I recently spent a recent afternoon with chief medical officers, chief population health officers, population health IT directors, and senior medical directors from leading health systems, at a roundtable hosted by Bright Spots in Healthcare and moderated by the organizations’ CEO, Eric Glazer, to talk about rising risk. Follow-up, team structure, technology, and cost dominated the conversation. As the models get better, that's where more of the work has moved — downstream of the prediction.
Most wish-list answers were about what happens after the model fires
Eric asked the group, one by one, what would most change the trajectory of rising risk management. The answers ran to clinical pathways, reimbursement for preventive work, capacity to reach patients, and the money case for automation. Prediction accuracy barely came up. What surprised me was how far along the case studies already were. One health system had an AI model scoring hospitalization risk across specific chronic disease cohorts. Another had a prognostic tool tied to capitation contracts. A third ran a remote-monitoring program that cut heart-failure hospitalizations by roughly 57 percent once it launched. One tiered-response model had rising-risk patients worked by a standing group — PCP, specialist, pharmacist, and care manager — doing proactive registry outreach together.
Continuous, proactive monitoring can lead to untenable volume
One aspiration came up again and again. Continuous, personalized monitoring, paired with intervention timed to match it, so a patient's risk updates the moment something changes and a care team can act that same week.
Nobody in the room had fully solved what that actually requires. One example stuck with me. An AI query at one organization returned 25,000 patients needing urgent attention by year end — a number no care team could act on. That list was the direct output of exactly the kind of higher-frequency, more sensitive monitoring the room said it wanted, and it broke the workflow built to receive it. A model that runs continuously produces more output, more often, and all of that still has to land somewhere a person can act on it. The programs that held up treated identification and follow-through as one loop, tracking who got flagged and whether the right intervention actually happened.
Getting there is a staffing and workflow problem as much as a modeling one — and, on the evidence in the room, still an unsolved one.
Staffing and workflow are as much a problem as risk modeling
Whenever the conversation turned to what was stopping people from doing more, the answer was financial, not technical. Preventing a disease doesn't generate a bill. The savings from an avoided hospitalization land on a payer's balance sheet, well after the intervention that produced them. Every dollar spent on outreach, monitoring, or automation shows up immediately on this year's budget; the dollar it saves shows up later. That gap is structural, built into how value-based care pays out. The ROI conversation around agentic AI in care management runs into the same wall. A business case built to pay off late still has to clear a budget cycle that wants returns now. Integrating a new tool into existing systems eats months of already-scarce IT and staff time, and nobody has settled whether an agent should run as one platform across the system or as a separate build for every department. One system found a way around the delay, though it depended on a starting condition most organizations don't have: as its prognostic model improved, capitation risk shifted from the health plan side to the care-delivery side, making the ROI tangible to its own CFO. Organizations that don't hold capitation risk directly don't have that lever. They still have to make the case for agentic AI the hard way, against a budget cycle built for near-term returns.
The patients you can't reach may be the ones you need most
One statistic stayed with me. Even the highest-risk patients are reached only about 30 percent of the time. One person in the room raised the harder possibility directly: the patients who can't be reached may be the ones most in need. In one hospitalist-at-home program, care teams doing video visits found people who weren't taking their medications, didn't understand their discharge instructions, or had no support at home — none of it visible in the record beforehand. That same program was eventually cut for reimbursement reasons. Any rising-risk program built only from what the data already shows will keep missing the group most likely to need it.
.png)
