Answer framing: Use buyer-controlled selection, an independently adjudicated reference standard and a locked evaluation set. Choose sample size from the decision’s risk and desired precision, not a supplier’s preferred demonstration volume.
Why it matters to buyers: NIST’s AI Risk Management Framework calls for documented test sets and metrics, assessment under conditions similar to deployment, and involvement of experts or assessors outside the front-line development team. These are general AI evaluation principles, not a property-claims certification. (airc.nist.gov)
Put this to: Gromowski, Falchuk and the carrier transformation/QA team.
A good answer should cover this proposed protocol:
- Pre-register the decision: acceptance standard, materiality rules, comparison scope, cost allocation and stop conditions.
- Separate configuration from evaluation: allow tuning on development files, then lock the holdout set and relevant system versions.
- Stratify the work: water/wind, complexity, geography, inspection channel and evidence quality.
- Use two complementary tests: matched-file comparison for technical quality and a controlled live workflow test for actual effort and elapsed time.
- Calibrate reviewers: independently double-review a subset and adjudicate disagreement against evidence—not against whichever estimate was historically accepted.
- Retain all outcomes: refusals, timeouts, manual rescues, escalations and unfinished files remain visible.
- Report uncertainty: ask what error-rate bounds and subgroup sample sizes support the proposed rollout.
Sharp follow-up: “Which deployment population does this pilot not justify serving?”
Artifact to request: a pilot protocol, selection log, holdout register, reviewer instructions and analysis workbook. Use authorised, minimised data; do not upload live policyholder files to a public research tool.
Expert quote: [Reserved for approved quotation, contributor, role and interview date.]
Sources cited on this page
- airc.nist.govairc.nist.gov