Healthcare providers are desperate. With clinician burnout hitting critical levels and administrative burdens eating up more of doctors' time than patient care, a full 68% have already adopted AI agents into their workforce, according to KPMG research. But most are still fumbling through point solutions—narrow tools that automate tiny slices of workflow without actually changing how hospitals operate. Hospital for Special Surgery in New York City is trying something different: treating agentic AI as foundational infrastructure rather than another checkbox on the digital transformation roadmap.

The Claims Processing Experiment

Dr. Ashis Barad, HSS's chief digital and technology officer, started with insurance claims because they're high-volume, rules-based, and currently a nightmare. Previously, handling appeals required both internal staff and a third-party contractor working for weeks to manage the workload. Since deploying AI agents nine months ago, HSS processes 1,100 claims per month entirely in-house. The appeals stage dropped from 45 minutes to five. Success rates climbed from 65% to 100%. Those aren't incremental improvements—they're an order-of-magnitude shift that frees human workers for actually complex problems.

Triage Goes Autonomous

Building on those backend wins, HSS is now rolling out patient-facing AI with Ema Unlimited, an enterprise agentic AI developer. The system handles scheduling and triage around the clock via web, text, or phone—asking clarifying questions about symptoms and routing patients to the right specialist based on location, insurance coverage, and physician availability. Dr. Barad describes it as completing "the whole loop," trained on all of HSS's context, rules, and knowledge base. The kicker: sensitive, complex, or ambiguous cases still escalate to human specialists. Every AI decision remains auditable, and staff can intervene at any point.

Guardrails Are Non-Negotiable

Given that healthcare decisions carry life-or-death stakes, HSS isn't treating safety as an afterthought. An AI subcommittee co-chaired by Dr. Barad and a senior nursing executive reviews all patient-facing deployments with far more rigor than backend processes like claims handling. The organization is also creating a dedicated AI lab at its Manhattan campus to democratize access—offering classes and one-on-one training so clinicians across departments can build and understand agents rather than treating them as black boxes handed down from IT.

Data Fragmentation Remains the Enemy

Dr. Barad frames agentic AI as "a general-purpose technology, analogous to electricity"—meaning it only delivers value when built on proper infrastructure. For HSS, that means unifying fragmented data sources across departments and legacy systems into a single source of truth. The problem is real: even basic metrics like "time to start surgery" have varied definitions across hospitals he's worked in. Without standardized, interoperable data, agents can't retrieve information from multiple sources or build the institutional knowledge that makes them genuinely useful rather than just expensive autocomplete.

Key Takeaways

  • Agentic AI at HSS handles 1,100 insurance claims monthly with appeals success jumping to 100%
  • Patient-facing triage service operates 24/7 and escalates complex cases to human specialists
  • An AI subcommittee scrutinizes patient-care deployments more heavily than backend processes
  • Deloitte research shows leading adopters redesign end-to-end workflows rather than deploying narrow tools

The Bottom Line

This is what healthcare automation actually looks like when it's done right—not chatbots that frustrate patients, but agents embedded deep enough in operations to matter. Dr. Barad's 90% non-clinical task automation target sounds ambitious until you realize the claims processing results are already proving it possible. The hospitals still dithering on AI strategy aren't behind the curve—they're running out of excuses.