Your AI passed the benchmark. Can it survive the bedside?

Scroll through one emergency department with a triage AI in it, and see the three places its benchmark never looked.

What your benchmark can’t see.

  1. A small, brightly lit box in the dark: six case cards on its floor, each with a check, read by a light-blue screen.

    Benchmark:Your triage AI passed its benchmark. Every case it saw sits in this box: complete, labeled and well lit.

  2. The lit box sits in the middle of a large, dark emergency department with a triage desk, a waiting room, beds and an exit.

    The department:The emergency department is bigger than the box. Much of what decides a visit happens where the data does not reach.

  3. A pool of light at the exit: two people walk out of the waiting room and through the door before anyone sees them. Marker one stands by the door.

    Exit:Some patients leave before anyone sees them, and they are more likely to come back, some of them sicker. A model trained on finished visits never meets them.

  4. A pool of light on one chair: an older patient sits with a card that says Feels weak and a light-blue middle score drawn from the box. Marker two stands beside her.

    Waiting room:One middle score can hold very different risks. An older patient who says only that she feels weak may be the sickest person in the room.

  5. A pool of light on the triage desk: a stretcher comes in and the nurse goes to it, while the AI’s advice glows unread on the desk screen. Marker three stands by the desk.

    Triage desk:A tool can be right and still fail. When the room fills, people stop reading what it says.

  6. The box walls fold down flat. Light-blue lines run from the AI to each of the three markers, and each marker turns to a check.

    Retest:Each blind spot becomes a test. Fix what fails, then test again in the room, not the box.

  7. The whole department in soft light, calm. The AI’s screen now sits at the triage desk with the nurse.

    The room:You can’t test for what you’ve never seen. We see it every shift.

Sample findings memo

  1. Exit

    Patients who leave unseen.

    Test: include visits that never finished.

  2. Waiting room

    Older patients with vague complaints.

    Test: check results by age and complaint, not only overall.

  3. Triage desk

    Advice unread when the room fills.

    Test: measure use under load, not only accuracy.

  • Go
  • Fix first
  • Stop

Fix first, then retest in the room.

Two emergency physicians. Still on shift.

Get a red-teamed clinical breakdown of your product.

Book a call

Not ready to book? Send a note.

I am:

Please do not include patient information. We cannot give medical advice or discuss anyone’s care. In an emergency, call 911.

How we handle your note