Talking AI — Evals: what proves an AI holds up, beyond the demo

An evaluation (an "eval") is a set of representative test cases that measures, with hard numbers, whether an AI answers correctly — before it goes live, and again at every update. A polished demo proves nothing: a model that shines on a few hand-picked examples can be wrong half the time on real cases, and drift silently when the vendor ships a new version. The right question to a vendor isn't "what can your AI do?" but "how do you measure its accuracy on MY cases, and can I see the test set and the score?". Before any rollout, build a set of real, already-validated cases with your teams, set a target accuracy score, and require it to be re-measured at every new version.

Source: Gartner, "40% of Organizations Deploying AI Will Use AI Observability to Monitor Model Performance by 2028," 12 May 2026 (share of software engineering teams using AI evaluation and observability platforms: 18% in 2025, a projected 60% by 2028).