Why Most Clinical AI Initiatives Stall Before They Start

Every hospital is investing in AI. Talk to any health system CIO in 2026 and you’ll hear about an AI roadmap, an AI committee, and at least one project that is supposed to demonstrate value no later than the end of fiscal year.

Then ask them a more specific question: How is the project going?

In a striking number of cases, the honest answer is that it has stalled. It’s rarely because the AI does not work. Usually, it’s the data underneath the AI that is not good enough to act on.

This is especially true in the operating room, where the data is famously inconsistent and the consequences of getting it wrong are clinical, not just operational. Understanding why this happens, and what to do about it, is the most important conversation a CIO can have with their AI strategy team in 2026.

Where Data Goes Wrong

In a typical perioperative environment running MEDITECH Expanse, the documentation flow works somewhat like this:

The patient is wheeled into the OR. Anesthesia begins. The surgeon arrives. The surgical team runs through a Patient Safety Time Out. The case proceeds. Throughout, a circulating nurse is responsible for capturing milestones like the time the patient rolls into the OR, when anesthesia starts, incision, closing, and patient out of room.

In most hospitals, the nurse captures these milestones from memory, often after the case ends, sometimes at the end of shift. The workstation may be across the room. The capture method may be a paper checklist that gets transcribed later. The cognitive load is high, the workflow is chaotic, and the documentation that ends up in the record is reconstructed rather than recorded.

This is not a failure of the clinical staff. It is the predictable result of asking humans to be data entry clerks while they are also delivering surgical care.

The data that lands in the EHR at the end of this process may be sufficient for billing and for retrospective analytics. It is not sufficient as a foundation for AI you use to make important decisions.

Evaluating Hospital Data for AI Readiness

There are general principles that apply to data requirements for clinical AI - especially the operational AI that helps perioperative leaders manage block utilization, first-case-on-time starts, and turnover times:

• Completeness. Every standard milestone for every case has to be captured, not just most of them. AI operating on partial data produces incorrect conclusions.
• Consistency. The same milestone has to be captured the same way across surgeons, services, and shifts. Variation introduces noise the AI cannot distinguish from signal.
• Timestamps. Real timestamps, not reconstructed estimates. The difference between "incision at 8:34" and "incision sometime around 8:30" is the difference between an AI that can find patterns and one that cannot.
• Role attribution. Knowing who did what, not just what was done. Without attribution, the AI cannot help leaders understand variation between surgeons or teams.
• Normalization. The same surgical milestone has to be encoded the same way across instances, regardless of how the surgeon or nurse describes it informally.

Most perioperative documentation today fails to meet most of these principles. Some of it fails all five.

Hospitals that try to feed this data into an AI get answers that look right but clinical leadership can’t defend when challenged. In other words, answers they cannot base their operations around.

And once a CMIO or perioperative director loses confidence in the AI, the project is functionally dead even if the contract is still in force.

A predictable response, when this data quality problem becomes visible, is to retrain the staff – maybe with monitoring and audits added on.

This rarely works, for a reason that is well understood in operations research but easy to forget under deadline pressure. You are asking humans to change behavior in the middle of high-stakes, fast-moving work, in service of a long-term outcome (better data) that they cannot see and do not personally benefit from.

“Generically trained models don’t cut it for enterprise applications and agents,” Google explains in a 2025 white paper. “That’s why many businesses ground foundation models in real-time web information, enterprise data … and other sources of relevant information.”

Relevant and current data is a start, but it has to be usable.

Designing Systems to Get the Right Data

There are IT solutions that make behavior change unnecessary. Specifically, solutions that capture the desired data easily and automatically, as a byproduct of work that is already happening.

This is the design principle behind LiveData's approach. Starting with a handheld Clicker device, clinicians capture every checklist step and timing milestone at the point of care. The nurse doesn’t need to remember, transcribe, or re-enter anything. Information flows into the operational store automatically, with timestamps and role attribution, in a normalized form in real time.

AI built on top of LiveData’s solution, which is currently used in over 90 hospitals and in over 2 million surgeries, gets the data foundation that it needs.

Making a Confident IT Leadership AI Decision

For a CIO evaluating AI vendors, the most important question is rarely about the AI itself. Many times, it is about the underlying data that the AI is basing its decisions on.

A vendor whose AI runs on whatever the EHR happens to contain, without the architectural layer that makes the data trustworthy in the first place, is selling a project that will stall. A partner whose AI can explain its answers, including which data was used and which was excluded, is a rare find.