice agents
AI voice agents: Build an evaluation dataset around the decisions the product must support.
Build an evaluation dataset around the decisions the product must support. Include ordinary cases, difficult cases, and known failure modes with documented provenance. Keep development examples separate from the examples used to judge a change.
Start with the decision
Define the unit being evaluated and the expected result before selecting examples. Record who labeled them, what disagreement means, and how the sample relates to actual use. Revisit coverage when the product or customer workflow changes.
The illustrative scenario
This is an educational workbook. The company or fund and all numerical inputs below are fictional. They are not ADIN results, customer claims, market benchmarks, or a forecast.
The fictional product handles appointment requests for a multi-location service business. A caller may change their request mid-conversation, and the scheduling system is the final record of an available slot.
A short call is not useful if it books the wrong appointment. Separate successful scheduling from hangups, transfers, and calls where the caller never confirmed the details.
An evaluation sample with a denominator
Imagine a permissioned evaluation set with 29 cases and 20 that meet a rubric defined before the test. The resulting percentage describes only this invented sample. Seventy percent is not a recommended target, a safety standard, or evidence about any real system. What matters is the rubric, sample coverage, and consequences of the remaining cases.
A short call is not useful if it books the wrong appointment. Separate successful scheduling from hangups, transfers, and calls where the caller never confirmed the details. Break the 9 remaining cases into failure categories before interpreting the aggregate. A recoverable wording issue and an unauthorized action may require different product behavior even when both count as a failed case. Keep reviewer disagreement and missing evidence visible.
The assumed time comparison is 9 minutes before and 6 minutes after, including the stated human step. Validate the same start and end points in a real study, then include retries, repair, and handoffs. Compare completed, correct bookings with the existing phone process. Include repeat calls and staff repair time instead of measuring call duration alone. A short average time is not enough if the result cannot be used or trusted.
| Item | Illustration |
|---|---|
| Invented evaluation cases | 29 |
| Cases meeting the fictional rubric | 20 |
| Cases requiring investigation | 9 |
| Observed proportion in this exercise | 69.0% |
| Assumed human minutes before / after | 9 / 6 |
Use this worksheet
Answer the questions with a source, date, owner, and next action. Mark an unknown as unknown. Use public or permissioned information and involve a qualified specialist when a question requires expertise beyond the team’s competence.
- What is the evaluation unit?
- Where did the examples come from?
- Who labeled the expected result?
- Which failures are underrepresented?
- What data must remain held out?
Questions specific to this scenario
- Who confirms the booking?
- What happens when a caller interrupts?
- How is a failed transfer recovered?
Collect permissioned test calls, interruption cases, confirmation records, and failed booking attempts. Keep different languages and noisy environments visible in the sample.
A concrete next test
Run permissioned tests with an interruption, an unavailable appointment, and a failed transfer. Ask the scheduling owner to verify the final records, then inspect what the caller was told. A correct transcript alone does not establish a correct booking.
Interpret the result
A larger dataset is not automatically more representative. An evaluation should reveal consequential behavior in the intended workflow. If repeated tuning teaches the team the answers, add independent examples before claiming a general improvement.
Compare completed, correct bookings with the existing phone process. Include repeat calls and staff repair time instead of measuring call duration alone.
Sources and method
These primary resources provide background for the concepts. The fictional scenario, calculations, and worksheet are ADIN educational examples; the sources do not report or validate those invented results. Reviewed October 11, 2026.