Fragmented Healthcare Data Still Blocks AI After TEFCA’s Surge

Nearly 500 million health records now move across the federal TEFCA network, yet 97% of hospital leaders still say data silos slow care. That gap is the bill coming due for healthcare AI.

Cedar Gate Technologies, an IQVIA business, used a Sept. 8, 2026 briefing to argue that model quality is no longer the constraint. The constraint is whether a health system can govern, trace and secure the record an algorithm is about to use.

Cedar Gate’s Pitch and the IQVIA Umbrella

Cedar Gate, founded in 2014 in Greenwich, Conn., sells value-based care software. On its own site it says it built that stack for payers, providers, and self-funded employers who are taking more financial risk.

The Sept. 8 brief is the same thesis in AI clothes. Data quality, it says, is table stakes after years of investment. The next spend is governance, sharing, lineage, transparency and early detection of bad data. Without those five, Cedar Gate warns, AI will amplify the mess instead of fixing care.

That is a vendor talking. It is also a fair description of what surveys keep finding. Innovaccer’s State of Revenue Lifecycle in Healthcare 2026 report, based on 150 U.S. professionals at 103 organizations and reviewed with Frost & Sullivan, found 63% already run AI in at least one live workflow. The same group put fragmented data first among barriers to scale, at 62%, ahead of staff, model opacity and budget.

Abhinav Shashank, Innovaccer’s cofounder and CEO, put the bind in one line: “AI is already in production, but most organizations are trying to scale it on top of fragmented data.”

Cedar Gate’s commercial answer is a single foundation that carries analytics, population health, and payment technology. James Dye, vice president and general manager of bundled payments, said one center-of-excellence client cut surgical rates in a brain and spine program to 85% below the industry average while holding patient satisfaction above 98%. That claim is the company’s. It is also a reminder of who writes the check: the party holding claims, not the internist staring at a missing outside lab.

TEFCA Moves 500 Million Records Without Bedside Trust

Washington has spent a decade telling hospitals the sharing problem was a rules problem. The 21st Century Cures Act, signed in December 2016, made withholding electronic health information the exception. Information-blocking rules followed. TEFCA, the Trusted Exchange Framework and Common Agreement, went live a little more than two years before HHS’s Feb. 11, 2026 tally.

On that date HHS, through ASTP/ONC, said the network had nearly 500 million health records exchanged, up from roughly 10 million in January 2025. Assistant Secretary for Technology Policy Thomas Keane, M.D., also said the office was sending its first “notices of potential non-conformity” to some certified health IT developers under 45 CFR 170.580, covering information blocking and API upkeep.

THE SHARING CLOCK THAT DID NOT CLOSE THE SILO

1. December 2016: Congress signs the 21st Century Cures Act and makes electronic health information sharing the default.
2. October 6, 2022: Information-blocking duties cover all electronic health information, not a narrower data set.
3. Late 2023: TEFCA goes live as the national exchange framework.
4. January 2025: HHS counts roughly 10 million records moved through TEFCA.
5. December 22, 2025: ASTP/ONC publishes the HTI-5 proposal to drop 34 of 60 certification criteria and revise 7, with an estimated $1.53 billion in developer savings.
6. January 1, 2026: Parts of the CMS prior-authorization interoperability rule, CMS-0057-F, take effect, with API duties still due Jan. 1, 2027.
7. February 11, 2026: HHS reports nearly 500 million TEFCA records and the first non-conformity notices to certified developers.

Volume is not a bedside chart. KLAS Research’s Arch Collaborative EHR Interoperability 2024 work found only 49% of clinicians had the outside integration they expected, against 46% in 2023 and 45% in 2018. That is 4 points in 6 years of FHIR APIs, networks and a national framework.

As we begin to realize the power of AI applied to healthcare, data liquidity will be a key defining need.

Thomas Keane, M.D., Assistant Secretary for Technology Policy, ASTP/ONC annual meeting

Liquidity, in that usage, is a firehose. It is not a guarantee that the identity, med list and last ejection fraction in the payload are the ones a model should treat as true.

Why Data Quality Stopped Being Enough

Cedar Gate is right that dirty data still wrecks an AI output. It is wrong if the brief is read as proof that quality work is finished. HIMSS Market Insights, in a January 2026 canvass of 100 healthcare leaders, found 86% agreed that sharing without data preparation is of limited use for analytics or AI. KLAS clinicians keep saying outside data is inaccurate, irrelevant or hard to find.

The unpaid work sits one layer up. A national network can copy a CCD all afternoon and still leave a model unable to say which hospital last changed the allergy list, how a lab unit was converted, or whether a claims diagnosis was coded for payment rather than care. Lineage and detectability are the names Cedar Gate puts on that gap. Hospitals already bought warehouses that do not answer those questions at the point of care.

EXCHANGE VOLUME VERSUS THE RECORD A MODEL CAN TRUST

Microsoft’s numbers come from a OnePoll survey run Jan. 7 to Jan. 14, 2026 among 500 decision-makers at hospitals with 400 or more beds in the United States, the United Kingdom, Germany, France, Australia, the Netherlands and Sweden. In that sample, leaders said data silos already affect timely care almost unanimously, at 97%, and about 62% blamed legacy technology. Fifty-eight percent said they were ready to put AI agents into care-coordination and administrative work.

Readiness, in a poll, is cheaper than a joined medication list. The same Microsoft brief points to Peterborough Regional Health Centre, which connected 18 production systems in Microsoft Fabric and then reported a 43% drop in wait time to inpatient beds and a 20% quarter-over-quarter fall in unneeded lab use. Lynn Mikula, the centre’s CEO, said the foundation is what lets the organization “build at scale, move at speed, and do it with a lot more confidence.”

That is the exception the rest of the table is waiting on. Most systems still reconcile in spreadsheets. A model trained on that habit will be fast, confident and wrong in the same places the spreadsheet was.

For Payers, Context Is the Product

Cedar Gate’s brief spends a long stretch on context, and that is where its business shows. A clinically integrated network wants a care gap and a risk score. A self-funded employer wants trend, leakage and whether last year’s bundle actually saved money. Cedar Gate’s 2024 analytics paper says its tools draw on a benchmark database of 15 million member lives. That is a claims-and-contract animal, not a nursing-unit animal.

The next buyer for “AI-ready data” is often the party that already holds the invoice. Payers and employers can demand a joined clinical-plus-claims feed as a condition of a risk deal. Hospitals that spent a decade standing up an EHR then get asked to stand up a second graph the EHR still does not publish cleanly.

CMS is applying its own pressure on the administrative side. HHS projects the HTI-4 rule on electronic prior authorization and real-time prescription benefit will save more than $19.2 billion in administrative cost over 10 years. Payers face remaining CMS-0057-F API duties on Jan. 1, 2027. Those pipes will move more data into the same contested middle layer. They will not, by themselves, tell a model which record won.

A pilot that works inside one EHR proves almost nothing about the next hospital, lab or plan. Scaling is a coordination job across parties that do not share a governance calendar, a patient-matching method or a malpractice carrier. The model score is the easy part of that sentence.

Only 45% Have an AI Ethics Structure

Production AI in U.S. healthcare is already here. It is just not the clinical oracle vendors keep sketching. KLAS Research’s Healthcare AI Update 2025, drawing on 3,370 people at 1,742 organizations, found ambient speech in use at 79% of organizations. Among more than 3,000 respondents, 17 even mentioned agentic AI, and one organization reported using it.

Innovaccer’s mix shows the same gravity toward throughput.

WHERE LIVE HEALTHCARE AI ACTUALLY LANDED

• Workflow automation: 52% of Innovaccer respondents put AI here first, in high-volume operational chores rather than diagnosis.
• Documentation support: 46% use it to cut note time, with leaders citing reductions of up to 40% when the tool sits in the core system.
• Scheduling and access: 41% apply it to the front door, where missing outside records still break the slot.
• Revenue cycle: 38% automate coding and billing, a domain that already runs on industry-specific codes Cedar Gate says generic open-source models mishandle.
• Governance: 45% have an AI governance or ethics structure, so more than half of the same sample is running tools without that scaffold.

Shashank told customers the next 12 to 24 months will sort platforms that unify data, rules and workflow from a pile of disconnected tools. Hospital IT leaders outside healthcare’s own bubble are having the same argument. Agents cannot tap systems that never left the machine room, or that moved to a cloud and then sat there unmodernized. Token budgets get locked a year ahead. The plumbing does not.

HHS is trying to make the raw export usable anyway. ASTP/ONC said it would launch EHIgnite, a two-phase prize through 2027, to turn a single-patient electronic health information dump into something a person can read. Draft USCDI v7, released Jan. 29, 2026, proposes 29 new data elements. Those are standards projects. They are not a lineage store in the EHR the nurse is using tonight.

HIPAA Has No Exemption for a Chat Window

Cedar Gate’s security chapter is the one part of the brief that does not need a survey to sting. Patient data is PHI. A prompt pasted into an open model can become a disclosure. The company argues for an architecture that keeps prompts and insights inside one organization’s boundary and never co-mingles them with another tenant. That is a product claim. It is also how HIPAA already thinks about business associates, audit trails and minimum necessary use.

ASTP/ONC’s HTI-5 proposal, published Dec. 22, 2025, goes a step further on the sharing side. The accompanying federal notice would treat automated means, including autonomous AI systems using health records, as inside the definitions of access and use. If that language holds, a model that quietly pulls EHI is in the information-blocking fight, not standing beside it.

Connecting every silo for AI also widens the blast radius of a stolen credential. The same joined graph that lets a care-management agent see claims, notes and labs is the graph a ransomware crew would like to copy once. Cedar Gate’s five pillars do not remove that trade. They price it: governance roles, lineage, detectability and a refusal to treat a black-box score as a clinical act.

Generic open-source models still fail the terminology test Cedar Gate describes, because a billing code, a nursing shorthand and a problem-list synonym are not the same object. Ambient scribes got a pass because a human still signs the note. A treatment recommendation that only saw half the chart does not get that pass, and neither does the health system that deployed it on an untraced feed.

The Cures Act, TEFCA and a generation of quality programs already moved the records. Clinical AI is now asking who will pay to make those records governable before Jan. 1, 2027, when more payer APIs come online, and before the next model is allowed to act on them.

Disclaimer: This article is news analysis of healthcare information systems and artificial intelligence tools. It is for information only and is not medical advice, legal advice, privacy counsel, or a recommendation to buy, avoid, or contract for any vendor’s software. Readers should consult a licensed clinician for care decisions and a healthcare privacy attorney or compliance officer before changing data-sharing, AI, or HIPAA practices. Figures, survey shares, and program statuses follow the company pages, research houses, and federal notices named above and may change as rules, products, and exchange volumes move.