Hawse Health operates six medical/dental facility locations across Eastern West Virginia.  In terms of behavioral health services, it has seven clinics, 14 school-based health centers and 13 behavioral health therapists.

Powers said that behavioral health documentation has been unsustainable for a long time, and behavioral health workforce shortages and rising burnout make it imperative to try to improve the situation. He noted that therapists doing after-hours documentation often becomes the norm rather than the exception. 

The problem is not just about notes taking too long, he said, but the clinician is trying to attend to their patient and have a clinical presence, while taking notes. “We need to ask whether technology can reduce cognitive load while preserving clinical judgment and maintaining that therapeutic relationship,” Powers said. “That’s the problem that we really wanted to focus on — more than just let's make it faster.”

Before the implementation of AI, Hawse behavioral health leaders had clear documentation standards. They had regular check-ins with providers to see what the barriers were and what they could do to minimize the friction points.
“We saw some benefit, but we hit a plateau,” Lyon said. “We still had some clinicians who were lagging behind. With us being in schools, a lot of our clinicians will prioritize seeing patients over completing documentation. At the school, they're going to see as many kids as they can while they're there, because that's the only opportunity these kids are going to have to get any sort of therapeutic intervention. So they end up prioritizing that, and then at the end of the day, they've got 13 notes to do.”

In thinking about a technology implementation to help solve this, Lyon said that clinicians are very good at being able to determine when something is being done for them versus to them. A productivity-first rollout is being driven by seeing more patients; a clinician-first rollout is being driven by reducing the burden and hopefully to improve quality in patient care, he said. 

The motive to use AI scribes was “trying to stop the bleeding with documentation burden,” Lyon said. Showing that you're trying to help the clinicians reduce the burden of documentation helps get buy-in, he said. "If the clinician feels like it's something that is either harmful to clinicians or harmful to patients, they're going to be against it, understandably.”

Hawse enrolled 13 clinicians in the program about a year and a half ago. They have done a little over 26,000 AI-assisted encounters since. “We're in rural West Virginia. We didn't have the ability to run any sort of an isolated pilot,” Lyon said. The first phase of the rollout was about a month of clinicians using it as a dictation aid, and then one or two months for consent building. By about six months, all the clinicians were using it with the majority of their patients. 

As they rolled this out, they learned from their mistakes. “Initially, paying attention to informed consent, we provided them with two pages and an explanation of ambient AI,” Powers said. “What we found is that patients and guardians got overwhelmed with that much information. We brought it back down to one page that was both an information and consent page, and then discussing that with them if they had any questions. We're now at about 99% of patients giving consent.”

Noticeable Impact on Document Completion Rates

Lyon spoke about some of the positive impact they have seen so far.

“Before we implemented ambient AI, our same-day completion rate was about 35% — not the worst, but far from like where we wanted to be. As of March of this year, our same day completion rate is now at around 55%. Within 72 hours we're having 85% of documentation complete, including psychosocial evaluations, which traditionally are a bit more documentation-intensive and require more clinician involvement to complete. So now notes are getting completed much closer to real time.”

There's also been a huge narrowing in performance across clinicians, both in number of encounters per month as well as timeliness of documentation, which is not something that they had anticipated. 

He said the improvement has had other positive side effects. “In the last year and a half, we had one clinician out for several months because they had to have emergency open heart surgery. We had another clinician out on maternity leave, and several other things. What was interesting is because the other clinicians were less burdened with documentation, and felt not spread as thin, they were more easily able to absorb that clinician’s patients until they got back, which means that patient access was preserved until their primary therapist returned. It was extremely helpful to our folks who have more severe levels of behavioral health problems and behavioral disorders.”

Powers spoke about the implementation from his own personal perspective as a clinician. “I’ve always been a clinician who is able to get my notes completed in a timely manner. When I began to use ambient AI as a scribe, I don't know that it necessarily helped me to be able to do it quicker, but what it did help me do was to use less brain power for the note and more brain power in the session. In other words, I was able to be more present. It really does help the clinician to focus on the modality that I'm going to use.”

When he reviews the note the scribe has created, Powers realizes the ambient AI documented the session in more detail than he probably would have if he would have tried to take notes and just recalled the session from his own memory.

Lyon described governance and regulation as still the Wild West. “When we started looking into this two years ago, there were not a ton of rules or regulations,” he added. “I think we started implementing this six months before the National Association of Social Workers even put out guidelines on what to look for. Early on our focus was heavily on privacy and compliance with HIPAA, and the fact that any sort of audio capture, even though it's not recording, it is transcribing, it does require careful consent and data handling. The regulatory landscape is still evolving.”