Turning AI’s Promise into Performance in Health Systems | Post 2 of 3
A surgical implant is going to run short in eleven days.
Nobody in the health system knows this yet. The signal exists, scattered across four systems that do not speak to each other. Consumption has been running above forecast in the cardiac service line for three weeks. A supplier shipment left the distribution center two days late. The perioperative schedule for the week after next has six cases booked that depend on this specific device. Somewhere in a purchase order queue, a replenishment order sits unconfirmed.
Each of those four facts lives with a different owner. The materials manager sees consumption. The supply chain analyst sees the shipment exception, if she happens to open that report. The OR scheduler sees six cases and has no reason to think anything is wrong. Nobody sees the fifth fact, which is the one that matters: these four things together mean that in eleven days, a surgeon will scrub in for a case that cannot proceed.
What happens instead is that the shortage is discovered on the morning of surgery. The case is cancelled. A patient who fasted since midnight goes home. An OR block worth roughly seventy dollars a minute sits idle. The preauthorization has to be reworked. Someone in patient access makes an unpleasant phone call. And the cost of all of it lands in six different cost centers, which is precisely why no single leader ever builds a business case to fix it.
This is not a technology failure. Every one of those systems is working exactly as designed. It is a workflow failure, and it is the specific kind of failure that AI point solutions cannot touch, because the failure lives in the space between the tasks rather than inside any one of them.
The unit of AI transformation is the domain, not the task. Value in health system operations leaks at the handoffs between functions, which means automating individual tasks preserves the exact seams where the money is lost. Rewiring end to end is not the ambitious version of retrofitting. It is a different project entirely.
The Premise Worth Questioning
McKinsey’s June 2026 report on the health system CEO imperative makes a recommendation that sounds modest and is not. Rather than spreading investment across a smattering of use cases, pick one or two high-value domains and rebuild them end to end. Redefine what work is actually required. Rebuild the roles, the processes, and the technology around that answer rather than fitting AI into the process you inherited.
The premise this challenges is widespread and rarely stated out loud: that retrofitting AI into existing workflows is the pragmatic path, and end-to-end redesign is the aspirational version you get to later, once the pilots have earned some credibility.
That premise deserves a serious hearing before I take it apart, because in healthcare it is defended by people with good reasons.
The Case for Leaving the Workflow Alone
Start with the obvious. Health system workflows are not arbitrary. They encode decades of accumulated response to real events. That redundant verification step exists because of an audit finding in 2014. That handoff to a clinical reviewer exists because a payer disputed a claim and the organization lost. That form gets faxed because one referring practice never joined the HIE and represents eleven percent of volume.
Redesigning a workflow means removing fences without always knowing why they were built. In an industry where the downside includes patient harm, regulatory action, and cash flow disruption in an organization operating on a two percent margin, that caution is not timidity. It is judgment.
Second, retrofitting is genuinely faster. Dropping an AI capability into an existing process requires no reorganization, no retraining of a department, no renegotiation of who owns what. It can be scoped, funded, and delivered inside a single budget cycle by a single leader. End-to-end redesign requires cross-functional authority that, in most health systems, exists only at the level of the CEO.
Third, and this one is underrated: the organization is tired. Most health systems spent the last fifteen years absorbing an EHR implementation that consumed enormous capital, disrupted every clinical workflow simultaneously, and delivered benefits that took years to materialize and that many clinicians still dispute. Ask that organization to reimagine a domain end to end and you are asking people who have been through a transformation to volunteer for another one.
All of this is real. The case for incrementalism in healthcare is stronger than in almost any other industry.
And it still does not work, for a reason that has nothing to do with courage.
The Value Lives in the Seams
Go back to the implant.
Now imagine a health system that has done everything right by the standard of incremental AI adoption. It has deployed demand forecasting in supply chain. It has an AI scheduling optimizer in perioperative services. It has automated prior authorization checking in patient access. Three tools, three vendors, three successful pilots, three leaders with a win to report.
The case still gets cancelled.
The forecasting tool improved its forecast. The scheduling optimizer filled the block efficiently. The prior auth tool verified authorization for a case that will not happen. Each tool did its job correctly, and the outcome is identical to the outcome with no AI at all, because none of the three could see the other two.
This is the structural point, and it generalizes well beyond the operating room. The workflows that consume the most money in health systems are the ones with the most handoffs, and the handoff is exactly what task-level automation preserves. You can make every station on the line faster and change nothing about how much work falls on the floor between stations.
McKinsey’s own example makes the same case from the revenue cycle side. A generative AI tool that drafts appeal letters is useful. But the value of coordinated agents working the entire back end of revenue cycle management, from claim submission through cash posting and reconciliation, is a different order of magnitude, because most denials are not appeal problems. They are eligibility problems, authorization problems, and documentation problems that became appeal problems through neglect at three earlier handoffs.
Which is why the report suggests denial prevention should begin before the patient is even scheduled. Verify benefit coverage and authorization requirements at the point of scheduling. Trigger proactive outreach when something is missing. Retrieve the required documents automatically rather than hoping someone chases them. Then make clean-claim rate a North Star metric alongside days in accounts receivable, because clean-claim rate is the number that measures the whole chain rather than any station in it.
Notice what changed there. The work did not get automated. The work got redefined. That is what end to end actually means, and it is why it cannot be reached by adding tools to the process you have.
Why Revenue Cycle and Supply Chain Come First
McKinsey nominates two starting domains, and the logic is worth stating plainly because it is the part most likely to survive contact with a risk committee.
The back end of revenue cycle and the supply chain share three properties. They are high volume, which means small per-transaction improvements compound into material numbers. They are dense with handoffs, which means the seams are where the value is. And they carry limited direct clinical risk, which means an agent making a mistake produces a rework queue rather than a patient safety event.
That last property is what makes them the right first domain rather than merely a profitable one. Denials management, underpayment management, accounts receivable follow-up, and cash posting are administrative and rule governed. They are the workflows where autonomy is most likely to work and least likely to hurt anyone. Care access and clinical workforce management carry more value, but they should be the second or third domain, not the first.
The measurement discipline matters as much as the domain choice. Initial denial rate, denial write-off rate, and days in A/R signal financial health early, which lets a health system decide what to scale next based on evidence rather than enthusiasm. Humans stay in the loop for exceptions, for model training, and for compliance. This is not a story about removing people from the revenue cycle. It is a story about what those people are doing with their attention.
What Coordination Actually Requires
Now return to the implant shortage, and consider what it would take for that eleven-day warning to actually surface.
A supply agent notices consumption running ahead of forecast and reconciles it against open purchase orders, delivery confirmations, and demand signals drawn from the clinical schedule. It concludes there is a shortage coming. It does not stop there. It triggers a perioperative scheduling agent, which identifies the six affected cases and coordinates rescheduling with the surgeons. That agent in turn notifies revenue cycle agents, which adjust the preauthorizations and the patient communications so that the moved cases do not generate a second wave of denials and confused patients.
That sequence is not one AI doing one thing well. It is three agents in different domains handing work to each other, each with access to a shared understanding of what the business is, each operating inside policy boundaries someone defined in advance.
At Datafi, this is the specific problem Runtime exists to solve. Deploying a single agent is straightforward and most vendors can do it. Coordinating many agents across a multi-step process that spans functions is a different engineering problem, and it is the one that determines whether a health system ends up with a collection of tools or an operating model.
Underneath it, the global business contextual layer holds the live representation that makes coordination possible: that this implant maps to these cases, that these cases carry these authorizations, that this supplier has this delivery history. It federates to your source systems in place. Your EHR, your ERP, your materials management system, and your claims platform stay exactly where they are, running exactly as they run today. Nothing migrates. This matters more in healthcare than anywhere else, because the answer to “we should rewire this domain” cannot be “so first replace Epic.”
Studio addresses a subtler problem in McKinsey’s recommendation. The report calls for cross-functional pods led by business leaders as product owners, on the argument that the person who understands the process nuances can learn the technology more easily than the reverse. I agree, and I would add a condition: a product owner who has to open an engineering ticket for every workflow change is not actually empowered, whatever the org chart says. Giving business owners the ability to build and modify agents and workflows without code is what turns the pod from a governance structure into a delivery capability.
Control Tower and Cyber hold the boundaries, and the measurement layer, which is where this series goes next.
One Domain Is Not a Big Bang
The objection I expect from any health system operator reading this is that end-to-end redesign sounds like a multi-year enterprise program, and those have a poor track record here.
It is the opposite. The recommendation is explicitly to narrow, not to widen. One domain. Not the enterprise, not every workflow, not a five-year roadmap. Pick the back end of revenue cycle, rebuild it around what the work actually requires now, measure clean-claim rate and days in A/R, and use what you learn to decide whether the next domain is supply chain or something else.
The scope is smaller than the portfolio of pilots most health systems are already running. What changes is that the work compounds, because the second domain inherits the context layer, the governance model, and the operating pattern that the first one built.
That is the difference between fifty experiments and an operating model. Not ambition. Architecture.
In the final post of this series, I take up McKinsey’s most counterintuitive recommendation: that health systems should stop waiting for the perfect data foundation, but should not wait on governance. Those two things get sequenced backward almost everywhere, and getting the order right is what makes iterative building safe in an environment where a governance failure is a front page event.
Datafi is the operating system for business AI. We help mid-enterprise organizations unify their data, deploy governed AI agents, and deliver measurable outcomes in weeks rather than years, without migrating off the systems that already run the business.
Turning AI’s Promise into Performance in Health Systems
Part 1: Health Systems Aren’t Short on AI. They’re Short on Impact.
Part 2: Rewire the Domain, Not the Task
Part 3: Guardrails First, Foundations As You Go
Written by Vaughan Emery, Co-Founder & Chief Product Officer at Datafi
Source: “The health system CEO imperative: Turning AI’s promise into performance,” McKinsey & Company Healthcare Practice, June 2026.
