Ask a health system where AI has produced unambiguous value and a surprising number of answers point to the pharmacy. It is not a glamorous domain and it generates almost no conference keynotes, which may be related to why it works.

The conditions are close to ideal. The data is structured, the rules are well-specified, ground truth is available quickly, the intervention is a recommendation to a pharmacist rather than a patient, and the failure mode is a rejected suggestion rather than a harm. Models in this environment can be evaluated continuously and cheaply.

What deployments report is steady, compounding improvement: interaction alerts with dramatically lower false positive rates than the rule-based systems they replaced, dose adjustment suggestions that pharmacists accept at high rates, and formulary substitution that runs largely unattended.

The lesson most frequently drawn — that clinical AI works where feedback is fast and the stakes of a single wrong suggestion are low — is not new. It is worth restating because the domains that get funded are frequently the ones where neither condition holds.