In late April, a healthcare executive stood at a podium in a conference room somewhere in Las Vegas with fluorescent lighting, name badges, and the subtle scent of hotel carpet. He said something that the majority of his industry still won’t say aloud. Optum Rx’s chief information officer, Santiago Abraham, acknowledged to a group of leaders in the health system that earlier iterations of AI-powered patient self-service were, in his opinion, mainly insufficient. The new ones are genuinely functioning, he continued. The fact that he drew it at all is what made the moment memorable, and the distinction is crucial.
It’s difficult to ignore how infrequently that level of candor appears in discussions about healthcare AI. The majority of public discourse remains at the level of promise: what AI might accomplish, what models are starting to do, and what a fully customized patient journey might eventually entail. The more difficult accounting—where the tools failed, why organizations became stuck, and what executives wish they had known before they started—is overlooked.
This more difficult accounting is beginning to take place. A first wave of chief AI officers and technology leaders, some of whom have only been in their positions for a year, are learning what it truly takes to implement AI in a real clinical setting across health systems. The solution is far less glamorous than the suggested demos.
The fact that the position isn’t what was advertised is one recurrent theme. According to Karandeep Singh, chief health AI officer at UC San Diego Health, there is a common misconception that his title indicates a desire to quickly and widely implement AI. The truth, he claimed, is more akin to the opposite: establishing credibility by being aware of when to avoid using AI. It’s odd to say that about a position known as Chief AI Officer, but leaders who have been present when pilots stalled, workflows failed, and clinicians resisted have come to a sort of consensus on the matter. It turns out that resistance is rarely just opposition. People who understand exactly what’s at risk frequently have valid concerns.
The unevenness of the data beneath the enthusiasm is striking. Ankit Jain, CEO of Infinitus Systems, presented data at a summit in late April that places the sector in the middle of ambition and execution: 69% of healthcare executives stated that AI is their top priority for 2026, but only about 8% of those companies have AI in production. That figure is higher than 90 in the professional services sector. It’s obvious that something is impeding things, and it’s not a lack of interest. In particular, 82% of pharmacists claim to be familiar with AI. Just 39% of people have used it.
Most of the time, data is the obstacle. In simple terms, Mouneer Odeh of Cedars-Sinai, a health system that has trained more than a thousand employees in AI literacy, stated that while functional prototypes can be created in a matter of hours, obtaining reliable and accurate underlying data takes much longer. He explained that over the course of ten years, thousands of job aids and training materials were reviewed in order to train an AI agent on EHR workflows.
In more direct terms, Alok Chaudhary of VCU Health made a similar statement. “AI is all about data,” he declared. “If you don’t have the great data foundation, AI is not going anywhere.” For the majority of his first year and a half, he built the infrastructure that could eventually make deployment of AI defendable rather than implementing it.
The picture of governance is also lacking. AI has been adopted in some capacity by 75% of health systems, but only 18% of those organizations have formal governance structures in place, and the majority lack both data policies and personnel who can assess what they’ve actually implemented. That ratio is unsettling in some way. Organizations are adopting things quickly enough, but sometimes not quickly enough to regulate what they’ve adopted.
When AI is successful, the results typically don’t match the original plan. One of his initial surprises, according to Ben Shahshahani of Cleveland Clinic, was the amount of value found in operational areas, such as revenue cycle management and documentation quality, as opposed to the clinical applications that receive the most media attention.
Prior authorization support has significantly shortened turnaround times at Optum Rx, allowing clinicians to continue making clinical decisions while eliminating friction from a traditionally sluggish procedure. Nobody anticipated that better self-service technology would result in a 20 percent decrease in live operator calls in just one year. It’s real, too. The industry seems to have spent years undervaluing the most feasible applications of AI while promoting its most spectacular ones.
At the end of his session, Jain presented a frame that is worth pondering. The first websites were digital newspapers with the same content but a different format when the web browser first appeared in the mid-1990s. There was a phone app when the iPhone first came out. Because they required people to truly internalize what the new infrastructure made possible, native uses—those that could not have existed before—took years to emerge. He contended that healthcare AI is at precisely that turning point. The majority of organizations continue to use it to mimic human behavior.
Future developments, such as hyperpersonalized care, AI that follows a patient throughout their entire course of treatment, and systems that identify what humans overlook on a large scale, are still a ways off. Whether most organizations will be prepared for it when it happens is still up in the air. However, the executives who took the lead, persevered, and persevered have at least made the way a little clearer for those who follow them.
