Building AI That Works Where Care Happens

AI made big promises this year. AI in healthcare created new resources for ambient documentation, virtual patient engagement, and synthetic trial populations. Technology became fancier with every presentation. New technology often starts with a promise and then the work begins. Progress does not happen on a stage or in the headlines. Progress happens in the quiet places, in the operational structure that integrates resources and balances workflow, staffing, and other variables.

These quiet places are where the real work is done. They determine what care a patient receives and when they receive it. Instead of asking what AI can do, the better question to ask is where AI is most useful. The first question asks a vendor to dazzle you. The second question demands that you know your system and direct AI to the most pressing problem.

Because so much friction exists in backlog offices, AI is most useful in those areas. When used thoughtfully, AI streamlines processes for claims adjudication and appeals and addresses gaps in patient referrals to other healthcare providers. AI should not be used to eliminate staff who perform these healthcare functions. The goal should be to accomplish low-value tasks and reserve high-value work where clinical judgment and expertise are most needed.

The usefulness of ambient technology can next be examined in clinical documentation. For years, medical professionals had to divert away from their patients in order to interact with screens. Now, with ambient technology, physicians can remain in front of their patients as the notes are taken in the background. The automation of clinical documentation is bringing improvements in the accuracy of health records, decreasing the burden of documentation on clinicians, and ensuring that clinicians can remain with their patients during the documentation process.

The third use is in population analytics, and here the function takes precedence over the math. More sophisticated personas to be included on an analytics slide is not the goal. Rather, the focus is on delivering the most appropriate and necessary care to the patients who need it the most, especially in vulnerable and underserved communities. When employed as a tool to maximize social good and impact, population segmentation can be a useful method to locate those who are most in need of the help and services.

Why So Many Pilots Stall

The majority of failed healthcare AI implementations stem not from poor model quality. Their failures occur because ownership of the problem was not assigned, the model was built in a silo, and success was measured against model performance rather than the improvement to practice. A pilot can pass every single test and not change anything. This situation can occur because the missing component was not technical, but rather the linkage of the tool to the people that the tool was supposed to help. If this is not designed properly, even the best model can become a much admired yet unused demo.

The Invisible Work That Makes AI Real

Most valuable AI is the AI that goes completely unnoticed. Value is created by the AI that is integrated to the already existing workflows. Progressing to the point of genuinely realizing value, be it financial or mission-related, from a given set of data is dependent on the purposeful infrastructure built by people. The tools that people build are important, but more important are the decisions that the organization makes to set a problem, assign ownership, and define success.

The aim was never for us to have control, but rather for us to have clarity, when our enterprise analytics framework was established. We invited heads of clinics, operations, and finance to decide what issues we should focus on and why. We did not introduce analytics and AI as new innovations lacking purpose. We described them as operational tools, with potential outcomes on issues such as throughput, workforce, care coordination and revenue cycle integrity.

The analytics framework established a new posture with an important distinction. Analytics could no longer be viewed as a finished product. Rather, teams began to evaluate analytics as a platform that enabled individuals to accomplish their objectives. Subsequently, teams became embedded partners as opposed to a factory for reports. They collaborated with business executives to transform pressing problems into actionable work.

It may not be readily apparent, but there are a number of steps that are actually required to carry out an incremental practice like embedding. For example, an analyst must be present during operational meetings for a duration of time sufficient to understand why a number changes, as well as to learn the workflow enough to understand the difference between a routine exception and a truly significant one. It is also important to develop sufficient trust so that the clinical lead of the meeting feels comfortable speaking to the real issue, rather than a polite version that would otherwise be discussed. None of this is a part of a tool evaluation, but it actually determines if a tool is utilized.

Redefining the Meaning of Success

Success required redefining the metrics. A successful use case was not the construction of a model that worked in theory. Measuring success was the reduction in the number of denials, a note completed in less time, a transfer that was less time consuming and disruptive, a patient experience that improved. When success is measured in that way, the discussion is no longer focused on the sophistication of the technology, but on the impact of the technology on the people who work within the system.

There are no shortcuts to that value. The work that creates it is invisible to everyone, the iceberg beneath the surface, the building of trust and the translation between people who do not speak the same language. That is what turns the abstract idea of artificial intelligence into the concrete use of a technology within a system.

What is Practical will be More Important than What is Powerful

AI will not change the practice of healthcare because of its power. AI will alter the practice of healthcare when it can become more practical and can integrate more with the actual delivery of healthcare and not with the envisioned delivery articulated within a demo. The most advanced technology cannot rectify the issues created by ineffective leadership. Sustainable improvements within healthcare practice will always be from people and not technology.

The next time someone wants to know how AI could be beneficial, do not discuss the tech. Instead, ask the following two questions. What problems do we actually want to solve, and who will have the desire to solve it? The responses will be much more informative about the likelihood of success than a comprehensive list of the features.

Christopher Hutchins Founder and CEO, Hutchins Data Strategy Consultants