The Care Stack

Interoperability after HTI-5 and a billion TEFCA records

Victor Wilson · · 3 min read

Two pieces of federal health IT news landed this year, and they point in different directions.

More patient records are moving between organizations than ever before. At the same time, the federal rules that asked vendors to explain their AI got lighter.

Both change what your data team has to own.

What changed

The national exchange network grew fast. In June, HHS said that more than 1 billion health records had been exchanged through TEFCA, up from about 10 million in early 2025. It also announced new oversight of the networks that carry that data.

Then, in late October, ONC finalized HTI-5. The final rule removes 14 of the 60 health IT certification criteria. That’s far fewer than the 34 it proposed last December. It also drops the AI “model card” and risk-management requirements for predictive decision support tools, and it puts off changes to the information-blocking rules. HHS projects $1.5 billion in savings.

CMS is pushing in the same direction as TEFCA. At the one-year mark of its health tech pledges, its readout listed new work on real-time benefits, scheduling and bulk FHIR. Payers also face a 2027 deadline for electronic prior authorization APIs.

So the short version: far more data will flow in, and less federal paperwork will come with the AI that reads it.

More records coming in is a data problem

A billion exchanged records sounds like a solved problem. For a health system, it’s the start of a new one.

Records from outside arrive in many shapes. The same patient may show up under slightly different names. The same medication list may come from three places, each a little different. Your clinicians need to know which version to trust, and so do your analysts.

AI makes this matter more. Agents and other AI tools are starting to read across the whole chart, including what came from outside. A tool that summarizes a patient’s history can’t tell a stale outside record from a current one unless your data says where each piece came from and when.

So the work here is plain. Keep track of the source and date of every outside record you store. Get patient matching right, and measure how often it fails. Decide, with clinical leaders, which outside data shows up in which views. None of this is new. There’s just a lot more volume now.

AI transparency is now your job

Under the old rules, certified health IT had to come with model card information and risk management for predictive tools. HTI-5 removed those requirements. Vendors may still share the information. Certification no longer makes them.

That doesn’t change what you need to know before you turn a model on. If you use a sepsis alert or a readmission score, someone in your organization should be able to answer a few questions. What does it predict? What data does it use? How well does it work on your patients, not the vendor’s test set? Who checks it, and how often?

Here’s what I’d keep doing in-house, whatever the federal rules say:

  • Keep your own list of every AI and predictive tool in use, with a named owner for each.
  • Ask vendors for model card details anyway. Put the request in your contracts.
  • Test each tool on your own data before it goes live, and recheck it on a schedule.
  • Track problems, like alerts that clinicians ignore or scores that drift over time.

You don’t have to invent this from scratch. The Joint Commission’s voluntary AI certification, launched in June, looks at exactly this: governance, monitoring and training. CHAI’s governance playbooks line up with it. Either one makes a good outline for your own process.

The tools can do some of the watching. An agent can pull performance numbers each month and flag a model that has drifted. Deciding whether to keep using it is still a call for clinical and data leaders, together.

Here’s the ask. This month, ask your data and clinical informatics leaders for one list: every predictive or AI tool running in your EHR and other systems, with its owner and the last time anyone checked how it performs. If that list doesn’t exist, HTI-5 just made it your job to build it.

Discussion

    Join the discussion