clinicians.build · July 11, 2026

The 40% Nobody Logged

A health system almost renewed a clinical decision-support contract on the strength of its validation study and FDA clearance. Then the CMO pulled the override log: physicians had reversed more than 40% of the tool's recommendations over six months. Nobody had flagged it. Here are 100 of those recommendations.

Primary source: Healthcare AI's next challenge isn't adoption, it's reliability (SmartBrief, Jul 2026) — Sina Bari, MD

100 recommendations. Six months. One override log nobody opened.

Each square is one recommendation the cleared tool made at the point of care. Navy means the clinician accepted it. Red means they silently overrode it. Press play — then ask the only question that matters: of the reversals, how many did anyone review?

1 cleared CDS tool · 100 recommendations Month 0
0
Accepted
0
Silently overridden
0
Overrides anyone logged
Press “Play six months” to watch the recommendations resolve.
Accepted Overridden Pending

The debate about whether doctors will use AI is settled — daily physician use hit 63% this year. The unanswered question is whether the tools already running still work where patients depend on them, and whether anyone would know if they didn't.

63%
of physicians use AI daily (2026)
40%+
of this tool's recs silently overridden
6 mo
it ran before anyone checked
“The system was technically live and practically ignored.”— Sina Bari, MD, on a 40%-override clinical decision support tool

An override isn't automatically a save

The comforting read is “40% overrides means the safety net worked — the doctor caught the tool's mistakes.” Sometimes. But an unmonitored 40% doesn't tell you humans are catching errors. It tells you nobody knows what the tool is doing: whether it drifted, whether it's crying wolf, whether the 40% is smart clinicians or a broken model. A safety net you can't measure is just a story you tell the renewal committee.

The half of the picture that never got drawn. “Cleared” and “validated” are point-in-time snapshots. “Deployed” is a moving target. Between them sits a layer almost nobody builds: override telemetry, drift detection, the per-recommendation audit trail. That's why this grid can turn 40 squares red and still show zero logged — the reversal data existed in the EHR, but no one had wired it into anything that would raise a hand.
Every red square above was a decision the clinician made against the tool. Reveal the monitoring layer and you see the real failure: not the overrides, but the six months of silence around them.
Physician-pull gets a clinical tool adopted. The override log is what gets it renewed — or quietly killed. If you ship clinical AI, instrument the override from day one: accepted, edited, or rejected, logged per recommendation. That single number is the thing a CMIO cannot get from your competitor's FDA certificate, and it's the difference between a one-meeting renewal and a tool that silently rotted for half a year.