AI EHR features IT leaders say they would turn off

As artificial intelligence spreads through electronic health records, a quieter question is emerging: which AI tools would health systems actually turn off? The answer matters because every major EHR vendor is moving quickly. Epic has more than 150 AI features and enhancements in development for 2026, including conversational search, autonomous coding and AI-assisted charting, in addition to ambient tools already in use. Oracle Health is also expanding its Clinical AI Agent to automate clinical orders such as labs, imaging, prescriptions and referrals. At the same time, third-party ambient scribe companies have spent the past two years connecting directly into Epic, Oracle Health and other EHRs.

Many CIOs and informatics leaders say embedding AI into the EHR can give clinicians meaningful time back. Muhammad Siddiqui, CIO of Reid Health in Richmond, Ind., told Becker’s in February that AI outside the EHR is like having a capable assistant who sits outside the room. But he also said embedded AI is not automatically better. If the AI is noisy, inconsistent or hard to govern, it can quietly undermine trust, he said. Clinicians are quick to disengage when tools feel unreliable or create rework.

That concern points to a less visible side of EHR AI: the features that look promising at first but end up disappointing users. According to Deepti Pandita, MD, vice president of clinical informatics, chief medical informatics and AI officer, and associate professor of medicine at UCI Health in Orange, Calif., chart and visit summarization is one of those tools. She said provider peers often report that chart summary AI, whether it comes from an EHR vendor or an ambient scribe company, is the feature they like the least, even if they were initially excited about it.

Dr. Pandita said the patient-level clinical summary is often not useful because it is usually generated primarily from aggregated chart content such as the problem list, history and narrative note text. What clinicians rely on, however, are structured encounter-level decision outputs, including medication changes, newly placed orders or referrals, and the specific follow-up or return plan. When those pieces are missing or blurred, the summary may sound plausible, but it is light on context.

In her view, the weakness is not simply that the summary is incomplete. It can also miss the parts clinicians most need in the moment: what changed today and what was decided today. That means clinicians still have to reread the actual visit documentation to find the missing plan, history of present illness and assessment nuances. Instead of saving time, the summary becomes extra work, she said.

That same concern about extra work applies to another class of EHR AI tools: alerting. Usman Akhtar, MD, associate vice president and chief medical informatics officer at VHC Health in Arlington, Va., told Becker’s that he would turn off AI-driven alerts that contribute to alert fatigue. Alert fatigue in the EHR predates generative AI by more than a decade, and adding AI-generated recommendations on top of the existing alert burden, without redesigning when and how those alerts fire, is the failure mode, he said.

Dr. Akhtar said that if he could eliminate one AI-driven EHR feature, it would be any tool that turns intelligence into an interruption. Broad AI-generated alerts or recommendations can surface at the wrong time, with too little context, and add noise instead of clarity. Clinicians do not need more digital nudges, he said. They need technology that quietly removes friction, reduces clicks and gives them time back. In his words, AI should feel like a trusted assistant, not another pop-up demanding attention.

The concerns raised by Dr. Pandita and Dr. Akhtar are not really about AI failing to do what it was designed to do. A summary that faithfully reflects the chart is performing its task, and an alert that fires reliably is also doing its job. The problem is a mismatch between what the tool outputs and what the clinician needs at that specific moment.

For a summary, the missing piece may be the one line that changed since yesterday’s note. For an alert, the issue may be a recommendation that appears mid-task instead of at the point of decision. In both cases, the result is the same: the tool may be technically functioning, but it does not fit the clinical workflow well enough to feel useful.

That disconnect helps explain why AI embedded in the EHR is attracting both enthusiasm and skepticism. Health systems want tools that help clinicians spend less time clicking and more time caring for patients. But leaders also appear to be drawing a line at features that interrupt, overwhelm or force clinicians to do the same work twice.

The broader EHR market is still moving ahead. Epic, Oracle Health and ambient scribe vendors are all pushing deeper into AI-enabled workflow support. Yet the experiences described by health IT leaders suggest that adoption alone is not enough. A feature that sounds useful on paper can still be turned down if it does not deliver the specific information clinicians need, when they need it, in a format that fits the task.

In that sense, the tools that health systems would turn off are also a reminder of what they want most from AI in the EHR: relevance, timing and governance. If a feature adds friction, lacks context or creates a new layer of rework, it loses the time-saving promise that made it attractive in the first place.

For now, the clearest caution from informatics leaders is that embedded AI should reduce burden, not shift it. When chart summaries omit the key decisions of the visit, or alerts interrupt without context, the technology can feel like one more thing clinicians have to manage. When it works well, it should fade into the background and support the clinician’s next step.

Becker’s notes that the conversation is still evolving as EHR vendors and third-party developers continue to add features. But the message from the leaders quoted in this story is straightforward: not every AI capability deserves to stay on. Some tools may be valuable only if they are redesigned to match clinical reality more closely.

That is why the question of what to turn off is becoming just as important as the question of what to turn on. In a crowded EHR environment, health systems appear to be paying close attention to which AI tools help clinicians move faster and which ones simply create another interruption.