Founders
How to build an evaluation dataset for AI customer support agents
AI customer support 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 team drafts replies to delivery and refund questions for an online retailer. A support manager owns the queue, while finance controls refund authority. Seasonal peaks change the mix of easy and difficult tickets.
A reply that sounds fluent can still promise the wrong refund. Keep suggested language separate from permission to change an order or issue money.
An evaluation sample with a denominator
Imagine a permissioned evaluation set with 32 cases and 22 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 reply that sounds fluent can still promise the wrong refund. Keep suggested language separate from permission to change an order or issue money. Break the 10 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 12 minutes before and 7 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 the draft-and-review workflow with the current queue, including time spent fixing errors and handing cases back to people. A short average time is not enough if the result cannot be used or trusted.
| Item | Illustration |
|---|---|
| Invented evaluation cases | 32 |
| Cases meeting the fictional rubric | 22 |
| Cases requiring investigation | 10 |
| Observed proportion in this exercise | 68.8% |
| Assumed human minutes before / after | 12 / 7 |
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 can approve a refund?
- Which issues require a human handoff?
- Does the test include peak-season complaints?
Use anonymized ticket histories, escalation reasons, reviewer edits, and repeat-contact counts. Record whether a customer issue was actually resolved rather than counting a generated reply as success.
A concrete next test
Choose one ordinary delivery question, one refund exception, and one repeat complaint. Have the support owner trace each from intake to a usable resolution. Record where draft quality, authorization, and customer follow-up require different evidence or different people.
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 the draft-and-review workflow with the current queue, including time spent fixing errors and handing cases back to people.
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.