KPI

Accuracy: AI benchmark

Share of outputs that are correct on a checked sample.

How to measure it

Human review of a random sample of outputs against a written standard.

Across the library

Median 91% across 33 deployments, reported range 42% to 100%, plus 2 "up to" values not pooled.

Higher is better. Unit: percent.

Reported values by use case

AI document intelligence for unstructured forms and documents

AI for IT incident triage and root cause analysis (AIOps)

AI clinical trial patient matching and prescreening

Real time AI assist for contact centre agents

Use cases that should track accuracy