Our Take
HSS has a concrete win, but the MIT piece conflates a working back-office tool with a mandate to overhaul clinical workflows—and buries the fact that patient-facing triage still routes complex cases to humans.
Why it matters
Two-thirds of health systems are already running AI agents in production (per KPMG). The question is no longer whether to deploy them, but whether your organization has the data governance and governance structure to avoid failures that will inevitably surface when these systems touch patient decisions.
Do this week
Health IT leads: audit your patient data fragmentation across departments and legacy systems this week. HSS's success hinged on unified data strategy before agents touched workflows—yours will too.
Hospital for Special Surgery deployed AI agents across back-office and patient-facing workflows
At Hospital for Special Surgery (HSS), an academic medical center in New York specializing in musculoskeletal care, AI agents now handle insurance claim processing that previously required weeks and manual coordination between staff and third-party contractors. The agents complete 1,100 claims per month. Appeals that took 45 minutes to process now take 5 minutes, and approval success rates rose from 65% to 100% over nine months (company-reported).
HSS has since deployed a second AI agent system in patient-facing settings: a scheduling and triage service built with enterprise AI developer Ema Unlimited. The service runs 24/7 via web, text, or phone. It asks patients clarifying questions about their condition, then books appointments with the appropriate clinician, accounting for location, insurance, and availability.
The triage agent escalates sensitive, complex, or uncertain cases to human specialists. Every decision is auditable and staff can override at any point. HSS is also building a dedicated AI lab to train all staff on building and deploying agents across the organization.
Data fragmentation is the real constraint—not agent capability
HSS's back-office win is real. But the triage service reveals the asymmetry: the agent can only be as smart as the data it can access. Dr. Ashis Barad, chief digital and technology officer at HSS, notes that patient data remains fragmented across departments and legacy systems, each with its own definition of basic metrics. For example, "time to start surgery" has a slightly different definition in each hospital he's worked in. This fragmentation means AI agents cannot reliably retrieve information across sources or assimilate the tacit knowledge that differentiates them.
The MIT piece frames agentic AI as analogous to electricity, a general-purpose technology that will overhaul health care. That's marketing language. What the piece actually demonstrates is that success requires building unified data infrastructure first. The back-office wins come cheap. The clinical wins require organizational redesign.
HSS's approach to governance matters more than the agent itself. Decisions about AI deployment flow through an AI subcommittee co-chaired by the CDTO and a senior nursing executive. Clinical-adjacent agents get far more scrutiny than back-office tools. That friction is a feature, not a bug.
Check your data strategy before you buy an agent
Two-thirds of health systems are already running agents (per KPMG). Most will deploy them against fragmented data and then blame the agent when triage fails or claims get routed wrong. HSS's lab is the right move: train everyone on the constraint, not the technology. Before you acquire or build an agent, audit which data sources it will need to access, whether those sources can talk to each other, and what definitions live in each one. If you do not have a unified data strategy, an agent will amplify your fragmentation, not collapse it.