Health Systems Are Being Told to Treat AI Safety as Core Infrastructure
A new policy analysis from the Margolis Institute argues that AI safety in health systems requires real infrastructure and stronger risk management practices. The key implication is that governance can no longer live at the margins of innovation teams; it has to be embedded into procurement, oversight, and daily operations.
The Margolis Institute’s focus on AI safety infrastructure is an important intervention because it reframes the debate away from isolated concerns about model performance and toward system capability. In practice, health systems need inventories of AI tools, clear lines of accountability, monitoring processes, incident reporting mechanisms, and governance that extends beyond initial deployment.
That shift is overdue. Many organizations now face a mixed environment of vendor models, embedded EHR features, department-led pilots, and informal employee use of external tools. Without infrastructure, even well-intentioned governance quickly becomes reactive. Safety failures are then more likely to emerge through workflow confusion, overreliance, bias, or poor escalation design rather than spectacular algorithmic breakdowns.
The report’s framing also reflects a broader reality: AI risk management in healthcare increasingly resembles other enterprise risk domains such as cybersecurity, infection control, and medication safety. It requires designated leadership, repeatable controls, and organizational learning loops. Treating AI oversight as an occasional committee exercise will not scale.
For executives, the strategic takeaway is stark. As AI becomes more deeply embedded in care delivery and operations, the absence of governance infrastructure itself becomes a risk factor. Health systems that move now can define standards internally before external mandates force a more hurried response.