Healthcare's AI problem isn't the model

Dr. Jaime Bland, CEO of Aquila Health

Healthcare's AI conversation has entered a new phase.

For the past two years, health systems have focused largely on identifying use cases. Ambient documentation, clinical copilots, revenue cycle automation, predictive analytics and generative AI assistants have dominated strategic planning sessions. Boards want AI roadmaps. CEOs want returns. Vendors continue unveiling increasingly sophisticated models.

But as provider organizations attempt to move AI initiatives from pilot projects into enterprise operations, a different reality is emerging. The organizations seeing the greatest challenges are not necessarily struggling because of the AI itself. They are struggling because of the data beneath it.

That shift is creating one of the most important strategic questions facing healthcare CIOs today: Is their organization's data infrastructure actually ready for artificial intelligence?

"Many healthcare organizations are approaching AI as a software deployment when it is really a data infrastructure initiative," she said.

The observation comes at a critical moment. As agentic AI, ambient listening systems and enterprise-scale generative AI deployments move from experimentation toward production, healthcare leaders are increasingly discovering that the limiting factor is not model capability. It is data quality, consistency, governance and trust.

From AI pilots to AI production

The healthcare industry has spent years hearing that data is a strategic asset. Yet many organizations are now confronting an uncomfortable reality: Years of data investments have not necessarily produced data environments capable of supporting enterprise AI.

"The industry is reckoning with years of data investment that has not provided the ROI that organizations were told to expect," Bland said.

One reason is that pilots often mask the underlying problem.

Successful proofs of concept typically rely on carefully curated datasets assembled specifically for testing. Data teams clean records, reconcile inconsistencies and prepare information before introducing it to the model. The result is often impressive performance.

Production environments tell a different story.

"Production, on the other hand, runs on the live feed, and the live feed is messy, non-standard, and incomplete in ways the pilot never surfaced, because it was never run on live data," Bland said.

For CIOs, that distinction has become increasingly important. Organizations that once viewed AI as a technology procurement exercise are discovering it is often an enterprise data modernization challenge.

Fragmentation remains one of the largest obstacles. Clinical systems, workforce platforms, financial applications and operational tools frequently maintain separate versions of the truth. Even within a single EHR, clinicians often document information differently, creating inconsistencies that become amplified when AI tools attempt to analyze information at scale.

"Most of the AI being deployed in health systems right now sits on top of that flawed, fragmented data, so the results it produces are poor," Bland said.

Interoperability's unfinished work

Healthcare has made significant progress in interoperability. Federal policies, information-blocking regulations and TEFCA have improved the industry's ability to exchange information.

But Bland argues that interoperability solved only part of the challenge.

"Interoperability investment did what it set out to do. It made data move," she said. "What it did not do is make that data worth retrieving."

That distinction highlights the next challenge for healthcare IT leaders.

Patient information increasingly can be exchanged between organizations. Yet assembling a complete, usable patient record remains remarkably difficult. Records arrive using different identifiers, coding systems and documentation structures. Consent requirements vary. Governance requirements differ. Reconciliation often requires significant manual effort.

For organizations pursuing advanced AI strategies, these inconsistencies become more than operational nuisances. They become barriers to scale.

The challenge is particularly significant for rural and under-resourced organizations, which often lack the staff necessary to normalize and govern information from multiple sources.

"We have spent years proving we can move data, but that's the easy part of the equation at this point," Bland said. "What we have not done is make it worth retrieving or usable when it lands."

The governance challenge hiding in plain sight

Healthcare leaders often discuss AI governance in terms of model oversight, transparency and safety. Bland believes many organizations are overlooking an equally important issue: data governance.

When evaluating AI readiness, she repeatedly encounters the same problems.

Patient identities do not align across systems. Clinical records conflict with payer data. Outcomes information is incomplete. Social determinants of health data may be sparse. Information about race and ethnicity is frequently inconsistent.

"The gaps tend to sit where they matter most, since outcomes, social factors, and race and ethnicity are usually the thinnest fields, which is also where bias hides," she said.

Those shortcomings frequently remain invisible during testing because models are evaluated against datasets containing the same limitations present during training.

The weaknesses emerge after deployment.

"AI cannot find a pattern that is not in the data," Bland said.

For CIOs, that reality changes the conversation around governance. AI governance increasingly becomes data governance. Questions about bias, reliability, explainability and clinical trust cannot be separated from questions about data quality and stewardship.

As healthcare organizations prepare for broader AI adoption, Bland argues that leadership attention should shift toward the infrastructure beneath the applications.

Rather than beginning with use cases and hoping the necessary data exists, organizations should first establish trusted, governed and accountable data environments.

That means treating data as a strategic product. It means assigning ownership. It means building governance structures designed to facilitate responsible innovation instead of simply restricting access. It means continuously monitoring data quality and model performance over time.

None of those initiatives generate the excitement associated with the latest AI announcement. Yet they may ultimately determine which organizations achieve sustainable value from artificial intelligence.

"The models will only get more capable from here, and that part keeps getting easier," Bland said. "The data underneath them is the slower, harder work, and it is what decides whether that capability turns into something a clinician can rely on."

For healthcare CIOs, the message is increasingly difficult to ignore. The next competitive divide may not emerge between organizations using AI and those that are not. It may emerge between organizations that invested in trustworthy data foundations and those that continued layering increasingly powerful AI tools on top of information they never fully fixed.

Follow Bill's health IT coverage on LinkedIn: Bill SiwickiEmail him: [email protected]Healthcare IT News is a HIMSS Media publication.

WATCH NOW: Emerging advancements in radiology AI