At Rochester, Minn.-based Mayo Clinic, which operates campuses in Minnesota, Arizona and Florida as well as a network of regional clinics and hospitals, AI strategy starts with the people closest to the work. The health system has focused on creating an environment where employees can identify problems in their own work and explore where AI could help address them.
That approach has helped shape an AI portfolio that now includes more than 500 models in various stages of deployment. Along the way, Mayo has been building the governance, workforce capabilities and oversight needed to support that growth.
Richard Gray, MD, vice president of Mayo Clinic and CEO of its Arizona operations in Phoenix, spoke with Becker’s about lessons the organization has learned as its AI work has expanded.
1. Give employees room to identify use cases
Mayo has largely taken a grassroots approach to AI, particularly during the earlier stages of its work, Dr. Gray said. The organization has encouraged clinicians and other employees to identify pain points in their work while expanding AI capabilities within individual departments.
“We want to create space for our people, our leaders, who feel comfortable exploring those new ideas, and they do so,” Dr. Gray said. “So we’ve largely, especially in the earlier stages of AI development, tried to engage a grassroots approach to AI and ask people to look at what are the problems that need to be solved, how do we better serve patients, what are the pain points, and how can AI play a role in that?”
Mayo is also embedding AI capabilities within departments so ideas can be developed closer to the work.
“We’re embedding and growing AI capabilities on the front lines within departments, so that they can bring those solutions forward and build a greater whole,” Dr. Gray said.
2. Build AI oversight into existing governance
Mayo’s governance model combines AI-specific oversight with adaptations of processes the organization already uses for clinical care and research, Dr. Gray said.
Many AI models move through a research protocol and Mayo’s institutional review board, which also reviews the safety of clinical trials. The results of that testing can then provide evidence as the organization considers whether a model should advance.
“We approach it in the same way that we approach bringing other evidence-based tools into our practice,” Dr. Gray said. “How can we be sure that they are evidence-based, that they have been tested, and that they are reliable?”
As the volume of AI work grew, Mayo added more AI-specific oversight. The organization appointed Micky Tripathi, PhD, former national coordinator for health information technology at HHS, as chief artificial intelligence implementation officer in 2025. Dr. Gray said Dr. Tripathi leads Mayo’s AI implementation and adoption team, which helps determine when evidence is strong enough for a model to enter practice, whether it can be deployed safely, who will own it once it is in use and how its performance will be monitored over time.
Other existing oversight remains part of the process, including department chairs and Mayo’s clinical practice leadership team.
3. Build AI skills both centrally and within departments
Mayo has taken a two-track approach to building the workforce needed to support its AI work, Dr. Gray said. The organization has recruited AI leaders and engineering talent at the enterprise level, including through Mayo Clinic Platform, its data and digital innovation ecosystem that connects clinical data, technology developers, researchers and healthcare providers. Individual departments and divisions are also hiring their own AI engineers and data scientists.
Mayo is also developing AI skills among employees already at the organization. Its education arm has worked with human resources to incorporate AI capabilities into existing job descriptions and degree programs across its five schools, while a voluntary upskilling and reskilling program supported by philanthropy from the Harper Family Foundation allows employees to select training based on their work area.
In 2025, 20,000 employees enrolled in the program, which Dr. Gray pointed to as an indication that employees see value in the training.
“So that rate of adoption and uptake, to me, is the test of whether we’re getting it right,” he said. “If we’re not forcing people into this education, but they are willingly choosing to enroll in it, finding it valuable enough that they’re telling others about it, who then enroll in it also, that is a sign of success.”
4. Look for AI that gives clinicians time back
One AI model Dr. Gray pointed to as a successful use case supports radiation treatment planning for patients with head and neck cancer. Developing those plans requires teams to determine how to deliver radiation to a tumor while protecting nearby nerves, blood vessels and the tongue.
The model was trained on treatment plans Mayo teams had developed from planning CT scans and can now automatically contour a plan at what Dr. Gray described as Mayo Clinic quality. He said the technology has reduced the work required from physicists and radiation oncologists by 75%.
“I think one of the reasons that’s an example of a win is because it took an issue of quality and consistency, and very high work effort, and converted it to something that was of great value to our physicists and radiation oncologists, who could then spend more time focusing on the patient,” he said.
5. Avoid relying on a single AI solution
One of Mayo’s lessons came from an AI project that did not move forward, Dr. Gray said. The organization was seeking a better way to review the extensive outside medical records brought by patients seeking second, third or fourth opinions, and two teams pursued different approaches to the problem.
One team focused on organizing the records and making them easier to search. The other pursued a generative AI approach that could summarize a patient’s full medical record and allow clinicians to ask questions about it in natural language. As a surgical oncologist, Dr. Gray said he was more drawn to the generative AI approach, but Mayo continued supporting both projects.
The generative AI project ultimately ended after the health-specific frontier model underlying it did not perform as expected and the technology partner decided not to continue investing in the model, Dr. Gray said. The other project remained funded and is now being used to address the problem.
“And so what we learned was, number one, not to have a single option, and certainly not to have vendor lock-in on a single underlying intelligence engine behind a product that we’re trying to use, because you will need that versatility, because some of those things change, but also because performance changes over time,” Dr. Gray said.
For health systems earlier in their AI work or operating with fewer resources, Dr. Gray said they do not need to replicate Mayo’s approach to get started. He pointed to ambient AI scribes as one area where hospitals can adopt commercially available technology without building their own AI infrastructure.
“And when we see these ambient listening technologies going into Mayo Clinic and other healthcare organizations, we see a restoration of that relationship and more joy coming back into medicine, and I don’t think you can put a price tag on that,” Dr. Gray said.
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.
