Ashish V. Shah, Chief Product and Technology Officer, Integrated Home Care Services (Integrated)
The first time we tried to solve care coordination with AI, the approach felt intuitive: build one capable assistant and let it absorb as much of the workflow as possible. That was health technology’s prevailing assumption: the better the model, the more work it could handle.
That was the wrong frame.
handle everything: provider matching, documentation, visit confirmation, fulfillment tracking, exception flagging. It performed well in a controlled environment. At production volume, thousands of referrals a day, it broke down.
Not because the technology was insufficient. Because the job we’d given it was incoherent.
Care Coordination Is Not One Job
When you look at how a real care coordination team operates, you find specialists, not generalists. One person focused on provider matching. Another managing authorization documentation. Another confirming visit completions and escalating exceptions. Another handling follow-up calls when something falls through.
Those roles are separated for a reason that has nothing to do with efficiency. It’s about accountability. When the same person handles everything, you cannot identify which part of the process produced a bad outcome. When a provider match takes four hours instead of forty minutes, was the problem the matching logic, the documentation step, or the confirmation call? In a generalist model, you cannot know.
AI deployment ran into the same problem. Asked to be a generalist coordinator, it gave us no way to diagnose what broke when something did, which sent us back to first principles and toward the model we run today.
Healthcare’s coordination problem is not primarily a communication issue. It is an ownership issue. The system has never been staffed for the amount of coordination the work actually requires. Access breaks when follow-up fails. Reliability breaks when ownership is unclear. Those are not edge cases. They’re what happens when coordination work isn’t explicitly staffed.
The Question We Had to Answer
If we were staffing this operation from scratch, how would we structure it?
The answer wasn’t “one excellent coordinator.” It was a set of discrete roles, each with a defined scope and a measurable output. Provider matching. Authorization management. Visit confirmation. Documentation review. Each a job with a job description.
So that is how we built it. A digital workforce, where each AI role had one job, one defined scope, and one output that fed the next step in the sequence. Not a generalist. A team of specialists, each accountable for a specific part of the work.
Credentialing verification. Referral fulfillment coordination. Documentation reconciliation. SLA tracking. Work that had always existed but rarely had a clear owner. Once each task had an explicit owner, the pattern became consistent in a way a general assistant never produced.
Provider matching became its own function, built for one job only. It matches referral requirements to provider credentials, geography, availability, and network status. The outcome measure is match time. Our current benchmark is 60 minutes or less. That number is achievable only because the matching function is not competing for attention with documentation review and follow-up calls.
Dina, that model came with it. The digital workforce approach was already running in production. What changed was the operating environment. Integrated processes thousands of referrals each day across dozens of health plan partners, each with specific performance requirements, operating in more than a dozen states. Scaling a proven model into that environment meant proving it could hold up under real constraints, at real volume, with no room for error.
What Health Plans Should Ask
When health plans evaluate AI solutions for care coordination, the questions tend to focus on capability. Can the AI handle scheduling? Can it process clinical documentation? Can it manage authorization workflows?
Those aren’t the wrong questions. They’re just not the right starting point.
The right starting point is job design. What specific role does this AI perform? What’s its defined scope? What happens at the boundary of that scope, and who handles it? Is this deployment running in production or in a controlled pilot? What’s the evidence of performance under real operational conditions, at real volume, with real members?
A general-purpose AI assistant that claims to handle all of care coordination is making the same promise we made with our first deployment. It may be technically capable. It may perform well in a test environment. What it can’t do is produce the accountability and measurable consistency that a role-defined digital workforce provides.
The staffing decision isn’t a technology question. It’s an operational design question. Health plans that get it right are the ones that ask vendors for the job description, not just the capability list.
Our answer to health plans asking about our AI is consistent: skip the slides and ask to see what it does in production, at real volume, with real members. That’s the only standard of proof that matters.
Ashish V. Shah is the Chief Product and Technology Officer at Integrated Home Care Services, where he leads the AI infrastructure and digital workforce strategy behind Integrated’s care coordination operations.
Learn more about how IHCS is innovating when it comes to home care benefits. Reach out to our team today.
