Who Owns the Outcome? The shift from ‘Capability’ to ‘Accountability’ for AI Healthcare platforms with Pritesh Patel, Chief Operating Officer, Andor Health     

Ask a health system executive what their newest AI tool has done for length of stay, and watch the answer be “it didn’t change”. The tool works. Clinicians use it. It generates notes, flags risk, drafts messages. But when the question turns from what the software does to what the number did, the conversation shifts from the vendor to the customer’s own staff. That shift, from capability to accountability, is the fault line running through healthcare AI right now. Pritesh Patel, chief operating officer of Andor Health, has a plainer way to name it.

“The only question that separates the field is who owns the outcome,” Patel says. “Everything else is a feature comparison. A health system can buy the best tool on the market and still miss the number it bought the tool to move, because the tool was never the thing that moved it. People moved it, working inside the tool. And if nobody on the vendor side is responsible for those people or that result, then the vendor has sold a capability and left the outcome on the customer’s desk.”

That framing sounds obvious until applied to the actual market. Software vendors sell a product and expect the buyer to staff it. Staffing firms send people and bring no technology. Both hand the result back to the customer. The gap between them, a company that supplies the technology and the clinical work as one accountable unit, is thin, and Patel argues it is the only place outcomes reliably move.

The Three-Layer Test

To make the distinction concrete rather than rhetorical, Patel uses a test he applies to any vendor claiming to change clinical results. It has three layers, and the point is how few companies clear all three.

The first layer is technology: the infrastructure, security, integration, and connectivity that let AI operate inside clinical workflows instead of beside them. This is table stakes. Most serious vendors can claim it, and buyers have gotten good at evaluating it.

The second layer is AI agents, software that does work rather than just displaying information. Checking a patient in, routing a case to the right resource, drafting documentation, monitoring a discharged patient, raising an alert when a reading drifts. A growing set of vendors now operates here too, and the second layer is where much of the category’s marketing energy currently sits.

The third layer is clinical resources: the physicians, advanced practice providers, nurses, dieticians, and case managers who actually deliver the care. Not a referral or a line on a slide, but part of how the result gets produced. “Layer three is where the field thins to almost nothing,” Patel says. “Plenty of companies will sell you layer one. A smaller group will sell you one and two. Almost no one operates all three as a single accountable unit, and that is exactly where the outcome lives. You cannot ship a lower readmission rate in a software update. Someone has to call the patient within forty-eight hours of discharge, and someone has to own whether that call happened and whether it worked.”

Why Layer Three Is Where Outcomes Move

The reason the third layer matters is structural, not sentimental. A generated clinical note is real work saved, but it is not a result a board reports on. The metrics leadership answers for are length of stay, readmissions, avoidable emergency visits, throughput, network retention, and clinical workforce capacity. None of those move because a piece of software was installed. They move when a specific clinical action happens reliably, at scale, for every patient who needs it, and when someone is answerable for the gap when it does not.

Consider transitional care after a hospital discharge. The technology can identify every discharged patient and open a task. An agent can place the first outreach call within minutes and escalate the ones that go unanswered. But the reduction in readmissions comes from a clinician actually reaching the patient, reconciling medications, catching the missed follow-up, and pulling the person back before a small problem becomes a return trip. When a company operates all three layers, outreach within forty-eight hours of discharge happens regardless of payer, across tens of thousands of encounters, and the readmission curve bends. When a company operates only the first two, the health system has a very efficient way to generate a task list, and the outcome still depends entirely on whether the customer’s already-stretched staff can work it.

Patel is precise about where the line sits, because the argument is easy to distort. “This is not the vendor doing everything while the customer does nothing,” he says. “Every real deployment runs on the customer’s own clinicians and operators, and that shared work is what partnership is. The failure is narrower. It is when the entire weight of the outcome lands on a stretched team while the vendor’s responsibility ends at go-live. When the number then fails to move, the vendor calls it an adoption problem and the customer calls it a vendor problem, and it turns out the outcome was no one’s job from the start.”

The Word That Hides the Gap

The word doing the most work to obscure this divide is platform. Nearly every vendor claims it. The website says platform, the deck says platform, the funding announcement says platform. Open the box and the contents are frequently a model, an interface, an integration, and an agent or two. That is a feature set, and the label does not change what is inside. The three-layer test is useful precisely because it cuts through the word. It does not ask what a vendor calls itself. It asks what the vendor brings across all three layers, and who is answerable when the result is due.

This also explains why the strongest version of the model does not depend on any single AI vendor’s software. The technology and the agents are built around generative AI from the ground up, and the frontier models underneath are interchangeable inputs, routed per task and swapped as better ones ship. The durable value is not the model. It is the workflow, the agents, and the clinical resources standing behind the result, wired together so the outcome is owned rather than handed off.

What This Means for a Buyer

For a health system or a state leader evaluating AI, the three-layer test reduces a crowded, jargon-heavy market to a few direct questions. What does this vendor actually bring across all three layers, rather than one or two? Once the system is live, who is accountable for the clinical outcome, and what does that accountability look like in practice? And what clinical capacity does the vendor contribute beyond software, so the weight of the result does not fall on the staff alone?

A vendor that answers all three with specifics is offering to own the outcome with the client. A vendor that cannot has, without meaning to, already said where the outcome will end up. “Adoption is not impact, and a generated note is not a result,” Patel says. “The companies that will matter in this next phase are the ones willing to stand behind the number, with technology, with agents, and with the clinical people who actually move it. That is a much shorter list than the platform claims suggest.”