Every health system I talk to is “doing AI.”
There’s a pilot in radiology. A GenAI scribe in ambulatory. A predictive model for readmissions – perhaps by someone who built two years ago, left the organization and now nobody can quite explain anymore. Then, usually there’s always a slide deck with the word “transformation” on it, and senior leadership has been promised something exciting by the next year.
And almost none of these systems have the foundation to do AI safely or at scale. Chances are, it may not survive the departure of the one analyst who built it.
Now, I don’t say that as criticism. As healthcare is only beginning the transformation journey early this decade, this is a natural landscape. I say it as someone who walks into these environments, opens the hood, and sees the same pattern repeat in system after system. The heart and ambition is real. The intent is good. The foundation is missing.
I am writing this post to help understand what that foundation actually is. Why is that most organizations don’t have it, and what an honest readiness assessment looks like. I am talking about the kind a CEO, CIO or CDAO can run in 30 to 90 days without a six-figure consulting engagement.
The gap between AI ambition and AI readiness
It seems healthcare has a unique and somewhat peculiar problem with AI. The intelligent automation using AI is becoming table stakes, as the clinical case is overwhelming – where the workforce shortage alone is projected to be 200,000 to 450,000 nurses by 2027. Meantime, value-based care is doubling the population we’re accountable for. Consumer expectations are also rising when it comes to digital experience, which most health systems are unable to deliver due to infrastructure gaps. Therefore, the pressure to “do AI” is not manufactured or hype, it is existential.
Unfortunately, pressure alone does not create capability.
What I see in most current-state assessments across health systems is a consistent set of structural gaps that prevent AI achieving true operationalization, and prevents it from being anything more than a collection of pilots. These gaps are not exotic, nor unique to any one system. They are, in fact, so common that we can assume that they are present until proven otherwise.
There are six of them.
The six gaps
1. Siloed data assets, no enterprise platform.
An SQL data warehouse barely held together by tribal knowledge. Departmental SQL or Oracle servers, with a variety of data storage, analytics tools. Perhaps multiple Epic instances or a variety of EHR systems that have never been linked. Perhaps a separate cloud warehouse that started but stayed siloed for special purpose. Maybe a research warehouse running on a separate infrastructure for few high-value projects. There is no single place for enterprise grade data, no governed highly curated, well-defined matured “gold” data layer with standardized KPIs; and let’s keep in mind roughly 80% of healthcare data is unstructured and effectively invisible.
2. No real-time or modern AI stack.
Most systems built their legacy data analytics systems before batch reporting, historically grew from there. Hence, their architecture is not ready in a world that now demands streaming clinical events, LLM-based interfaces and agentic workflows. Then there is no place for a data scientist to work with accessible data that is both governed and usable. Many systems begin working on an “AI roadmap” assuming an infrastructure that simply does not exist.
3. Governance is absent, not weak.
Let’s get real here – No data dictionary, no lineage, no quality control or automated monitoring. No HITRUST or SOC-2 readiness. For research or commercial partnership, de-identification is done study by study, case by case, by some trusted hand. A good check point is, when someone asks, “Where did this number come from”, the answer often is a person’s name, not a trusted enterprise system.
4. Organizational and talent fragmentation.
Yet, it’s not that data analytics talent investment is scarce, rather most likely the opposite except these talents are scattered across 10 to 20 teams, reporting to as many different leaders without any common CoE collaboration (may or may not be centralized). Finance has its analysts. Pop health has theirs, so does research or quality; most likely various parts of hospital administration as well. None of them work for the same person. None use the same tools or produce the same comparable numbers. They grew sporadically for some business urgency or frustration with IT. There is no center of excellence, there is only a federation of cottages.
5. Tool sprawl without strategy.
Often, there are a hundred-plus disconnected tools. Similarly high number of vendor data feeds, platforms, quick analytics solution for specific business. Each procured to solve a specific problem at a specific moment, none integrated, operates on systems feeding specific limited siloed data feeds; hard to consolidate because now they are part of critical business operations, all renewing on different cycles; almost certainly they are collectively costing more than a modern unified platform would.
6. Operational immaturity.
Obviously, IT cannot commit to any SLAs. No runbooks. No on-call rotation for critical data products that support clinical operations. Latency is measured in days. Issues are known only when discovered by users, not monitors or automated operations. While the data team appears to be heroic, they are often exhausted. This is a sign of a broken process, not a strong data culture.
If in any system three or more of these are present, the organization is not ready for enterprise AI. It may be ready for more pilots, which is a different thing – as it can only do idea validation, but without final successful adoption.
Why pilots succeed and enterprise AI fails
Often boards and senior leadership are confused by this disconnect. Their view is simple and reasonable – the pilot worked; the vendor demo was impressive. The numbers show that the model in the published study outperformed clinicians on a benchmark. So why hasn’t anything scaled?
The answer is often simple as well – because pilots succeed on the strength of individuals; may be a motivated clinical champion, a smart clever engineer, a vendor doing the heavy lifting on data extraction to prove a point or gain new business. But enterprise AI succeeds on the strength of infrastructure. Pilots can easily route around broken governance and fragmented data. Enterprise deployments and adoption cannot.
A robust enterprise deployment takes a lot more rigor and lifecycle management. The very moment you try to take a successful predictive model and run it across every facility, perhaps multiple service line, in real time, integrated into clinical workflow, with proper monitoring for drift and bias (hopefully with audit trails for regulators), with retraining when the underlying population shifts – every one of the six gaps above becomes more visible and a blocker. Maybe the model that worked in the pilot now needs data that lives in three places, governance that doesn’t exist, an MLOps platform that was never built, and an organization that doesn’t have a clear owner.
This is why systems can simultaneously have impressive AI demos, and yet embarrassing AI maturity. The two are measured differently for obvious reasons.
What “ready” actually looks like
A ready organization has, at minimum:
- A unified data platform at its core, cloud-based, organized in layers (raw, modeled, curated) so that every data analytics and AI workload draws from the same governed source.
- A real-time ingestion and integration capability for HL7, FHIR and operational events – not only dependent on nightly batch.
- Along with the above, an enterprise data catalog with clear lineage and automated quality monitoring, and HITRUST/SOC-2 readiness as a continuous program, providing robust protection and anytime ready-to-audit.
- A consistent de-identification engine for research, commercialization and any regulated secondary use, governed centrally rather than negotiated per study or use case.
- A data analytics AI organization with a clear center, federated spokes in the business, and governance bodies (data council, AI council) that meet and make decisions.
- An MLOps capability — with model registry, automated monitoring, with retraining workflow — so that AI in production is operationally managed, governed and not orphaned.
- A self-service environment that lets power users experiment, heavy analytics users like finance, HR, pharmacy, and research faculty work with governed data on their own, without waiting in an IT queue.
Honestly, this is not exotic. None of it requires invention. All it requires is proper sequencing, investment and the leadership willingness to consolidate what has been allowed to fragment.
What a 30-day honest assessment looks like
You are a CEO, CIO or CDAO and you want to know where you stand. Honestly, you do not need a heavy consulting engagement to start. You need four areas of conversations and a spreadsheet.
Week 1 — Inventory. Ask your CIO to produce, in a single document, every data platform, warehouse, vendor data feed and analytics tool in use across the system. Most organizations cannot do this in a week, which is itself a finding. Aim for a list, not a database — exhaustiveness is less important than the act of looking.
Week 2 — People. Ask HR for a list of every employee with “analyst,” “data,” “informatics,” “BI,” or “scientist” in their title, and the leader they report to. Map the org. If the resulting chart has more than five reporting lines into analytics work, you have a fragmentation problem.
Week 3 — Governance. Ask three questions of your data team. Where is our data dictionary? Show me our data lineage for the top five enterprise KPIs. What is our HITRUST/SOC-2 readiness today? The quality of the answers — and the speed at which they arrive — is your governance score.
Week 4 — Use cases. Ask clinical and operational leaders to name the three AI or analytics capabilities they would most want to deploy in the next 12 months. Then ask your data team whether the foundation exists to support them. The gap between the demand list and the supply capability is your readiness gap, in concrete terms, in business language.
At the end of 30 days, you will have a current-state picture honest enough to act on. Remember, not a perfect one. An honest one. That is more than most organizations have.
The leadership act
One key thing to remember, the hardest part of this work is not so much technical. The architecture is well understood. The tooling is mature. The talent exists in the market. Almost always, what is missing is the leadership willingness to look at the current state honestly and name what is broken. But visionary transformative leaders cannot be shy knowing that some of what is broken was procured, stuffed, or sponsored by not so distant people; though likely in response to the need of another time and urgency.
The systems that will do AI well over the next five years are not the ones with the most impressive pilots today. They are the ones whose leadership had the realization, audacity and discipline to build the foundation first.
That work is not glamorous. Not shiny, or immediately as visible. But it is the work.
This is my first in a five-part series on my thoughts about what it takes to move a health system from data fragmentation to AI-native. Next: the Bronze-Silver-Gold conversation every CEO needs to have with their CIO.
I would welcome reactions, disagreements and war stories from your own organizations in the comments. I believe the more honest the field gets about this, the faster we all move.
At the Becker's 11th Annual IT + Revenue Cycle Conference: The Future of AI & Digital Health, taking place September 14–17 in Chicago, healthcare executives and digital leaders from across the country will come together to explore how AI, interoperability, cybersecurity, and revenue cycle innovation are transforming care delivery, strengthening financial performance, and driving the next era of digital health. Apply for complimentary registration now.