Shadow AI Report: Key Findings

2. Clinicians are more likely to experiment with unsanctioned tools

While burnout and administrative burden have been lessening, they remain at untenable levels.¹ Clinicians are looking for the most efficient ways to support patients and get clinical answers—which often includes using free AI tools that could jeopardize outcomes, security, and possibly introduce patient risk.

One interesting finding was that 26% of providers who reported using unsanctioned AI tools did so out of curiosity and experimentation, whereas only 10% of administrators reported the same. This could indicate a higher risk tolerance among providers in getting quicker answers or casually testing out tools that could improve their workflow.

However, providers also ranked patient safety and inaccurate outputs as the top two AI risks, showing there may be hesitancy over the tools’ ability to meaningfully support their daily clinical work. Interestingly, providers placed a lack of source transparency as the last concern, and for providers with less than five years of experience, bias was ranked the lowest overall. Ranking these topics low in concern is at odds with evidence-based medicine and highlights the need for thorough AI solution literacy training.

3. Administrators believe AI policies are strongly communicated—providers, less so

A final gap identified in the survey is how AI policies are created and communicated. Nearly a third (30%) of administrators—including those with clinical backgrounds—indicated they were involved in reviewing, developing, or updating AI policies, while only 9% of providers were.

Administrators were more likely to say they strongly agreed (42%) that policies were clearly communicated compared to providers (30%). However, 21% of providers said they disagreed or were neutral on the clarity of communication polices compared to 11% of administrators. These differences show that AI policies need to be clearly communicated in multiple locations, not only by email or enterprise communications, but also in point-of-care locations such as the EHR. Training sessions are even more critical as AI is an emerging and constantly evolving technology, and even the most technology-savvy employees may not understand the latest risks and opportunities. Training sessions can also support active learning and policy reinforcement among providers as enterprise tools are established.

Shared AI priorities: Patient safety, security, and accuracy

Understanding where the respondents agreed can help inform policies and strategies addressing shadow AI. When assessing AI risks, both providers and administrators ranked patient safety as the top concern, along with privacy as #3 and #2, respectively. When asked about the top preferred features in AI tools, providers ranked accuracy, security, and reliability as the top three, and administrators selected security, accuracy, and ease of use.

Appealing to these shared values can help when messaging AI policies and enforcing enterprise-wide tools. Ultimately, many within the industry are seeing incredible possibilities with AI. Already, it can improve diagnoses beyond human abilities, such as image scanning, spotting bone fractures, and early disease detection, and support administrative tasks like assessing ambulance needs.⁹

With the explosion in technology and innovation, leaders have plenty of tools to consider and test before making enterprise-wide selections.