Morgan Jeffries, MD, Medical Director for AI at Geisinger and his colleagues, are reshaping how a large integrated health system organizes, oversees, and educates around AI — moving from a small, model-building shop to a cross-functional product organization with formal risk review and a rising emphasis on workforce literacy for AI tools.

Within Geisinger’s AI function, Jeffries and peers have restructured around product teams—pairing product managers with data scientists, engineers, and design—to convert high-potential ideas into supported solutions. He described a two-track operating model: building internal models end-to-end when Geisinger owns the problem, and “enablement” when vetting third-party offerings to ensure model transparency, local validation on Geisinger data, monitoring, and policy compliance before rollout.

As demand accelerated, Jeffries and colleagues worked with executive leadership to define a shortlist of enterprise priorities and use those as a gate for scarce capacity. He noted that program selection (what to build or enable) is distinct from risk governance (how to ensure safety and accountability). The former aligns investment with strategic goals; the latter applies to any AI-bearing program—homegrown or vendor-supplied—that enters the environment.

Governance that Scales with Risk

Asked what “good” looks like, Jeffries pointed to a systemwide intake and committee review that sorts programs by risk. Low-risk uses receive best practices and light touch. Higher-risk uses require proof of due diligence—owned by the program’s business sponsor and reviewed centrally. “They need to do an assessment of the risks associated with the AI system… and then an equity assessment… and then a monitoring plan… and then an escalation plan,” he said, emphasizing that local context matters even when vendors provide documentation. He also noted the importance of bringing multiple perspectives—technical, clinical, and ethical—into the committee’s deliberations.



He differentiated governance from procurement: new vendors are flagged through IT intake and contracting forms that explicitly ask about AI, but the tougher challenge is volume and drift—longstanding software that starts shipping AI features through routine upgrades. That reality, he argued, forces leaders to think beyond case-by-case review and toward patterns and platforms.

Preparing for Embedded AI Everywhere

The market has shifted from “we are an AI company” to “AI is inside the product,” and Geisinger is adapting its oversight model accordingly. For established platforms that enable many AI-powered workflows, Jeffries and colleagues are working with platform owners on a shared governance cadence so the oversight happens closer to where solutions are assembled. On sheer scalability, he was frank: “the centralized governance, that model is going to fall apart.” Instead, central teams should define policy, coach on risk patterns, and step in on select high-impact or ambiguous cases, while business and product owners take day-to-day responsibility for compliance.

He also entertained a question that many leaders are asking: can AI help review AI? He cautioned against over-reliance—automation bias, complacency, and de-skilling are real—but acknowledged that assistant-style tools could help less-experienced project managers consider risk and equity angles they might otherwise miss. In that sense, a well-designed copilot may be “better than the alternative,” provided subject-matter experts remain engaged and final accountability stays human.

Educating the Workforce, Not Just the Committee

The third pillar is education. Jeffries said the organization is building curricula that do two things: teach employees to use generative tools safely and effectively, and normalize a culture in which frontline staff share small wins and cautionary lessons. He sees this cultural diffusion as a practical risk contr...