When AI Is Pointed at the Right Data, the Record Keeps Itself Accurate
The average physician practice maintains its information for 20 health plan contracts. Keeping those listings current cost about $999 a month when we last measured it in 2019, the equivalent of one staff day every week, and $2.76 billion a year nationally, without contributing to care. ¹
For years, the only dependable way to know whether a clinician had moved, closed a panel or changed a phone number was to ask. Practice managers answered. Health plan teams called. Credentialing specialists confirmed and reconfirmed. The method was sound. It just could not keep up.
The part I find hardest to accept is that healthcare has never been short of data. Licensure, enrollment, credentialing and network participation all generate information about where clinicians practice, and the signal that a clinician has moved is often already in the system long before a directory changes. We had the data. We picked up the phone instead.
AI changes that, if it is pointed at the right data and given clear limits. Where independent sources agree, it can act on its own. Where they do not, a person decides.
The finding that changed how I think about this problem came from a 2024 study in The American Journal of Managed Care. Researchers went back to 1,802 provider listings in Pennsylvania's ACA marketplace that an earlier survey had already found to be wrong. When they called again, about 40% were still wrong, an average of 540 days later. Just over 13% had been fully corrected. ²
Every one of those errors was found with a single phone call, and the listing stayed wrong anyway, because the record kept moving while everyone worked.
What strikes me is that the industry has spent years treating provider data quality as a discovery problem. The evidence increasingly suggests it is a maintenance problem. We know how to find inaccuracies. What we have lacked is a system capable of keeping pace with change. That distinction matters because it changes where organizations invest, what they measure and, ultimately, how they define success.
Every health plan I have spoken with is piloting AI agents somewhere in operations, and across those conversations one pattern holds: AI agents perform where they have complete context and clear authority to act, and they stall where they are asked to reason about a partial picture.
Most of the AI now being applied to provider data sits downstream, inside one organization's systems, working on a copy of the record. That copy has already been transformed and separated from its origin. An agent working on it can describe what the record says with great sophistication. It cannot see the event that made the record stale.
I believe this marks the next major shift in provider data management. For years, the industry invested in making downstream systems smarter. The bigger opportunity is to point AI at the source record, which is built from the most authoritative data source for each field, and where every new signal can be checked against everything else known about that clinician, including what the clinician has confirmed. It is the one place an AI agent can see the whole picture.
The first is an understanding of how provider data decays. Each field changes at its own pace. A clinician's name and specialty rarely change. Practice location changes more often. Panel status changes most of all. That means every field can be checked as often as it actually changes, instead of everything being re-checked on a fixed schedule.
The second is AI that can reason across many sources at once and record why it reached each decision. That was the missing piece. The information existed; the capacity to act on it, and to show the work, did not.
Enjoying this article? Join DataSpring for an upcoming webinar exploring how more accurate, connected provider data can reduce administrative burden and improve healthcare operations.
Register NowThe rule I would set for any AI working in a regulated function is simple. Its authority is bounded, and the boundary is agreement. Where the authoritative data sources for a field agree with each other, an agent can update the record autonomously and document why. No person needs to confirm it. A person confirming two sources that already agree is checking a box, not adding knowledge.
Agreement only counts when the sources are independent. Two directories that copied the same stale address are one source, not two, and an agent at the source has to be built to tell the difference. That is why pointing AI at the right data matters more than pointing it at more data.
Everything unsettled goes to a person:
Conflicting sources: a field where authoritative data sources disagree.
Clinician-only facts: a field only the clinician can establish.
Significant change: a change that suggests more than a routine update, such as a practice closing rather than a suite number changing.
When that happens, the evidence is already assembled and the question is already narrowed. The agent decides whether a change is routine. When it is not, it does not guess.
Change is observed, not reported. Agents continuously compare the record against the authoritative data source for each field, so a change is picked up when it happens rather than at the next scheduled check. The record no longer waits to be told.
Effort follows change. Fields that have not moved are left alone. The traditional cycle re-verifies everything to find the small share that changed. Monitoring inverts that.
Every change carries its evidence. Each update records what was found, where, when and why the agent acted. No health plan should hand a regulated data asset to a system it cannot interrogate, and someone always signs the network filing. When every change comes with its reasoning, the compliance team has the evidence it would otherwise spend a week assembling, and the record is verified far more often than the No Surprises Act's 90-day minimum. Auditability, rather than trust, is the right frame, and it is the only basis on which autonomy belongs anywhere near this data.
The clinician has the final word on what only the clinician knows. Attestation does not go away. Clinicians are asked fewer, more specific questions, about a record that is already current.
The record learns from its corrections. Every exception a person resolves, and every review that finds an error, feeds back into how the agents weigh sources and watch for change. The record does not just stay current. It gets better at staying current.
Together, these are the difference between a directory that is accurate on the day it is checked and one that is accurate on the day a patient uses it.
At DataSpring, our review team has called more than 14,000 providers using a method modeled on CMS's directory reviews, scoring records at both the field and form level. The clearest lesson for me is how much can be settled without asking. Six of the eight things CMS reviewers check can already be established from authoritative data alone: name, practice name, specialty, address, phone number and plan participation. ³
The other two, whether the clinician is seeing patients at that location and whether they are accepting new patients, are the hardest and the ones that matter most to patients. Data can narrow them, and AI can focus outreach on the clinicians where something has most likely changed. The clinician still confirms.
Today, that outreach lands on practices. In DataSpring's Call Verification Insight Brief, more than a quarter of practices said they receive 15 or more verification calls a month, 40% of those calls last more than 15 minutes, and 82% of respondents said the calls leave less time for patient care. ⁴ When the data settles everything it can before anyone picks up the phone, the call that remains can take under 60 seconds, and practices will feel that difference.
The larger opportunity is for health plans. When plans draw on a single continuously maintained record, one well-targeted call can do the work of many, and reliance on plan-by-plan outreach can fall significantly. DataSpring already maintains more than 7.4 million provider records and supports some 1,000 health plans and customers in credentialing through a shared model. Directory can work the same way: a shared, highest-accuracy record, available to every plan through data feeds and APIs.
The practitioner gets back the staff day each week spent answering the same questions from 20 different plans, and is asked only about what genuinely requires them.
The health plan moves capacity from reconciling records and chasing confirmations back to designing and managing networks, with the exceptions that need real judgment reaching the people best equipped to exercise it.
The patient finds a phone number that works, an office where the directory said it would be, and a clinician who is actually taking new patients. A 540-day error is not something a patient waits out. It is a specialist appointment that never gets made.
For most of my career, provider data quality has been measured by how effectively organizations could find and correct errors. That model was not a failure of effort. It was the only design available, and an enormous amount of skill went into making it work as well as it did.
Healthcare has always had the data. What it has not had is a way to turn that data into decisions as fast as the world changes. The next era will be measured by how little correction is required in the first place. When records can continuously observe change, evaluate evidence and maintain themselves at the source, integrity becomes part of the system rather than an outcome of constant intervention.
Point AI at the right data, and the record can maintain itself. We have started, and I hope the rest of the industry joins us.
SOURCES
CAQH, "The Hidden Causes of Inaccurate Provider Directories," survey of 1,240 physician practices, November 2019.
Haeder SF, Zhu JM. "Persistence of Provider Directory Inaccuracies After the No Surprises Act." Am J Manag Care. 2024;30(11):584-588. doi:10.37765/ajmc.2024.89627
Centers for Medicare & Medicaid Services, "Online Provider Directory Review Report," November 2018.
DataSpring, Call Verification Insight Brief, 2026.