From point solution to platform: Rethinking AI infrastructure in healthcare
August 31, 2026
Every hospital system in America is now running an AI experiment. Almost none of them are finished. That single, uncomfortable fact, which surfaced in Teladoc Health and Becker's Hospital Review's ninth annual survey of 167 U.S. health systems, is the real story of where virtual care and clinical AI stand in 2026. It isn't a story about hesitation. It's a story about architecture.
The data makes the point starkly. Ambient documentation adoption jumped from 20% to a projected 70% in a single year. AI-enabled clinical decision support moved from 26% to 78% over the same period, a 50-point swing that took telehealth itself nearly a decade, plus a pandemic, to achieve. And yet no single AI use case has gone live in more than 37% of the organizations surveyed. Adoption is universal. Completion is rare.
That gap has a name now: the architecture gap. It's the space between an enthusiastic pilot and an enterprise capability, and closing it is turning out to be the defining challenge for health system technology leaders and for the vendors who partner with them.
Everyone's In. Nobody's Done.
Teladoc's Tim Wright described the paradox during a recent webinar with Danny Sanchez, Teladoc's VP of Technology and Innovation, and Suki CTO Joe Chang: "everyone's in, but nobody's done." The distance between those two states isn't a failure of appetite. It's a failure of infrastructure to keep pace with ambition.
Three in five surveyed health systems are already making, or planning to make, major changes to their telehealth platforms and devices, not because their original programs failed, but because the assumptions underneath them didn't anticipate an AI-native future. Point solutions, deployed for a single service line, create what Sanchez calls a "patchwork", which is technically functional, but organizationally unsustainable. Every new use case spins up its own data silo, its own support burden, its own integration debt.
The fix isn't slower adoption of new technology. It's smarter capture from the beginning.
One Signal, Many Applications
The biggest idea to come out of the conversation is deceptively simple: capture a signal once, and let many applications consume it. A camera that detects a patient trying to leave a bed shouldn't exist solely to power a fall-prevention app. That same signal can trigger a virtual sitter alert, inform the nursing workflow, and feed a broader patient-risk model — if, and only if, the underlying architecture was built to share it.
This reframes the buying decision hospitals are making right now. The right question when evaluating new hardware isn't "does this solve my current use case?" It's how many future use cases the device can support without being ripped out and replaced. Sanchez put it plainly: "You should design for optionality, standardize on open interfaces... so you can change models, vendors or workflows and not have to replace the entire stack when you have new applications."
That principle extends to where computation happens. Latency-sensitive, safety-critical signals — a fall event, a patient safety alert — belong at the edge, close to the patient. Longitudinal pattern-matching and enterprise-wide intelligence belong in the cloud, where larger models can reason across populations rather than moments. As Sanchez summarized the hybrid model in one memorable line: "if possible, you should react locally but reason globally."
Validation Isn't a Milestone. It's a Discipline.
Perhaps the most consequential — and least discussed — theme in the conversation was model governance. A system a hospital's committee approved in March may be running on entirely different underlying models by September. That's not a hypothetical; it's the operating reality of any AI vendor iterating at the pace the market now demands.
Suki's Joe Chang described what that looks like from inside a vendor's engineering organization: automated and manual evaluation pipelines, adversarial models acting as judges on new outputs, and a clinical-informatics-built framework tracking hallucination rates, omissions, and workflow fit before any new model or prompt set ships. Post-release, the signal comes from usage — five-star in-app ratings normalized against specific model and retrieval strategies, reviewed weekly to catch regressions the pre-release testing might have missed.
For health systems, this reframes governance from a checkpoint into a standing function. Sanchez was direct about what that requires contractually: hospitals should demand "notification of material model changes, the right to defer an update... and the ability to rerun their own acceptance testing before you deploy or redeploy a revised model." In other words, the vendor relationship has to be built for continuous change, not a one-time procurement decision.
Encouragingly, none of this requires a standing army. A five-hospital system doesn't need a new department — it needs, per Sanchez, "a small three-to-five person AI enablement team" spanning clinical informatics, IT security, and quality, with one person holding clear coordination accountability across them.
The Metric Nobody's Tracking
One of the sharper observations of the discussion concerned measurement itself. In the Becker's survey, clinician burnout ranked just 13th of 17 goals health systems are pursuing with AI — arguably an undercounting of the very problem ambient documentation was built to solve. Chang's response reframed the ROI conversation entirely: the metrics that matter to both a CFO and a CMIO are shared ones — coding accuracy, time spent per note, user adoption, and patient volume growth without added clinician pressure.
Health systems want to be able to measure ROI from AI. They expect to see that reporting, with APIs that can pull that data into their internal reporting systems, allowing them to evaluate these outcomes. This is especially applicable in things like RCM, patient volumes, and then, ultimately for all of us, it's quality of care.
Joe Chang
Suki CTO
But the more interesting risk he raised was behavioral, not technical: clinicians who trust a tool stop scrutinizing its output. Early pilot users review every AI-generated note closely. By week six, many are skimming. That's not a documentation problem — it's a monitoring problem, and it points toward a near-term product category: real-time adversarial "verification agents" designed specifically to catch quiet quality drift before it becomes a pattern of rubber-stamping.
Scale Tests Different Things Than Pilots Do
Forty enthusiastic pilot users loving a tool tells you almost nothing about what happens with four thousand. The technical scaling — compute, load, reliability — is the easy part. The harder truth, as Chang noted, is that AI output is inherently probabilistic: "it's outputting to the mean," which means variability across specialties, accents, room acoustics, and clinical styles is not a bug to be eliminated but a condition to be engineered around. Sanchez's advice was consistent throughout: validate locally, against real clinicians and real patient populations, rather than trusting vendor benchmarks alone.
What This Means for the Next Twelve Months
The organizations that close the architecture gap won't be the ones that move fastest. They'll be the ones that build three things in order: shared governance and ownership across clinical, nursing, technology, and operations; an integration standard — grounded in open APIs and frameworks like HL7 FHIR — that keeps first-party data accessible rather than locked inside a single closed platform; and a validation function that treats every model update as a live event, not a settled fact.
And perhaps the most quotable line of the entire session came from Chang, describing where the real leverage now sits inside health systems experimenting with AI: "the hottest programming language right now is actually something called English." It's a glib line with a serious implication — the barrier to meaningful AI experimentation inside a hospital has dropped from a specialized IT project to something a motivated informaticist or clinician can drive directly, provided the underlying platform is open enough to let them.
The pilots have proven the value. The architecture is what proves the system.
