Clinician Voices on AI That Actually Helps

When healthcare professionals engage with technology, there is generally one overriding consideration: "Will this help me, or will this get in my way?" It is a brutally simple and unforgiving test, and most technologies designed to aid healthcare have failed it. There is no shortage of well-intentioned solutions that promised much-needed simplicity and added complexity instead.

The issue is not whether AI is powerful. AI is undoubtedly powerful, but this speaks to its capability rather than its intention. It is therefore crucial to differentiate whether AI is powerful because it serves a purpose or whether its power is exhibited because it was built to showcase the capabilities of its underlying technology. Identifying which of the two it is matters, and clinicians can usually tell almost instantly.

In order to appreciate the challenge that healthcare technology faces, it is important to consider that time is literally a clinical resource. A technology that saves the time of healthcare professionals is almost guaranteed to be trusted. However, technologies that consume time, no matter how impressive they are, will be distrusted.

Context is Everything

When judging AI tools, most evaluators adopt criteria that fail to account for the nuances of real-world application. Most evaluators cite performance during a demo in isolation. This is not how a tool should be judged. Just as important is how the tool functions within the real-world care delivery process, coping with all of the disruptions and hand-offs that are absent from the sales pitch. When decisions regarding the use of a tool are made far from the care delivery process, the clinical staff is forced to adapt to a system that is not designed with their needs in mind.

Fit is not a final step in implementation; it is the prerequisite. Healthcare is not a single operating model that a single tool can address. There are numerous specialized environments that each operate independently and are unified only by overwhelming systemic pressure. A tool that does not acknowledge the existence of the diverse operating environments is not helpful to the work, as it was designed for a hospital that does not exist.

Analyzing the Same Functions, Different Results

The requirement for context is concrete. Take the same alerting function and insert it within two different units. For the first, we configured the alerting function with the people expected to engage with it. As a result, it activates infrequently, and when it does, it has meaning, allowing the team to develop reliance on it. For the second unit, the function was activated in its default setting and therefore alerts on everything; as a result, within one week, everyone learned to ignore it without even looking at it. Same system, different use, different result, and the only difference was whether or not anyone thought to tailor it for the location. A function can be classified as useful or nonuseful by the context of a system and the people processing it.

Things That Work

If you examine AI that gets integrated into clinical practice, you will see the same few characteristics, and again, they are not especially about the AI itself.

The first characteristic is that it is integrated into the existing workflow. The most effective tools are those that do not draw attention to themselves. They integrate seamlessly into the processes that clinicians already have, and allow the work to be done without bringing attention to something new. The moment a tool requires a distinct process, it has begun to provide diminishing returns.

The second is that it brings clarity rather than complexity. Optimal AI should help a clinician decipher intricacies by identifying noteworthy patterns or recognizing pertinent risks, thereby streamlining the decision-making process. If it adds complexity, it simply becomes another system to control, as opposed to a solution.

The third is that it honors clinical judgment. The most trusted tools are those that alleviate the administrative burden on a clinician, while still allowing that clinician to make decisions. The intention was never to make the clinician obsolete. Instead, it was to allow the clinician to reclaim their time and cognitive resources to perform those tasks that are exclusively theirs.

These are not technical specifications. They are design choices, and they reveal what a tool was actually built for.

Support, Not Standardize

There is tension and we need to name it. Some technologies are built to support clinicians, and some are built to standardize them, to flatten the variation in how people work into a single approved path. Standardization can look efficient, even as it eliminates the variation of clinicians dealing with real situations that design teams will never understand. Support provides a better means and encourages clinicians to hold their judgment. Standardization removes judgment and rationalizes it as consistency. Clinicians understand the difference from the first day, and respond as expected.

From Usage to Utility

There is a quiet shift happening in how the better organizations judge these tools, and it is worth naming. For years the measure of success was adoption. Are people using it. That question is easy to answer and almost beside the point, because a tool can be used all day and still make the work worse. Usage is especially misleading when a tool is mandated, because then people use it whether it helps or not, and the adoption number climbs while resentment climbs right alongside it. A high usage rate on a tool no one would choose is not a success. It is a workaround everyone is forced to perform.

The harder questions to ask deal with utility. Are we actually helping? The organizations that answer this correctly have started to view the experience of the users of a tool as legitimate data rather than a soft bonus. These experiences contain data that can help organizations uncover insight that traditional metrics may overlook. These organizations rely on clinical leadership to help with tool design. They understand that a tool constructed with the help of frontline staff is likely to meet the needs of the staff. They are also starting to consider whether or not a tool restores time vs. simply automating a task.

Reconceptualizing Success

Almost all dashboards are able to tell you if a tool shows up. Very few can tell you if it provides any relief. This is the gap that most healthcare technology fails to bridge. These technologies appear to be successful in user adoption but fail to assist the very users they were created to serve.

This means that the questions must be structured differently. The point is not simply to determine whether or not a tool was adopted. Rather, the question must focus on what was returned as a result of the tool. The most common answers to this question are time and confidence. Sometimes the answer is simply the relief of a completed shift with fewer, if any, open tasks. Given the level of complexity and the amount of stress in this environment, the most useful evaluation of a tool is based on the experience of using the tool rather than its raw performance. Most of the time, this is the only indicator of whether the tool is functioning as intended.

The Long View

The development of Artificial Intelligence (AI) technologies in healthcare will not be a race to produce the best technology. Rather, it will be an evaluation of how well these technologies understand and integrate into the reality of healthcare. Healthcare professionals are not looking for more technology. They are looking for technology that is appropriate for the setting and helps without being a burden or an obstacle.

The best solutions integrate seamlessly and stay unobtrusive, built to enable the people using them. Their true worth is not in the promises they make. It is in the support they deliver to the caregivers who bear the responsibility of caring for others.

Christopher Hutchins Founder and CEO, Hutchins Data Strategy Consultants