Governance dashboards count model invocations without recording the exact internal data slice or fine-tune checkpoint used for each one.
Procurement teams sign off on AI platforms that expose only invocation counts and latency metrics. The contract never requires the vendor to expose the training subset identifier attached to each inference.
When a compliance recommendation later proves false, investigators receive a timestamp and user ID but no pointer to the data window or prompt template. Reproducing the error becomes impossible.
Shadow deployments accelerate the problem. Business units spin up direct API access through cloud credits, bypassing the central registry that was supposed to enforce lineage capture.
Board packs then present aggregated accuracy figures that collapse once the underlying slice is questioned. No one can isolate whether the failure came from outdated reference data or an unapproved fine-tune.
Fix the procurement clause first. Require every approved model endpoint to return a deterministic slice ID and prompt hash with every output before the first production token is accepted.
Audit teams must then sample those IDs against the actual source tables rather than trusting summary dashboards. Only then does accountability attach to a concrete decision instead of an abstract invocation total.