Writings
Writings on AI, data, and the organizational realities behind success and failure.
Cleared for Deployment?
Before an aircraft enters controlled airspace, someone explicitly authorizes the crossing. Clinical AI has no equivalent moment. Models routinely move between hospitals without anyone formally assessing whether the patients, documentation practices, and workflows resemble those they were trained on.
Software Has Become Cheap. Institutional Knowledge Has Not.
Agentic coding has made software dramatically cheaper to build. But it has also exposed a different bottleneck: understanding how organizations actually work. As implementation costs collapse, institutional knowledge, not code, becomes the scarce resource that determines whether transformation succeeds.
The Feedback Loop Nobody Built
Most healthcare AI governance focuses on deployment, validation, and compliance. Far less attention is paid to what happens afterward. When nobody owns the process of monitoring real-world performance, clinician adaptation becomes the only feedback loop that reliably functions.
Why the Person Holding the Map Still Matters
Models are useful precisely because they simplify reality. The mistake is expecting them not to. From clinical scores to AI systems, the real challenge is not building perfect maps, but developing the judgment to know where their edges are.
The Validation Trap
Validation exists to reduce uncertainty, not eliminate it. Yet many organizations quietly transform validation into something else entirely: a mechanism for avoiding decisions. The result is not caution. It is a different kind of risk.