The Obstacle Is the Way: What Healthcare AI Can Learn from the Stoics
kdavis
Thu, 05/14/2026 - 12:20
On the flight home from HMPS, I was listening to Ryan Holiday's Daily Stoic podcast. He was riffing on the central idea of his book The Obstacle Is the Way, drawn from Marcus Aurelius: "The impediment to action advances action. What stands in the way becomes the way."
It's a line that gets quoted as motivational shorthand, but the deeper Stoic discipline underneath it is more useful—and more demanding. Epictetus called it the dichotomy of control: focus relentlessly on what is yours to shape, and stop spending energy on what isn't. The obstacle isn't a problem to get past so the real work can begin. The obstacle, properly understood, is the real work—because it sits in the small territory of things you actually control.
That idea kept surfacing as I reflected on the conference. Because there is a version of this playing out right now across every health system trying to do something meaningful with AI.
What Marketing and Digital Leaders Don't Control—And Won't, Anytime Soon
Sit in on any health system's digital leadership meeting and you will hear a long list of things the team is waiting on. Epic to release the next set of patient-facing capabilities on its roadmap. Central IT to free up resources from the clinical project queue. Legal and compliance to finish their review of the new use case. Leadership to publish a formal AI policy so the team can stop running shadow experiments. Capital budget cycles to come around again. Agency and vendor partners to ship what was promised last quarter.
None of those concerns are unreasonable. Every one of them is real. But none of them are within a marketing or digital leader's control. Waiting on them is, in the strict Stoic sense, a category error—energy spent on the wrong domain.
Meanwhile, the one thing that is within your control the structure, quality, and governance of the data that powers your digital experience)is the thing most often deferred. It's treated as a precondition, scoped into a future phase, handed to a team without the authority to fix it across silos.
That framing is exactly backwards.
What HMPS Kept Telling Us
AI dominates the conversation at every healthcare event now, and HMPS was no exception. But the most instructive moments were when speakers stopped treating data as an obstacle in front of the work and started treating it as the work itself.
The scaling problem is a data problem. The "AI in Action: Product Strategy Lessons" panel cited an MIT report finding that 95% of in-house AI projects fail to deliver ROI or scale beyond pilot phase. The common thread in the failures was rarely the technology—it was AI deployed on top of unstructured, inconsistent, ungoverned data. You cannot build intelligent systems on a foundation that isn't.
The governance problem is a data problem. In "Beyond the Digital Front Door: Governing Patient-Facing AI in a Post-Website World," the discussion made clear that most large health systems have AI governance bodies, but they are heavily IT-focused and often constrained by primary vendor alliances. The deeper issue: fragmented data across systems prevents effective AI implementation regardless of how good the governance structure looks on paper. Governance without clean data is process theater.
The AI search problem is a data problem. "AI-Ready Patient Search: Beyond SEO" made the GEO point bluntly: as search shifts from Google to AI-driven tools, the health systems that surface in results will be the ones with structured, consistent, machine-readable provider data. AI tools cannot recommend a provider they cannot accurately read.
I'd offer one honest qualification: not every AI challenge in healthcare is a data problem. Some are genuine model-selection problems, change-management problems, or clinical-workflow integration problems. But strip those away and what remains, underneath nearly every stalled initiative, is a foundation that was never built to support the weight being placed on it.
Beneath the Surface: What Patients See vs. What AI Sees
Here is the frame that has changed how I talk about this work.
For 20 years, healthcare digital teams have optimized the presentation layer—what patients see. The website. The Find-a-Doctor experience. The portal. The mobile app. The brand. The campaigns. That work matters and it isn't going away.
But there is now a second layer that matters just as much, and most health systems have never optimized it: the data architecture layer—what AI sees. Generative search engines, agentic care navigators, voice assistants, and AI copilots don't read your website the way patients do. They read your structured data. Your provider taxonomy. Your location records. Your specialty codes. Your real-time availability feeds. Your listings across 70-plus healthcare publishers.
If that layer is fragmented or inconsistent, AI will surface the wrong provider, recommend the wrong location, or skip your system entirely in favor of a competitor whose data is cleaner. That outcome has nothing to do with how good your AI strategy deck is. It has everything to do with what AI can actually see when it looks at you.
The Real Failure Mode
The standard objection I hear from CMOs and digital leaders is: "We're not waiting—we're running pilots in parallel while we work on data." That's the polite version of what's actually happening at most health systems. Pilots on top of messy data, AI work and data work running on separate tracks, with the data track perpetually under-resourced because it doesn't have a demo at the end of it.
That's the trap. The two tracks never converge, the pilots never scale, and three years later there's a graveyard of proofs of concept and a data estate that's still fragmented—right as the AI capabilities you were waiting on finally arrive and you can't deploy them.
The Stoic move is to collapse the tracks. Stop treating data as preparation for AI and start treating AI use cases as the forcing function that determines which data gets cleaned, structured, and governed first—and to what standard.
This is the operating model shift. Not "fix the data, then do AI." Not "do AI in parallel and hope the data catches up." Instead: pick the AI use cases that matter most (Find-a-Doctor, scheduling, agentic search, clinical decision support) and let those use cases drive the data work in priority order. The use case sets the bar. The data work clears it. The AI capability ships on a foundation that was built to hold it.
In our work at Phase2, we've found four data requirements that consistently determine whether AI will actually work for a health system: structured provider data with consistent formatting across every profile, location intelligence normalized across every listing and publisher, search optimization built on a consistent taxonomy for specialties and conditions, and real-time accuracy automated across every source system. Get those right and most AI use cases become tractable. Skip them and no model in the world will save the pilot.
What This Looks Like in Practice
We saw this play out with Tufts Medicine and Yext—work I had the privilege of presenting at HMPS this week alongside Tufts and the Yext team.
Tufts had the classic obstacle: 3,500 physicians across 10 websites on 5 CMS platforms, data pulled from five separate sources, a fractured patient experience. Rather than layering AI on top of a broken foundation, we went straight at the obstacle—pulling data into a unified warehouse, verifying it across all 17 departments and four entities, achieving 94% consistency across 3,137 provider records pre-launch.
The shift in goal was telling. The original goal was a deliverable: launch the brand, consolidate to one platform. The new goal was a discipline: drive access, drive visibility, optimize the site, reposition the brand—continuously. That's the move from project to operating model.
The use cases drove the data work. Find-a-Doctor needed accurate specialty taxonomy, so the taxonomy got fixed. Scheduling needed real-time availability, so the integrations got built. AI-driven search needed structured, machine-readable provider data, so that's what got prioritized.
For context on scope: this was a multi-year engagement spanning brand strategy, data architecture, platform consolidation, and ongoing optimization—not a one-time project. The work continues today, which is the point. The shift from project to operating model means the investment compounds rather than ends.
The results, year over year:
- 711 % increase in booked appointments (modeled at approximately $8.14M in additional patient revenue)
- 183 % increase in Find-a-Doctor visits
- 65 % increase in search sessions year over year
- 53 % click-through rate vs. a 25 percent industry benchmark, with 92.5 percent of searches surfacing a relevant doctor against and 85 % industry standard
- 88,642 listing updates across 1,600-plus listings on 70-plus healthcare publishers, representing approximately $188K in operational savings
And then, the part that matters most for what comes next. Because the data foundation was built right, Tufts is now the first healthcare system to monitor brand presence and consumer sentiment across AI-driven search platforms at a hyper-local level—running at 47% positive and 50% neutral AI brand sentiment, with a 112% increase in search engine impressions year over year.
That's not a Find-a-Doctor outcome. That's an AI-readiness outcome. The structured data layer they built isn't just powering today's experience. It's the infrastructure for whatever AI requires next—generative engine optimization, agentic care navigation, AI-driven scheduling, the patient-facing workflows that don't yet have names.
The Window Is Narrowing Faster Than People Think
This work can't wait, and it's the part of the conversation most health systems haven't caught up to yet.
AI is developing memory.
Not just the consumer feature you've seen rolled out in ChatGPT, Claude, and Gemini—though that's part of it. The deeper phenomenon is path dependence in how AI systems learn what to recommend. When a generative search engine answers "who's the best orthopedic surgeon near me?" and surfaces a particular health system, that answer trains two things at once: the patient's mental model of where to seek care, and—through real-time retrieval and feedback loops—the system's own pattern of which sources to trust on the next query. Every query that resolves to a particular health system reinforces that system as the answer to the next query. The defaults harden.
The health systems that are visible to AI today are building compounding visibility. The ones that aren't are accumulating an invisibility deficit. And the gap widens every quarter.
This is a meaningful break from how digital trends used to work. For most of the last two decades, if you were late to a digital channel—SEO, paid search, social, content marketing, even programmatic—you could spend your way back. You hired an agency, allocated budget, executed a 6-to-12-month catch-up plan, and you were back in the game. The economics of catching up favored money.
AI doesn't work that way. The data infrastructure required to be discoverable and recommendable by AI—structured provider data, normalized listings, consistent taxonomy, real-time accuracy across every source system, integration with every relevant publisher—takes 18 to 24 months to build well, even with a competent partner and a committed leadership team. While you're building it, the systems that already have it are training the models and the patients to default to them.
A skeptical CTO might push back: "Models retrain. Knowledge cutoffs roll forward. We'll be in the next training run." That's partially true and entirely insufficient. Three reasons:
First, real-time RAG and agentic retrieval are increasingly the dominant pattern. Those systems pull from live structured data, not just training corpora. Health systems with clean structured data win those queries continuously, in real time, regardless of training cycles.
Second, even when models retrain, the patterns of trust they encode—which sources they cite, which providers they surface—get reinforced by the corpus they were trained on. Late entrants face a steeper signal-to-noise problem than first movers did.
Third, and most importantly, the patient query layer is changing in real time. Patients are forming habits now about which AI tools they trust to find care and which answers feel authoritative. Health systems that show up well in those queries today are building durable mindshare. Those that don't are accumulating an invisibility deficit that compounds.
You don't have to believe AI search will fully replace traditional search to take this seriously. You only have to believe that the systems discoverable by AI in 2026 will be the defaults of 2028. That's a claim the data already supports.
The Stoic discipline cuts both ways here. Focusing on what you control is liberating—but it also removes your last excuse. You can't wait for the AI roadmap to settle. You can't wait for the budget cycle. You can't wait for vendor consensus. Waiting isn't neutral. It's a decision that gets more expensive every month, and the hill gets steeper, not flatter, the longer you wait.
What This Means for Your Organization
If you're a CMO or chief digital officer at a health system, you already know the version of this happening inside your walls. The AI strategy deck looks great. The pilots are running. The roadmap is approved.
And underneath it all, the real fights are about something else entirely. Budget constraints that force you to consolidate where possible and prioritize only the highest-impact investments. Scheduling complexity from inconsistent open-scheduling practices across departments. Evolving brand positioning that requires constant re-alignment with clinical and operational stakeholders.
But the one that matters most for AI readiness, and the one that gets the least executive attention: vendor selection. Most health systems have inherited a martech stack that's a federation of point solutions—a CMS that doesn't talk to the provider data system, a scheduling platform that doesn't sync to the listings tool, a CRM that lives in its own world, a Find-a-Doctor experience built on top of an EHR data feed that updates on a delay. Each of those vendors has a roadmap. None of those roadmaps are coordinated. Building AI-ready data infrastructure on top of that landscape isn't a procurement question—it's an architecture question. And it's the question almost no one is asking out loud.
Underneath all of it is the quiet anxiety that when it's time to scale AI, the same fragmentation that broke Find-a-Doctor and the patient portal will break the AI work too.
You're right to be anxious. And you're spending energy on the wrong things—vendor roadmaps, model benchmarks, regulatory timelines—when the territory you actually control is the one you keep deferring, even as the window to do something about it narrows.
The question isn't when will our data be ready for AI? The question is which AI use cases are important enough to force the discipline of getting our data right—and what does our system look like beneath the surface, in the layer AI actually sees?
That's the Stoic move. Focus on what's yours to shape. Let the work you control define the work that follows.
The obstacle isn't in the way. It is the way through.
Phase2 helps health systems do this work—turning fragmented provider data, scheduling, and digital experience into the foundation that makes AI actually deliver. If the parallel-tracks problem sounds familiar, or if you want to talk through what your data architecture layer looks like to AI today, I'd welcome the conversation.
Publication Date
Thu, 05/14/2026 - 11:43
Marshall Schoenthal
Healthcare Industry Principal
Featured Blog Post?
Yes
Has this blog post been deprecated?
No
Summary
Most healthcare AI projects aren't failing because of bad technology. They're failing because the underlying data is a mess, and health systems keep waiting on things they can't control, like vendor timelines and budget approvals, instead of fixing the data problems they actually can fix. Tufts Medicine proved what's possible when you tackle the data foundation first, and the results speak for themselves.
Topic
Artificial Intelligence
Blog Post Graphic Banner The Obstacle Is the Way.png
Promo Image