Monday, 14 September 2026
Artificial intelligence is beginning to cross an important threshold in healthcare. The question is increasingly moving beyond whether algorithms can perform impressive medical tasks and towards whether hospitals can integrate those capabilities safely into everyday patient care. Mayo Clinic believes that transition could represent one of the most significant changes in modern medicine. Speaking at AI4 2026, Micky Tripathi, Chief AI Implementation Officer at Mayo Clinic, described an organisation attempting to move artificial intelligence beyond research projects and individual algorithms into clinical workflows used by doctors, nurses and patients. The strategy encompasses administrative efficiency, medical imaging, cardiovascular medicine, pathology, cancer detection and potentially much more personalised models of care.
The scale of Mayo’s programme demonstrates how quickly healthcare AI is developing. Tripathi said the organisation currently has roughly 450 AI-related solutions either deployed or progressing through its development pipeline, with 128 already implemented in practice. Around three-quarters of the systems have been developed internally, according to the presentation, reflecting Mayo’s strategy of allowing clinicians and employees to originate solutions to problems they encounter directly. The financial commitment is also substantial. Tripathi said Mayo invested approximately $450 million in technology during the previous year, of which roughly $100–125 million could be directly associated with AI and agent-based technologies. These figures were presented by Mayo at AI4 and should be regarded as organisation-reported investment figures rather than independently audited expenditure.
The investment reflects Mayo’s view that the largest risk associated with artificial intelligence may not simply be deploying it incorrectly. There is also a risk that healthcare organisations move too cautiously and fail to capture improvements in patient outcomes that the technology could make possible. That position does not mean abandoning controls. Instead, Mayo is attempting to create a governance system that allows lower-risk applications to reach clinical practice relatively quickly while subjecting systems with greater potential consequences to more extensive scrutiny.
The starting point for Mayo’s strategy is the enormous digital transformation healthcare has already undergone. Two decades ago, large parts of the US healthcare system still depended heavily on paper medical records. The widespread adoption of electronic health records subsequently digitised enormous quantities of clinical information, but digitisation alone did not make that information easy to interpret. Modern patient records can contain structured information, physician notes, laboratory results, medical images, audio and data imported from other healthcare organisations. For patients with complicated medical histories, the quantity of information can become enormous.
This is particularly relevant for Mayo because a significant proportion of its patients travel from elsewhere seeking specialist treatment for serious or complex conditions. They can arrive carrying years of records generated by multiple hospitals and physicians. Before providing treatment, Mayo clinicians must reconstruct what happened previously and identify the pieces of information most relevant to the patient’s current condition. Artificial intelligence potentially changes the economics of that process. Instead of requiring clinicians to search manually through hundreds or thousands of pages, AI can classify, organise and summarise outside medical records before an appointment.
Mayo has developed an internal system for this purpose that processes incoming medical material and creates structured summaries for clinicians. The objective is not to make the diagnosis independently but to ensure that physicians can spend more of their time interpreting the information rather than locating it. This distinction between improving healthcare efficiency and expanding medical capability runs throughout Mayo’s AI strategy. Some applications remove administrative work that humans currently perform. Others attempt to identify patterns that human perception may not be capable of detecting.
The first category includes increasingly familiar technologies such as ambient clinical documentation. With appropriate consent, AI can help capture information from conversations between clinicians and patients and prepare documentation for review. Tripathi said adoption of ambient documentation across Mayo has become widespread enough that the technology is beginning to resemble an ordinary clinical productivity tool rather than an experimental AI application. Nursing provides another example. Shift changes require outgoing nurses to transfer both documented information and practical knowledge about patients to incoming staff. Mayo nurses have been involved in developing an AI-supported virtual assistant intended to make this handover more efficient and reduce the amount of manual work required to transfer information between shifts.
The organisation is also experimenting with patient-facing digital assistants in areas where specialist capacity is limited. Genetic counselling is one example. Rather than requiring every initial interaction to take place directly with a genetic counsellor, Mayo is testing virtual interfaces that can conduct preliminary conversations and help determine what type of specialist interaction should follow. The potential benefit is not simply labour productivity. Where highly specialised healthcare professionals are in limited supply, technology can potentially increase the number of patients who can access their expertise.
Medical imaging offers another relatively straightforward productivity opportunity. Mayo has worked with Microsoft on technology that can assist radiologists by identifying lines and tubes appearing in chest X-rays. These objects can obscure anatomy and require radiologists to distinguish medical equipment from clinically relevant structures before interpreting an image. Tripathi said the technology can save approximately 30–45 seconds when reviewing an individual chest X-ray. That sounds relatively small until it is multiplied across radiologists who may interpret more than 100 such images in a working day. The example illustrates an important feature of healthcare AI: significant productivity improvements may come from accumulating many small reductions in repetitive work rather than one dramatic technological breakthrough.
Pathology represents a much larger transformation. Mayo has been digitising millions of traditional pathology slides, converting physical tissue samples into images that can be analysed computationally. Once pathology becomes digital, artificial intelligence can search for patterns within tissue that would be extremely difficult to identify consistently through human observation alone. Mayo is developing foundation-model technology around these digital pathology resources and is increasingly integrating AI-enabled pathology into clinical workflows. The result could eventually change both the speed and the depth with which tissue samples are analysed.
Tripathi highlighted the potential application during surgery. When surgeons remove a tumour, they sometimes need rapid confirmation that the edges of the removed tissue are free of cancer. Traditional frozen-section analysis can require tissue to be prepared, stained and examined by a pathologist while the patient remains in surgery. Mayo is researching digital approaches that could substantially shorten parts of this process. Tripathi described the possibility of reducing a procedure that can take approximately 20–25 minutes to closer to five minutes through digital imaging and virtual staining. If such systems prove clinically reliable at scale, the implications extend beyond diagnostic speed. Reducing time during surgery can affect operating-room utilisation, staffing requirements and the period a patient remains under anaesthesia. In a large hospital, even modest reductions in procedure times can accumulate into meaningful additional capacity.
The more significant frontier, however, is using AI to identify disease before it becomes visible to clinicians. Pancreatic cancer provides one of Mayo’s strongest examples. Researchers developed an AI system capable of analysing routine abdominal CT scans and detecting subtle signs associated with pancreatic cancer before tumours become apparent through conventional visual interpretation. Research published in 2026 validated the model across data designed to reflect real clinical environments and indicated that signs of disease could potentially be identified as much as three years before clinical diagnosis in some patients.
That distinction is potentially critical because pancreatic cancer is frequently diagnosed after the disease has progressed considerably. Earlier identification can expand the possibility of intervention while the cancer remains treatable. The research remains an example of why clinical validation is essential. Detecting a statistical signal associated with future cancer is not equivalent to diagnosing every patient accurately years in advance. Mayo is continuing to evaluate these technologies across broader patient populations before they become routine clinical tools.
Cardiovascular medicine provides another area where Mayo has accumulated significant AI experience. Researchers have been developing algorithms capable of extracting information from electrocardiograms that extends beyond what the test was traditionally designed to reveal. A conventional ECG records the electrical activity of the heart. With AI, the same relatively inexpensive examination can potentially reveal patterns associated with conditions such as heart failure, atrial fibrillation and cardiomyopathy before those conditions become clinically obvious.
Researchers are also exploring the concept of physiological age. Instead of looking only at a person’s chronological age, algorithms can analyse cardiovascular information and estimate whether the heart appears biologically older or younger than expected. The broader principle is important. Healthcare organisations already possess enormous quantities of diagnostic data collected for specific purposes. Artificial intelligence may allow additional medical information to be extracted from tests that patients are already undergoing.
The same concept is being explored with brain imaging, genomic information and other medical data. By comparing historical patient information with eventual outcomes, researchers can search retrospectively for patterns associated with diseases that clinicians could not identify at the time. Genomics could eventually help determine which patients are most likely to respond to particular medications, potentially reducing the trial-and-error process involved in selecting treatments. Rather than prescribing a standard therapy and waiting to determine whether it works, AI could potentially combine genetic, clinical and other information to estimate which treatment is most appropriate for an individual patient.
Mayo is also examining entirely new sources of clinical information. One of the more unusual examples presented at AI4 involved the human voice. Subtle changes in speech may contain physiological information that the human ear cannot recognise reliably. Certain cardiovascular conditions can affect nerves associated with the vocal system, while neurological or cognitive changes may alter cadence, language and other characteristics of speech. Mayo researchers are investigating whether AI can detect these signals. Tripathi said cardiology patients are beginning to provide voice samples that can establish baselines against which future changes can potentially be measured. Neurology researchers are also examining whether changes in speech could provide additional indications of cognitive decline.
The concept illustrates where medical AI may ultimately become most transformative. Data that previously had little diagnostic value can potentially become another clinical signal once algorithms become capable of recognising patterns invisible to human perception. The physical hospital itself is also becoming a potential source of medical data. Mayo is undertaking billions of dollars of investment in new facilities and expansion, particularly around its major campuses, and Tripathi described how new patient rooms are being designed with more sophisticated ambient monitoring capabilities.
Instead of relying solely on pressure-sensitive bed alarms for patients considered at risk of falling, sensors could potentially monitor movement more passively and identify changes indicating that intervention may be required. Over time, these systems could produce information about how patients walk, move and recover. Changes in gait, for example, could become another measurable signal for assessing fall risk or deterioration.
The strategic implication is significant for healthcare property. A hospital room is traditionally physical infrastructure containing a bed, medical equipment and connections to hospital services. Increasingly, it could also become a data-generating environment capable of continuously supporting clinical assessment. That could gradually change hospital design. Sensors, computing infrastructure, connectivity, cybersecurity and data architecture may become as fundamental to new healthcare facilities as conventional mechanical and medical systems.
Mayo’s longer-term ambition goes considerably further. Most AI applications today remain specialised tools. One system examines an ECG, another interprets pathology, another analyses an image and another summarises medical records. Each provides a narrow view of the patient. The eventual objective is to connect these separate capabilities and develop a more integrated model of an individual’s health. Instead of analysing the cardiovascular system, pancreas, brain or genetic profile separately, AI could potentially combine multiple sources of information into increasingly sophisticated models of the patient as a whole.
Tripathi described this as moving towards the possibility of digital twins: computational representations that could help clinicians model how an individual patient might respond under different diagnostic or treatment scenarios. Such systems remain an ambition rather than a production capability at Mayo. The organisation is beginning to think about the architecture and interfaces required to combine specialised AI models, but Tripathi explicitly cautioned that the complete digital-twin concept is not currently deployed in clinical practice.
If eventually realised, however, it could significantly change precision medicine. Rather than relying primarily on averages derived from large patient populations, clinicians could increasingly model treatment decisions around the characteristics of an individual patient and compare possible outcomes before choosing a course of action. Turning that research into routine medicine remains the difficult part. A highly accurate algorithm is not automatically a usable healthcare product. It must fit into clinical workflows, present information in a way doctors and nurses can understand, integrate with existing systems and continue performing reliably after deployment.
This is why Mayo increasingly treats AI governance as an implementation mechanism rather than simply a compliance requirement. Tripathi’s role is specifically focused on moving promising systems from research into clinical practice while maintaining appropriate safeguards. The organisation has introduced a risk-based approach. Lower-complexity systems operating in relatively low-risk environments can move through an accelerated process combined with monitoring after deployment. Technologies capable of directly affecting consequential clinical decisions receive more rigorous evaluation before reaching patients.
The approach recognises that applying the same approval process to every AI application could itself become a barrier. A tool summarising administrative information does not necessarily present the same risks as an algorithm influencing cancer treatment. At the same time, Mayo is encouraging clinicians and employees to experiment and develop ideas. The central governance process becomes more important as those experiments approach actual patient care, where evidence, integration, safety and accountability become essential.
This creates an unusual innovation model. Instead of relying entirely on technology companies to supply healthcare AI, Mayo is combining internally developed clinical tools with partnerships involving major technology groups. Its collaboration with Microsoft, for example, extends into advanced AI applications for healthcare. The strategy reflects the economics of the sector. Even a large healthcare organisation cannot match the billions being invested by global technology companies in foundation models and computing infrastructure. What Mayo possesses instead is clinical expertise, medical data, specialist workflows and the ability to test technologies in real healthcare environments.
The emerging healthcare AI ecosystem is therefore likely to depend on partnerships between technology companies and medical organisations, with each contributing capabilities the other cannot easily reproduce. There is also another major force changing healthcare: patients themselves are gaining access to increasingly sophisticated AI. Consumers can already obtain copies of their medical records and upload information into general-purpose AI systems to ask questions about laboratory results, imaging reports or potential treatment options. Tripathi argued that healthcare providers should recognise that this behaviour is happening rather than simply discourage it.
For healthcare organisations, that could fundamentally change the relationship with patients. A person arriving for an appointment may increasingly have used AI to analyse their records, research their condition and prepare detailed questions before meeting the clinician. The challenge becomes helping patients understand where consumer AI can be useful, where it may be unreliable and when professional medical judgement remains essential.
Ultimately, Mayo’s programme suggests that the next phase of healthcare AI will be less about individual algorithms and more about integration. The technology already demonstrates remarkable capabilities in research environments. The harder problem is connecting those capabilities to medical records, hospital infrastructure, clinical workflows and human decision-making in ways that consistently improve patient outcomes.
For Mayo Clinic, that is the standard against which AI ultimately has to be measured. Saving a radiologist time is valuable, but the organisation wants to understand how that saving translates into better care. Automating a nursing task is useful, but the important question is whether patients receive more attention or experience better outcomes as a result. That philosophy provides an important counterweight to the speed of AI development. Healthcare does not need technology simply because it is impressive. It needs technology that produces measurable improvements in how people are diagnosed, treated and cared for.
The transition now underway at Mayo Clinic therefore represents something larger than the adoption of another generation of hospital software. After decades spent digitising medicine and accumulating enormous quantities of clinical data, artificial intelligence may finally provide healthcare organisations with tools capable of extracting much more of the knowledge contained within it. The decisive question is no longer whether AI can discover something useful. Mayo Clinic’s experience suggests the next challenge is whether healthcare organisations can turn those discoveries into trusted products that doctors can use and patients can actually benefit from.
Source: CIJ.World Research & Analysis Team
