Three AI Bets We Keep Getting Wrong

The AI results coming out of major EMR vendors right now are genuinely impressive. Documentation faster. Prior authorizations cut nearly in half. Early cancer detection rates running well above national averages. I am not arguing with the numbers. They are real and they matter.

But I keep watching organizations make the same three bets on their AI strategy — and I keep watching all three fall short in the same predictable ways. The frustrating part is that the failure mode is the same each time. Not the AI. The data underneath it.

Healthcare does not have a model problem. It has a data problem. And the three most common approaches to solving it are each broken in a different way.

Bet 1: trust your EMR vendor's AI

The instinct is understandable. Your EMR vendor knows your workflows. Their AI is embedded in the system clinicians already use. And the results look compelling on a conference slide.

But EMR-based AI can only reason from what the EMR contains. And for most organizations, the EMR is not the complete record. It is the record of care delivered inside your walls, through your workflows, in your system. That is an important subset. It is not the whole picture.

Here is what typically sits outside it:

• Prior EMR history: most health systems have been through at least one migration. Historical data often sits in an archive, partially mapped, largely invisible to AI running on the current system.
• Care delivered elsewhere: patients seen at competing systems, community clinics, out-of-state providers. That history does not come back into your instance in any structured, usable way.
• Payer claims data: a longitudinal view of every provider a patient has seen, across your system and everyone else's. Most EMR-based AI never touches it.
• Behavioral health and social records: intentionally siloed. The social determinants driving a disproportionate share of outcomes are frequently the least accessible data in the building.
• Independent labs, imaging archives, wearables: results appearing as scanned PDFs rather than structured data — present in the chart, invisible to the model.

And beyond what is missing, there is the question of what is there but unreadable. ECGs. Imaging studies. Pathology scans. Physiological waveforms. Clinical notes full of reasoning that most AI treats as a blob rather than structured signal. These are not edge cases — they carry some of the heaviest diagnostic signal in medicine, and most EMR-based AI cannot reason from them.

The patients where the data gap matters most are exactly the patients where AI has the highest potential value. High-risk, high-cost patients are disproportionately the ones whose records are most incomplete. The AI is most blind precisely where it needs to see most clearly.

Bet 2: build a data warehouse and layer AI on top

So the natural response is: fine, we will fix the data problem ourselves. Build a data warehouse, pull everything in, clean it up, and bring in best-of-breed AI vendors to work from a complete, normalized dataset.

This sounds right. It is also where a significant number of well-funded AI initiatives go to die. I have watched this pattern play out more than once. A health system spends 18 months and considerable budget building a warehouse. The data is cleaner. Some things are more connected. And then the AI vendor plugs in — and the problems start.

• The warehouse captures structured data reasonably well. Clinical notes, imaging, ECGs, waveforms — the multimodal data carrying the heaviest clinical signal — rarely make it in. The result is a more organized version of the same incomplete picture.
• Schema changes break everything. EHR vendors update their data models. Partner data arrives inconsistently. Fields shift across versions. Pipelines built for reporting were not designed for continuous AI workloads — and they break in ways that are expensive to fix and slow to detect.
• The warehouse is stale by design. Built for reporting cycles, not real-time clinical decisions. The AI vendor gets clean data from yesterday, or last week. For high-acuity use cases, that lag is the difference between useful and useless.
• Governance does not transfer. The warehouse normalizes data but rarely captures where it came from, how it was transformed, or what it does not include. When the AI makes a recommendation, nobody can trace the reasoning back through the data lineage.
• The AI vendor becomes a dependency. Every model update, pipeline adjustment, or new data source integration sends you back into a lengthy implementation cycle. The flexibility you thought you bought does not exist in practice.

A 2026 survey of 150 healthcare organizations found that 62% cite fragmented data systems as the top barrier to scaling AI — ahead of staffing, model transparency, and budget. Most of those organizations have data warehouses. The warehouse is not the solution to fragmentation. It is a cleaner container for it.

Bet 3: buy a portfolio of AI point solutions

The third bet tends to happen when the first two have already disappointed. Rather than waiting for the data problem to be solved, the organization starts buying AI tools for specific use cases — one vendor for documentation, another for prior auth, another for risk stratification, another for patient engagement.

Each tool works reasonably well in isolation. Then the compounding starts.

• Every tool has its own data model. None of them share a patient identity layer. The same patient looks different across systems with no mechanism to reconcile the views.
• Governance multiplies. Each AI vendor has different data handling practices, different compliance postures, different audit capabilities. Managing that portfolio becomes a full-time job — and the gaps between tools are where the risk accumulates.
• The workflow problem gets worse. Clinicians have multiple AI surfaces to check, each with its own interface, each trained on a different data slice. Cognitive load increases rather than decreases. Adoption suffers.
• There is no learning across tools. A risk model and a documentation tool trained on different data, maintained by different vendors, with different update cycles cannot learn from each other. The intelligence stays siloed even if the data does not.
• You renegotiate contracts forever. As the vendor landscape evolves — and it is evolving fast — each point solution becomes its own negotiation, integration project, and migration risk when the vendor pivots or gets acquired.

One health system CIO said it plainly in a recent industry interview: 'I don't want to put together point solutions and end up with a thousand different AI solutions. That's going to be chaos.' That is the destination of Bet Three, taken to its logical conclusion.

What all three bets have in common

Each addresses a real problem. EMR vendor AI delivers genuine workflow efficiency. Data warehouses improve data quality for reporting. Point solutions solve specific use cases. None of that is wrong.

What all three get wrong is the same thing: they treat the data problem as a solved prerequisite rather than the central challenge. They assume that if you have data somewhere, AI can work with it. But healthcare AI does not fail because organizations lack data. It fails because the data is fragmented across systems, locked in unreadable formats, stale by the time the model sees it, and ungovernable once the AI acts on it.

The organizations actually getting AI to scale in clinical settings are not the ones that picked the best vendor. They are the ones that built the data infrastructure that makes any vendor work — a unified patient record across every system, multimodal processing that makes imaging and waveforms and notes readable, event-driven pipelines that deliver intelligence when it matters, and governance that makes every output traceable and trustworthy.

That infrastructure is not a data warehouse. It is not an AI vendor. It is not a portfolio of point solutions. It is an intelligence layer built specifically for the complexity of healthcare data — one that sits alongside your EMR, reaches the data your EMR cannot touch, and makes all of it available to AI in a form it can actually reason from.

Posts 2 through 7 walk through what that layer looks like and how to build it.

So what: Before your next AI investment decision, ask which of the three bets you are currently making and whether it is delivering at the scale you need. If the honest answer is "no," the problem is almost certainly not the AI. It is the data architecture underneath it. Our Data Readiness Assessment, a 4–6 week scoped audit, delivered as a board-presentable report, is built to answer exactly that question.