There’s an odd mismatch in healthcare AI right now. Models keep getting better at discrete tasks, yet the operational burden inside health systems remains stubbornly familiar. People still move between portals, chase documentation, reconcile exceptions, and manage handoffs across systems that were never designed to work together.
Dr. Nigam Shah, chief data scientist at Stanford Health Care, has made a related point while discussing Stanford’s ChatEHR initiative, arguing that AI is only useful when it’s embedded into the workflow and grounded in real medical context. It’s a useful outside check on the argument, not the source of it. Consider a workflow like claims or eligibility review: the model can draft a summary or flag a discrepancy in seconds, but the model was never the hard part. The hard part is everything around it.
The Proof Is in the Gap
HealthAdminBench, a 2026 Stanford-led benchmark, tested AI agents against realistic administrative work, prior authorization, appeals, durable medical equipment ordering, across simulated EHR, insurer portal, and fax environments. The strongest system hit 82.8% success on individual subtasks. Its end-to-end completion rate was 36.3%.
That gap is the whole story. Enterprise work is a chain of decisions, integrations, policy checks, and exception paths. A model can execute several links well and still produce nothing usable if one link breaks.
A System Can Work and Still Be the Problem
This is a lesson we see play out across every modernization effort, not just AI ones: a system can be technically functional and still be the constraint on the business built around it. Nothing has to be broken for a workflow to be holding an organization back. It just has to have accumulated enough friction, enough manual reconciliation, enough handoffs, that people are working around it instead of through it. That kind of drag is easy to miss precisely because nothing is failing.
It also means workflows don’t need AI simply because AI is available. That reframes the question away from where can we add AI? toward something more specific: what does the organization need this workflow to accomplish next, and what combination of existing technology, integration, modernization, automation, AI, and human judgment actually gets it there? Sometimes that answer is a model. Just as often, it’s cleaner data access, a rebuilt integration, or removing a handoff entirely, with AI playing a smaller role than the initial ask assumed.
The Workflow Is the Product
Once you accept that, the design principle follows: the workflow is the product. The model is a component inside it, not the product itself. Enterprise value is experienced through the completed workflow, not the API call.
That’s why the unglamorous parts of the build carry the weight. Which system of record does the AI pull from, and is it one the adjuster already trusts? What are the permissions and data governance around it? What happens on exception, and where does a human still need to sign off? Skip these, and you’ve added intelligence without removing work, an AI-shaped demo sitting on top of the same fragmented process.
Judge It Like an Operator, Not a Demo
Once something is in production, the right questions are operational, not architectural. Did turnaround time actually fall? Were fewer manual touches required? Did staff adopt it, or route around it? Did the process get easier to audit, or did the effort just move somewhere less visible? Those questions, not model benchmarks, are how an organization should judge whether AI created real value in a given workflow.
Where This Leaves Healthcare Leaders
Healthcare doesn’t need fewer AI experiments. It needs a clearer discipline for applying them: understand what the organization needs a workflow to accomplish next, preserve what’s already creating value, fix what’s creating friction or constraint, modernize what needs to evolve, and use AI specifically where it materially improves the outcome.
Have a workflow that’s technically working but increasingly holding the business back? Tell us where the friction is; we’ll help you figure out what actually needs to change.
