In January 2025, roughly 10 million health records moved through the Trusted Exchange Framework and Common Agreement (TEFCA), the national framework for health information exchange. By early 2026, that number was nearly 500 million.
Most operations leaders never saw that shift happen. No announcement landed in your inbox. No vendor called to walk you through it. But it's part of the reason the AI pilot that stalled last year might suddenly be viable this year, and it's also why the AI pilot that worked in one department still can't talk to the one next door.
The AI Conversation Has Been Pointed at the Wrong Layer
For the past two years, most AI conversations in healthcare operations have centered on the tool. Which ambient scribe. Which prior auth assistant. Which chatbot for patient scheduling.
Adoption numbers back up how fast this has moved. AI adoption across healthcare organizations went from 72% to 85% in a single year. Two thirds of physicians reported using AI in 2024, up from under 40% the year before.
But here's what doesn't get discussed as often: roughly 80% of hospitals access their AI primarily through modules built into their electronic health record (EHR). That means for most organizations, "AI strategy" isn't really a strategy. It's whatever the EHR vendor decided to ship this quarter.
That's not necessarily bad. But it does mean the thing actually determining what AI can and can't do in your operation isn't the AI. It's the data layer underneath it, and whether that layer can talk to anything outside your EHR.
Why the Plumbing Matters More Than the Model
71% of non-federal acute care hospitals used predictive AI integrated with their EHR in 2024, up from 66% the year before. A pattern worth sitting with. Steady growth, embedded directly into existing systems.
Meanwhile, at Mass General Brigham, a sepsis detection model now runs across 13 hospitals. That's not a single-site pilot. That's a model pulling consistent data across a dozen different facilities, which only works if those facilities can exchange data in a usable format in the first place.
Same with their revenue cycle AI, launched in 2025 to improve coding accuracy. Coding accuracy depends on clean data flowing between billing systems, claims systems, and eligibility checks. If that data is inconsistent or locked behind proprietary formats, the AI is working with bad inputs no matter how good the model is.
This is the part that doesn't show up in vendor demos. The model is the easy part to demonstrate. The integration work, getting data to move cleanly between systems that were never designed to talk to each other, is where projects actually stall. And it's rarely the vendor's problem to solve. It's yours.
What's Actually Changing, and What Isn't
The real shift is in the exchange layer. TEFCA's growth from 10 million to nearly 500 million records exchanged in about a year is a meaningful signal that the infrastructure for moving data between systems is maturing faster than most operations teams have clocked.
In a recent Becker's survey on AI adoption, 63% of respondents ranked interoperability as the single most appealing AI capability, ahead of any specific clinical or administrative use case. That's a notable thing for operators to rank above the tools themselves.
What hasn't changed, and what tends to get glossed over: "connected systems" doesn't mean one unified platform. In practice, most organizations are running EHR-native AI, a handful of point solutions, and some exchange network connectivity, all at once. That's not a failure state. It's just the reality. But it creates real ongoing work: vendor management, data governance, security review, and someone keeping track of which system is the source of truth for what.
That operational overhead tends to be underestimated relative to how much attention goes to what the data flow itself will eventually enable.
The Gap Between Early Movers and Everyone Else Is Widening
This isn't a future problem. The U.S. healthcare system avoided an estimated $258 billion in administrative costs in 2024 through better electronic data exchange, according to the Council for Affordable Quality Healthcare (CAQH) Index released earlier this year. That's not a projection. It's what already happened, and the savings went disproportionately to organizations that had already invested in their exchange infrastructure before layering AI on top, not the other way around.
The legacy systems issue is why. Proprietary data formats and inconsistent standards remain the top-cited barrier to integration. Most teams report needing 6 to 9 months for standard interoperability implementation, and multi-site consolidation often runs 12 to 18 months. Organizations that haven't started that work aren't behind on a single deadline. They're behind on a body of infrastructure work that takes a year or more to complete, while TEFCA exchange volume keeps climbing and the organizations already connected keep pulling ahead on what they can do with AI.
Regulatory pressure is real here too, even if the specifics are still shifting. CMS interoperability reporting requirements expanded in 2026, with new reporting obligations taking effect this year. The exact compliance dates for some of the underlying API rules have moved before and may move again. What hasn't moved is the underlying trend: the bar for what counts as "connected" keeps rising, and the gap between organizations that built for it early and those that haven't is measured in the savings CAQH is now reporting, not in a single missed deadline.
Start With the Data Layer, Not the Tool
If you're evaluating an AI tool right now, the first question probably shouldn't be about the tool. It should be about what your data layer can and can't do.
Can this tool pull data from systems outside your EHR? Does your organization have any visibility into your Qualified Health Information Network (QHIN) connectivity, the network that determines whether you can participate in frameworks like TEFCA? If a use case spans departments, sites, or payers, has anyone checked whether the underlying data can actually move between those systems, or is that an assumption baked into the pitch?
None of this requires rebuilding your entire infrastructure before touching AI. EHR-native tools will keep working the way they do now. But any use case that crosses systems, sites, or organizational boundaries is going to run into the limits of a closed data layer fast, and that's exactly the category of use case with the clearest documented ROI so far.
The infrastructure bet for 2026 isn't about picking the right AI tool. It's about understanding whether your organization can move data well enough for the AI tool to matter.
What would you find if you asked your IT team how your systems currently exchange data with anything outside your EHR? For a lot of operators, that's a conversation that hasn't happened yet, and it might be the most useful one available right now.
FAQ
We already have an EHR. Isn't that "connected" enough?
For single-system, single-site use cases, often yes. But most cross-department or cross-site use cases need exchange-layer connectivity that the EHR alone doesn't provide. The EHR is one node, not the network.
Do we need to rebuild our data infrastructure before we can use AI?
Not entirely. EHR-native tools will keep functioning as they do now. But any use case that spans departments, sites, or payers will hit the limits of a closed data layer quickly, and that's where some of the clearest ROI has shown up so far.
What's actually required by CMS in 2026, versus what's just best practice?
CMS interoperability reporting requirements expanded this year, and the specific compliance dates for some of the underlying API rules have shifted before and could shift again, so treat any single deadline with caution. What's not in question is the direction: reporting obligations and scoring expectations are increasing year over year, and broader "connected systems" work beyond the minimum is discretionary but increasingly something peer organizations are ahead on.
Is TEFCA something we need to actively join, or is it background infrastructure?
For most organizations, it's becoming background infrastructure, the way payment networks are. But if no one has checked your organization's QHIN connectivity, that's worth a direct conversation with IT.
If interoperability is the real bottleneck, why does every AI vendor pitch focus on the model?
Because the model is the easy part to demo. The integration work is where projects stall, and vendors generally don't own that risk. That's left to the organization.
How long should we budget for this kind of work?
Standard interoperability projects run 6 to 9 months. Multi-site consolidation runs 12 to 18 months. Any AI rollout that depends on cross-system data should be planned against those timelines, not the AI vendor's timeline.
Sources: Becker's Hospital Review, HHS.gov on TEFCA, CAQH Index via GlobeNewswire, Becker's on healthcare AI ROI, IntuitionLabs on hospital AI adoption, Nature Health on AI implementation, Medisolv on 2025 interoperability requirements, Invene on CMS interoperability, Becker's on TEFCA vs. CMS networks