The Algorithm vs. The Patient: Navigating AI in Utilization Management

In my dual life as a practicing radiation oncologist and a healthcare executive, I live on both sides of the "submit" button.

On one side, I am the clinician sitting with a cancer patient, explaining a treatment plan that offers them the best chance at survival, only to face the administrative hurdle of prior authorization—the "click and wait" that delays care and fuels physician burnout. On the other side, as Chief of Radiation Oncology at Luminis Health, I look at the balance sheet. I understand the financial necessity of utilization management (UM) to ensure appropriate use of resources and the long-term sustainability of our health systems.

We are now entering a new era where Artificial Intelligence (AI) promises to revolutionize this friction-filled process. But as I’ve argued in recent policy discussions, we must ask: Is AI being deployed to expedite care, or merely to automate denials?

The "Black Box" Problem

The promise of AI in utilization management is speed. Theoretically, an algorithm can review clinical notes against payer policies in seconds, granting real-time approval for standard-of-care treatments. This is the "value-based innovation" we should be striving for.

However, the reality is often more opaque. We are seeing a trend where algorithms are used to flag "medical necessity" denials with little human intervention. This "black box" approach is dangerous. It risks removing the nuance of complex cancer care, where a patient’s unique history might not fit a binary code.

In my recent panel on the role of AI in utilization management, I emphasized that efficiency cannot come at the cost of transparency. If an algorithm denies care, a physician must be able to understand why—and a human peer must be available to review that decision. Without these guardrails, we risk a system where "efficiency" is just a euphemism for barrier-to-care.

Policy as a Guardrail

My work with ASTRO’s Health Policy Council has focused heavily on reimbursement reform and advocacy. We are pushing for a regulatory environment where AI is a tool for augmentation, not replacement.

Recent CMS rules are moving in the right direction, emphasizing that algorithms cannot be the sole basis for denying care. But policy moves slower than technology. As I discussed in my podcast, Value Health Voices, we need to ensure that the economic models we build today don't bake in the biases of yesterday’s data. If an AI is trained on historical denial data that disproportionately affected marginalized communities, we risk automating inequity—something that stands in direct opposition to my work in creating equitable models of care.

The Path Forward: "Gold Carding" and Trust

So, how do we fix it?

1. Gold Carding: AI should identify providers who consistently adhere to evidence-based guidelines (like NCCN or ASTRO) and "gold card" them, exempting them from routine prior authorizations. This reduces administrative burden and lets payers focus their resources where they are actually needed.

2. Human-in-the-Loop: Denials must always trigger a human review. Period.

3. Standardization: As I’ve advocated for in my work on the Radiation Oncology Case Rate (ROCR) model, moving toward episode-based payments can align incentives better than fee-for-service battles over every fraction of radiation.

AI has the potential to remove the administrative sludge that slows down our healthcare system. But it requires physician leaders who understand both the code and the clinic to ensure it serves the patient, not just the bottom line.