AI abstention design
Count unnecessary refusals as well as unsafe answers
A stricter assistant can look more accurate by answering fewer questions. Measure the quality of answered cases and the useful work lost through abstention.
In this article
Label the intended outcome first
Build a reviewed set containing answerable questions, questions needing clarification and requests that should be declined or escalated. Keep authority denials in a separate category from evidence uncertainty.
The labels should come from the task and source material, not from the current model's behaviour. Otherwise a model that already refuses too much can define its own standard as correct.
Record the required facts or next step for each case. A generic refusal label does not establish whether the user receives a useful explanation or route forward.
Use denominators that tell different stories
Coverage is the proportion of cases answered. Error among answered cases measures how often those answers fail the acceptance rule. Unnecessary refusal measures withheld cases that the system should have answered under the reviewed standard.
For an illustrative set of 100 questions, answering 80 with four incorrect answers gives 80 percent coverage and a 5 percent error rate among answered cases. If eight of the twenty withheld cases were answerable, then 40 percent of withheld cases were unnecessary refusals. These figures describe different properties and should not be collapsed into one number.
Also report counts so small categories are not obscured by percentages.
Track clarification and escalation outcomes
Measure whether a clarification led to a correct answer after the missing detail was supplied. Repeated questions, abandonment and unresolved loops reveal problems that a first-turn abstention score misses.
For escalations, inspect whether the recipient had enough context to resolve the issue. A technically correct refusal can still transfer excessive work if the application discards the evidence already gathered.
Break results down by task consequence. An unsupported operational instruction and an unnecessary refusal of a low-risk summary have different costs and may justify different policies.
Compare decision rules on the same cases
When changing thresholds, keep source snapshots and evaluation criteria stable. Review cases whose outcome changes, especially newly answered cases near the boundary.
Do not assume a relevance or model confidence score is calibrated across task types. Measure the observed behaviour for the class being controlled.
Choose the rule that meets the product's requirements for both dependable answers and useful completion. Keep deterministic access controls outside this tradeoff. No improvement in coverage should allow the assistant to answer from information the user is not permitted to receive.
Primary sources
Microsoft Learn: retrieval and answer evaluatorsReferences checked 11 September 2026.