For Healthcare Leaders, Build vs. Buy Determines ROI on Enterprise AI

The pressure is on to deliver ROI in healthcare AI. In a world where up to 95% of AI projects reportedly fail, the deep-seated instinct to build custom, in-house tools no longer serves the industry. Between shrinking margins, fierce competition, and a lack of talent, most healthcare leaders are already confronting a hard truth about their AI strategy.

Matthew Crowson, MD, brings a unique blend of clinical and technical experience to the table. As Director of AI/GenAI Product Management at Wolters Kluwer, a practicing ENT surgeon, and a faculty member at Harvard Medical School, he operates at the intersection of technology and patient care. Today, he credits his medical training for shaping his direct, data-driven approach, as evidenced in more than 90 peer-reviewed papers and patents. When asked about the build versus buy problem, Crowson cut straight to the core.

Turning to economics for clarity, Crowson draws inspiration from an almost century-old theory that continues to influence how organizations decide what to build and what to buy. The framework, first proposed by economist Ronald Coase in 1937, offers a straightforward method for determining whether developing an AI solution in-house is the most effective approach or if partnering with a vendor is a better option instead.

• Build vs. buy cheat sheet: "It starts with a few simple questions. How often are your requirements likely to change? What level of integration does it need? Is it mining sensitive, patient-level data, or does it operate more like a bolt-on service? And how often is it used? Something you check quarterly can be managed very differently from a tool that needs daily, hands-on support," says Crowson.
• A textbook buy: Crowson points to ambient dictation as an example, a technology that has become widely accepted in US hospitals. Its requirements tend to be stable, and it typically doesn't require deep integration, he explains. That makes it a prime candidate for a "buy" decision under this framework. "It's one of the most accepted AI systems in hospitals now. How often do the requirements change on that? Not very frequently. It's a transcription service."

The choice, however, isn't just a strict binary. For Crowson, the most promising paths forward lie in a nuanced gray area, where partnerships evolve the choice beyond simply building or buying. Rather than single, hard-coded solutions, he forecasts a future filled with customizable platforms. "You're buying the platform, and it's more of a two-way transfer of knowledge and data," he explains.

By allowing health systems to grant secure, API-based access to a vendor's platform, the hybrid model is designed to address the "trust moat," he says. That arrangement is what enables them to customize models with the assurance that sensitive data remains behind their own firewall.

• The data stays put: It's a technical solution to a trust problem that already has a powerful real-world precedent, Crowson explains. "The main risk comes when data leaves the barn, so to speak. That concern is eased when it stays behind the firewall. In those cases, what’s shared isn’t the raw data but secure artifacts like model weights. There’s already a clear precedent for this with Epic’s API access, which powers wearable integrations and Apple Health. It shows that innovation can happen without patient data ever leaving the system."
• Headline risk:  The risk of a data breach can be so severe that it creates an intense environment of risk aversion, Crowson explains. "In a heavily regulated industry like health, one disclosure of PHI can land you on the front page of the New York Times for a data breach. So they are incredibly risk-averse, and there's a massive trust moat that prevents a health system from trusting a vendor."

So what's the fix? Crowson lays out a pragmatic starting guide for any healthcare leader looking to cut through the vendor noise and ground their AI strategy in reality.

• Diagnosis first: The first rule for any AI initiative is to resist the urge to start with technology, Crowson says. "You shouldn't be a hammer looking for nails. Instead, start with old-fashioned problem-solving and ask what you are fundamentally trying to solve. What is the organization's most significant pain point right now?"
• Talent check: Organizations need to be realistic regarding their ability to manage solutions in-house, he continues. "Even the brand-name health systems are talking single-digit margins right now," Crowson explains. "They don't have the money to be recruiting top-end AI talent, and they just can't compete for it. The same companies competing for that talent are the familiar names: Meta, Google, Microsoft, Amazon. If you don't have the talent to maintain or build it in-house, it really becomes a simple buy decision."
• Data before deals: No AI strategy can succeed without a solid data foundation, Crowson advises. "Before you can even engage with these vendors and platforms, your own data house has to be in order. We still haven't figured out how to leapfrog poor-quality data or databases."

For healthcare leaders, Crowson's final message is clear: the cultural habits that have long defined academic medicine may no longer be sustainable. "There’s a bit of bravado in how healthcare tends to approach problems, driven by a belief that we can design the best solutions ourselves. It’s a natural reflection of who’s in these organizations: academics used to leading their own research and not necessarily working closely with industry." But achieving successful innovation often depends more on embracing pragmatic partnership than an instinct for bespoke bravado, he cautions. "You don't want to bet the firm on an unproven, high-risk AI automation project."