Ambient AI in Healthcare: Cutting Documentation Burden Without Ceding Control

Most of the burnout problem in healthcare is paperwork. In the inaugural episode of The Signal Room, Dr. Barry Chaiken, a physician leader and former chair of the HIMSS Board of Directors, who has authored a book on the future of healthcare, is one of the few people to pinpoint the cause of clinician burnout to something that's usually overlooked: the electronic health record, and the documentation requirements associated with it. He makes a disruptive and uncomfortable point about why the systems were designed the way they were. The primary purpose of healthcare organizations is to document patient care in order to get paid. In a US healthcare system that relies on documentation for payment, there would be no market for a system designed to improve patient care.

This context is important because it helps us understand the role that ambient AI can play in the future. The requirements for documentation in healthcare are unlikely to change in the foreseeable future, without a fundamental change to the reimbursement model. Therefore, as we look to the future, the more appropriate question to ask is not how to remove the burden, but how to shift that burden away from the clinician, and in this regard, ambient AI is the most promising answer that Chaiken has provided.

The Documentation Burden Is the Practical Place to Start

Chaiken has placed ambient listening, which is the first practical use of AI that he has identified, in the practical side of the artificial intelligence balance. The reasoning is based on something that every healthcare system is already measuring: clinician burnout driven by the completion of care documentation.

Ambient AI provides documentation based on the encounter audio, which allows the clinician to step away from the keyboard and engage with the patient. In Chaiken's account, this is more than just a minor improvement. Properly integrated ambient listening technology can document a visit as completely as the clinician would, and often in more detail.

That last point, about completeness, is reliant on proper integration of the technology into the caregiver workflow. The capability is not a property of the tool sitting on a shelf. It is a property of the tool placed inside a workflow designed for it.

A Note the Patient Can Actually Use

Chaiken noted many examples of improvements and enhancements in the quality of care possible due to the improved documentation, rather than just the time saving for the clinician.

The patient's note need not be the clinician's note. It can be modified to a summary that the patient may understand given their language, level of education, profession, and the like. That way, the patient takes home a meaningful document, rather than a clinical record that they cannot understand. The same logic was applied to preparation for the surgery. Preoperative instructions can be tailored to address the patient's individualized needs and prepared in a language that is uplifting and assuring.

The same is true for referrals. When a patient is referred to a specialist, the information that the specialist requires can be synthesized and moved to the beginning of the note, rather than being buried in the content of the note. To illustrate, for a patient going for a hip replacement, the surgeon would want to know the patient's clotting factors and hemoglobin, while the cholesterol level would be less relevant. Chaiken viewed this as a type of personalization that is not feasible for humans to do due to the time and effort that would be required, from patient to patient, but that AI can achieve. It transforms documentation from a burden that the clinician bears to something that benefits the patient and the next clinician in the line.

Where the Risk Lives: Automation Bias

The same conversation refused to let the positive aspects stand alone. The moment AI moves from documentation to serving as a therapeutic or diagnostic assistant, Chaiken called it a specific type of failure: automation bias.

His account displayed a lack of sentimental bias towards the subjects. Clinicians are not absent of humanity. They too can become preoccupied and miss things and make mistakes. In fact, given the nature of humanity, people can be expected to make mistakes. Take a recommendation from an AI, and place it in front of a clinician who is working under those conditions and is experiencing a workflow that was never meant to keep them involved, and the recommendations will go through the process without justification. Since AI can fabricate and provide an answer that is factually incorrect, a process that does not have a manual review has a legitimate problem embedded in it.

The solution he suggested is more complex, and more than a single thing in the process. First, the process should be designed in such a way that a clinician is an active participant in the process in order to not implement automation bias. Second, build in a review step by using AI paired with another algorithm that focuses on the AI's work and the clinician's judgment. Lastly, implement an oversight process. Regularly check the system for actions that are out of the norm, and use those unexpected results to correct the process and the underlying design. These approaches are all required, and are not interchangeable.

Surveillance Without the Stopwatch

Chaiken anticipated the response that word invites. Surveillance sounds like someone looking over the clinician's shoulder, ready to punish. This perspective is unwarranted and has good reason to be distrusted in healthcare.

His reimagining began with two examples. Crew resource management in commercial flight, which permits co-pilots and supporting staff to challenge a captain, has been shown to improve safety in aviation. The same can be said for the operating room, which has been shown to improve safety through the use of a checklist and establishing open permission to anyone in the room to stop a procedure should they find something that is not correct. A measure in his view of surveillance is the opposite of a stopwatch, which shaves a nurse's call from four minutes to three minutes and forty-five seconds and is the exact opposite of how to do it right. Instead, he states it is a way to ensure the work is within reasonable bounds, and surmises that the work may be pushing the bounds for improvement for the work in progress.

He was clear that there is a negative connotation for some with regard to the use of surveillance in for-profit, fund-owned health care practices, which has been implemented to shorten the duration of patient visits. This is the direct opposite of what surveillance is meant to achieve. The focus of surveillance is on the patient, and if that care is overlooked, then the focus has clearly been lost.

Build It With the People Who Do the Work

The thread tying the optimism and the caution together is design discipline. Chaiken returned more than once to a conviction shaped by years of EHR rollouts: technology is only a set of tools, and tools tied to no outcome can absorb enormous time and money while achieving nothing, or making things worse.

To ensure your initiative is unsuccessful, issue orders to the staff before consulting directly with them. If you characterize clinical staff as mere consumers of outputs, you not only risk their trust, but also your initiative. Trust is not given, but earned. Patients also have an important stake in the process, with a say in how AI reaches them through messages and summaries. Collect the information from everyone first, and then make decisions.

How Hutchins Approaches Ambient AI

Our work meets ambient AI where Chaiken locates the real difficulty: not in the model, but in the workflow and oversight built around it. We help health systems design the encounter flow so the clinician stays an active check on the output, define the monitoring that catches drift and automation bias before they reach a patient, and establish the review that treats oversight as a search for better practice rather than a clock to enforce. This is continuous with how we think about responsible AI in healthcare and how we approach designing clinical AI for real conditions, where the same questions of human control and validation decide whether a deployment is safe. These themes run throughout The Signal Room podcast, where clinicians and operators describe what it takes to put AI to work without losing the human judgment it is supposed to support.

Authoritative sources

• FDA — Good Machine Learning Practice for Medical Device Development
• NIST AI Risk Management Framework
• WHO — Ethics & governance of AI for health
• HIMSS
• NEJM AI