NEW YORK — What happened
A bitter public dispute has erupted between commercial AI evaluators and enterprise procurement departments over how model performance is audited. Internal memos leaked from Fortune 500 tech firms reveal that up to forty percent of promised productivity gains fail to materialize in production environments.
Why it matters
Enterprise buyers are discovering that high scores on curated industry leaderboards do not translate to seamless automation of messy, unstructured enterprise workflows. This discrepancy is triggering a budget re-evaluation across financial services and healthcare sectors.
Deep dive
The root of the issue is Goodhart's Law: when a measure becomes a target, it ceases to be a good measure. Vendors design models to excel at specific benchmark questions while failing basic edge cases in proprietary databases. Enterprises are responding by building custom, private evaluation pipelines tailored strictly to their own historical data logs rather than relying on public scorecards.
Report check (claims vs what is verified vs still rumor)
Vendors claim standardized evals are sufficient for risk assessment. Auditors have verified that private, task-specific testing suites yield significantly lower success rates than vendor-published metrics. Unverified rumors suggest major insurance carriers are preparing liability exclusions for unsupervised AI workflows.
Open questions
Will enterprise demand force the establishment of an independent, government-backed bureau for AI safety and performance auditing? How will procurement teams adapt to software that changes its behavior via silent cloud updates?