Healthcare AI Pilot Purgatory Keeps Hospitals From Scaling

There are plenty of AI initiatives in hospitals. What they lack are projects that make it past the test phase. According to USAII, this is healthcare AI pilot purgatory, with lots of small experiments that appear to be working well but never scale up to be used across the organization. But the actual issue isn’t the technology, says the author, Elizabeth Enyart Allen, principal consultant at Enyart-Allen LLC. It’s getting it to scale.

While Most Health Systems Use AI, Few Govern It Well

According to those numbers, 92% of healthcare organizations are allocating more resources to artificial intelligence, but only 1% report that they have reached maturity. Be careful with those two numbers. They also appear in McKinsey research across all industries, as cited by Rightpoint, so they may not be specific to healthcare.

The same is true for healthcare-related data. According to eFax, 88% of health systems are already using AI within their organizations, while just 18% have fully developed governance and a comprehensive strategy. According to the same report, 70% of executives reportedly have at least one pilot that never moved beyond limited use.

Staff Are Using AI Before They Are Trained

Allen reports that half of healthcare workers are using AI tools, while only 16% have received formal training. As approved tools become slow or restricted, people resort to personal accounts, often referred to as “shadow AI.”

According to Black Book’s survey of 228 health professionals, shadow AI has become a routine occurrence. Front-line staff at organizations with mature governance were 30% less likely to use generic AI tools every week.

Messy Data and Risky Write-Back Slow Everything Down

Records should be made available in common, standardized formats such as FHIR and HL7, writes Allen. In addition, she mentions that OpenAI and Epic have integrated ChatGPT into the EHR space. FHIR alone isn’t enough, she says. Hospitals still need to establish policies, access restrictions, and protocols for escalation.

Her most succinct point is about reading versus writing. Summarizing a patient’s chart is low risk. Writing orders or altering a care plan isn’t, so until these steps are tightly controlled, clinicians will continue to copy information between screens. Stale data is also a safety concern because a patient’s chart may change in just a matter of minutes.

A “Zero Trust” Model for Clinical AI

Her idea is borrowed from the field of cybersecurity. A policy engine evaluates all requests in relation to the clinical context, an administrator manages each session, and a gateway strips out personal data before it reaches the AI. She adds that automated audit trails can help meet HIPAA, GDPR, and EU AI Act requirements, while doctors still have the final say.

Three Moves Leaders Will Want to Make This Quarter

Allen’s list is brief. Audit data pipelines against FHIR and HL7, implement an AI ethics committee with audit trails and automatic privacy scrubbing, and measure financial ROI in addition to clinical safety. She mentions a $3.70-per-dollar benchmark but reminds readers that pursuing it without safety measures creates risks.

The takeaway is simple. The easy part of the equation is buying AI. Whether it ever gets out of the pilot phase depends on training, governance, and clean data.

Similar Posts