Duke's sepsis AI worked by flipping hospital hierarchy: nurses alert the doctors

Sepsis kills fast, and Duke Health's answer was Sepsis Watch: an AI delivered as an iPad app that scores every emergency-department patient on an hourly basis for the likelihood of developing sepsis, flags medium- and high-risk patients plus those already meeting criteria, and routes them for rapid treatment once a doctor confirms. In its first two years, Duke's hospital managers and clinicians reported the tool dramatically reduced sepsis-induced patient deaths, and it entered a federally registered clinical trial expected to report in 2021.

The technical victory is only half the story, according to a two-year Data & Society study by cultural anthropologist Madeleine Clare Elish and Elizabeth Anne Watkins. Making the tool work required what they call 'repair work': designing new communication protocols, writing training materials, and navigating workplace politics. The deepest change was authority. Physicians traditionally make the diagnoses but could not monitor another app; rapid-response nurses could — so a nurse now calls an attending to challenge their read on a patient.

That protocol reversed the typical chain of command ('Are you kidding me?' was one nurse's first reaction), and the nurses and attendings involved often had no prior relationship, having worked in different sections of the hospital. Duke did the repair anyway — and the repair, not just the model, is what the case credits for the outcome.

The model was the easy half: it worked because Duke let nurses watch the scores and call attending physicians directly — reversing the hospital's chain of command.

It names the hidden cost of deploying AI in institutions: 'disruption is also a form of breakage' — old protocols die and hierarchies scramble, and someone must repair them.

The design move is counterintuitive for medicine: nurses carry the monitoring burden and get standing to call doctors — the only workflow that survives an hourly scoring cadence.

Real-world evidence exists on both halves — reported reductions in sepsis deaths in practice, plus a registered clinical trial — making it a rare honest AI-in-production case study.

AI adoption is social labor: redesigning protocols, training staff and navigating power dynamics turns a model into outcomes. Give the tool to whichever workflow fits, even if that inverts seniority.

Sepsis Watch became part of a federally registered clinical trial expected to share results in 2021, and the Data & Society case study published the repair-work findings — new communication protocols, training materials and negotiated power dynamics — as the documented price of getting an AI tool into daily clinical use at Duke Health.