AI strategy v. AI in strategy: Health system success hinges on a two-letter word

Sara: In our last episode, we discussed how your AI vision needs to be an intentional strategy. In today’s episode, we examine what this means and how it manifests itself in the market today.  

Some organizations are looking for the “right” AI strategy, but there’s no standard AI strategy that applies to all health systems. Just like every health system needs to tailor its enterprise strategy to be successful as a business and in its mission, its market, and its unique competitive environment, it also needs to tailor its AI strategy. 

Tom: When a new capability comes up and has all this transformational hype around it, the challenge is that a lot of health systems are asking, “What’s our strategy for this?”  

But this thinking creates an artificial separation. Instead of thinking about the role AI plays in every dimension of their business, it creates this box that they’re going to plan around.  

What I’ve seen in some health systems is this “me too” strategy, where they say, “We’re going to do all these AI things other people are doing.” This produces an undifferentiated position, which intrinsically invalidates the entire point of strategy. You also get disconnected projects: AI-based projects and business projects. 

We see this a lot with information technology (IT) departments. Because AI sits in the IT budget, you get this pressure to shrink AI investment when the organization faces budget constraints. You get fragmented governance, separating AI governance, IT governance, and strategic governance.  

At the end of the day, you get strategic dissonance. There’s a structural flaw in thinking of AI strategy as its own distinct thing. It’s incoherent to say, “This is what we’re going to do with AI, and this is what we’re going to do in our business. Now, how do those two pieces relate?”  

Sara: In a way, AI becomes a solution looking for a problem rather than solving inherent problems in the business—problems not just in the technology but in the organization and in how it’s meeting its internal needs as well as those of  its community and market.  

If you’re thinking about it the way you just described, Tom—as this artificial separation—you’ve already lost because you’re taking a technology-first, vendor-centric approach rather than a problem-centric approach.  

That is a key pivot that organizations need to make: Become problem-centric and think about AI through the lens of the problem you’re trying to solve. Then you can think about how the technology can meet those needs. 

Tom: I try to capture how executives should think about this with a simple change in the mnemonics: It’s not “the AI strategy,” it’s “AI in strategy.” It may seem minor, but it changes the orientation. The reality is that AI impacts every dimension of strategy. And if it doesn’t, you’re thinking about it the wrong way. 

Even if you took the most generic strategy, such as the Quadruple Aim, every element of cost, experience, quality, and the care team’s experience would be impacted by AI. When you reframe it that way, your AI strategy is just, “Where does AI impact your strategy writ large?” 

Then you get into a chicken-and-egg question: Do you take the strategy you’ve built and say, “Let’s think about how we can use AI to do all these things?” Or you could say, “How do we use AI in our service line strategy?” That could have direct, meaningful benefits.  

But the flip side is that AI has such transformational potential in some domains that it could actually change what your strategy is.  

So you need a structural view: “What is our posture? What role do we play in our market? What assets do we have?” That’s the starting point; those are the foundational elements of the strategy.  

Then you think about how you’re going to invest in this and change the organization, its posture, and its portfolio to differentiate it in the future. Ultimately, you need to think, “What capabilities do we need?” Then across this entirely new domain of AI, you need to be asking, “How do we establish trust: Trust in the process, in the outcome, in the role it plays?”  

AI is a technical component, but it’s also a business process component, so it inculcates every dimension of strategy. In every part of strategy, we should be asking, “Is there an AI application?” And the answer, almost invariably, will be yes. 

Sara: What you just described is that by starting with AI strategy instead of starting with what you’re trying to accomplish, you naturally get constrained.  

Let’s take the growth example. Independent of the technology, really focus on the strategy you’re trying to accomplish. For instance, “We want to reach every individual in our community in a personalized way.” Independent of the technology itself, that is a vision; that is a strategy that you can then translate into how you will accomplish certain things that human beings can’t do or that would be very expensive or take them a really long time to do.  

So you might say, “We are going to deliver n-of-1 personalized care or outreach to every human being in our market. Through that, we could pinpoint individuals who need specific services from us and convert them into our health system.”  

And then you can apply AI and say, “What are the components that I need? I need an identity component so I can know people. I need a whole life-cycle and tailored performance marketing component so I can drive specific campaigns to individuals. And all of this would be powered by compute and AI in the background.” 

But you don’t start your strategy with the AI.  

Tom: We’ve talked previously about how the costs of cognition, engaging with patients, and responding to patients can all decrease with AI capabilities. You could think about your strategy in a legacy way: “How do we do what we’ve always done to run a hospital-centric business and make it easy for patients to get an inpatient procedure?”  

But what you’ve just described is a different future healthcare economic model: “How do we think about engaging consumers? How do we get to a different spot in our consumers’ decision-making process, in their disease exploration, and in prompting what they need?” When you change the underlying math of that, you can create a new business model. And I think that’s the power AI opens up.  

It’s not that we can cut waste margins out of the things we’ve always been doing. It’s not that primary care is going to be 5% more profitable. Thinking about it as, “We’re going to squeeze a little bit more and do a little bit better on access, getting from a 5-week wait time to a 4-week wait time,” is the wrong way to view it. 

But when you take a holistic view, all of the cost structures of healthcare can change. All the operating models can change.  

Health systems can’t take the traditional view of making money on commercially paid services and losing money on everything else. The macro hydraulics are not working in our favor. There’s less and less commercial population in most markets. There’s more and more Medicare, Medicaid, and uninsured. We have to think differently about that entire structure. 

Imagine one context chain: If you could say to any health system in America that their primary care business would break even tomorrow, the entire economics and strategy of what they do would be different. You’re not going to sit here and ask, “Do we hire another primary care doctor?” You would do it because it allows you to own the population, grow your reach, drive growth. 

That’s one example. And it’s not unfathomable to say we don’t have to lose money on primary care in a more digitally and AI-mediated operating model. But the same is true of all our businesses.  

Increasingly, if you visualize a matrix showing all the sites of care of a traditional health system and the different financial payer classes of your consumer patient populations, that board is increasingly turning red. And it’s increasingly turning darker and darker shades of red. And the pockets of green are like a lake with water that’s evaporating faster than it’s raining in. The lake is getting smaller and smaller and shallower and shallower.  

If all we’re doing is focusing on how we retain this one little puddle, we’re missing the point that we can change the overall context of that lake.  

AI can change the economics so that we don’t lose as much money on every patient. Ultimately, the silver bullet is answering the question, “How do we craft a solution that knows what a patient needs and effectively delivers it in the most cost-effective, timely, consumer-centric way?” Which is easy to say but is very complicated to accomplish. 

Sara: It’s very complicated to do because it’s not just how many minutes we need for an in-person visit and how we can squeeze it down. And it’s not just bringing virtual, real-time visits into the mix. 

Those are two potential levers and degrees of freedom, but there are many others. Like, what happens when you start to bring in asynchronous care that is informed by an intelligence layer with information about the patient’s intent, clinical need, and relevant demographic factors? Then you might apply these things to a bot-driven framework.  

It gets complicated quickly, and it requires a combination of your strategy and the operational and financial nuances that enable something like this. The incentives, the clinical reasons, and many other things need to be orchestrated relative to one another. 

Tom: A key issue that every health system faces is, “How do we know what each patient needs?”  

A patient can come through any channel: By phone, in person, through a digital platform, or online. We generally know something about them. As soon as we identify them, we know their medical record and history. Just off that, when a patient comes in—even when we don’t know why—we can predict the likelihood that they will need to be admitted. Or, as we’ve done in some of our hospital at home work, we can predict the likelihood that they will be eligible for care at home as their site of care.  

The potential of how we understand what a patient needs is a huge lever. And at the end of the day, we can get better insight based on what we know about their past, their personality, their preferences, and their perceptions. For instance, my mom’s use of healthcare is very different from mine. She engages in a much different way than I do, and understanding that is part of defining her need.  

We’ve done some really cool work with clients that use AI to identify and think about what patients actually need. Now, if you let AI do this on its own, using AI as a medical device, there is a whole set of approval recommendations and Food and Drug Administration (FDA) oversight.  

But if you use AI to inform the creation of rules-based deterministic models that are overseen by a clinician, that’s really compelling: You can establish a deterministic spine that routes care based on what patients need. 

The AI can ask the patient questions that explore whether other information should be added to the deterministic rule base. And what you actually get is something that looks like AI and feels like AI to the consumer—it feels empathetic, thoughtful, and customized.  

But behind it, there’s a rules engine that the AI leans on and validates to make sure it does not send a patient off the rails by hallucinating or sending them in the wrong direction. No, the rules dictate this.  

When health systems start to realize they can get a different level of granularity, they can have a high degree of confidence in knowing what patients need. 

Think of the AI as a Plinko board of supply options. It’s a permutation of a care team member: Is it a super-super-specialized physician? Is it a more junior physician? Is it an advanced practice provider? Is it a nurse? Is it a tech? Is it a medical assistant? Is it just an administrative resource? Is it a bot that is going to be initiating an action? And then think about the setting it’s going to operate through.  

So you can click on the care team member, you can click on a setting, and you can calculate a cost for that interaction. And now you’ve got a robust set of cost platforms and cost profiles with different clinical capabilities. You’re not going to expect the same level of capability from a bot over a chat interface as from a physician in an in-person interaction. They’re different clinical capabilities.  

But AI triage based on a deterministic backbone can appropriately route patients, giving them options and saying, “This is what we recommend.” You’re getting better engagement with the patient. You’re giving them more timely information, and it’s likely at a lower cost. But again, that requires a fundamental shakeup of how we think about the economics of different sites of care. 

Sara: I want to connect some points that you made. One is this deterministic backbone, which is critical for providing guardrails and constraints for AI. Connecting that to your trust point from earlier, every organization will have a different tolerance, a different risk posture, a different set of parameters that define and quantify trust in a system for them.  

When exploring these new care models and operating models, those trust factors must be reflected in the solutions—whether they are clinician-facing or patient-facing.  

In a past role, my organization at the time built a solution that allowed patients to ask questions to get a degree of screening and triaging. We just put guardrails around the large language models but did not include the deterministic backbone. And there were some instances, like two out of millions of conversations, where the AI—which was non-deterministic (meaning we couldn’t predict the output based on the inputs)—went outside of those guardrails. 

I had an experience at a previous organization, for example, where we had built a solution where patients could come in and ask all sorts of questions and in a way do what you're describing—get a degree of screening and triaging as appropriate. But we learned we had put guardrails around the LLMs without the deterministic backbone. 

One example was a patient who was traveling and came in and asked what time she should take her birth control in this new time zone. And, technically, that is a piece of clinical advice and is a boundary that shouldn’t have been violated. But the AI actually did answer. Based on our risk posture, we decided to eliminate all non-deterministic answers. 

All of these things are really critical to factor into a trust platform because we actually wouldn’t have known that had happened if we didn’t have the underlying platform, the audit logs, and the mechanism to understand all interactions and to be able to go back and course correct. So that’s a really critical point. There’s a depth and nuance associated with AI, with how many different ways it could manifest. So you need a really strong foundation of trust. 

Tom: If any health system looks at the top two layers of their strategic plan, every single dimension should have something they can deploy AI substantively on. For most, this exercise alone will cause them to rethink their strategy. 

For instance, if a goal is to take the subsidy for primary care down by 10% and you apply an AI lens to it, that’s an easy target. What if instead we said an AI-based primary care model could eliminate the subsidy entirely? That changes the cascading effect of the entire business. 

That’s a tangible takeaway everyone can do: Look at your strategy today as an executive team. Do that with a facilitator who understands the art of the possible with AI, and you can get some really powerful insights to help determine which applications will be the most impactful to prioritize.