238: Essentia Health’s ACMIO Dr. Dave Kunstman: The Voice Agent Epic’s Agent Factory Built in Under a Day, Why Faster Builds Mean Faster Mistakes, and Why Ambient AI Is Care Redesign, Not Documentation

Subscribe on: RSS | SPOTIFY | APPLE PODCAST | GOOGLE | BREAKER | ANCHOR

On this episode of The Digital Patient, Dr. Joshua Liu, Co-founder & CEO of SeamlessMD, and colleague, Alan Sardana, chat with David Kunstman, MD, Associate Chief Medical Information Officer at Essentia Health, about "The Voice Agent Epic’s Agent Factory Built in Under a Day, Why Faster Builds Mean Faster Mistakes, Why Ambient AI Is Care Redesign, Not Documentation, and more..." Click the play button to listen or read the show notes below.
‍

Audio:

‍

Guest(s):

• David Kunstman, MD, Associate Chief Medical Information Officer at Essentia Health
• Joshua Liu, MD (@joshuapliu), Co-founder & CEO at SeamlessMD

‍

‍Episode 238 - Show Notes:

[00:00:07] Episode preview

[00:04:53] What Epic's Agent Factory actually is — Epic's move from AI that generates information to AI that helps you do work. Dr. Kunstmandescribes it as a visual workflow builder: you take inputs like chart review, apply logic, and let it make decisions and perform actions across documentation, orders, and messaging. "It shifts AI from a tool that you open to one that actually provides workflow support in the background."

[00:06:05] How a post-discharge voice agent got built in under a day — at Epic's build-a-thon in Verona, Dr. Kunstman's group started on whiteboards, identifying which patients needed follow-up and why, and within less than a day had a voice agent that contacted the patient, answered questions, and scheduled the next appointment. The prerequisite was base knowledge: the first full day covered what initiates a workflow in Epic, and at one point there were more Epic developers in the room than participants.

[00:08:03] Why the long-term opportunity for a rural system is scale — one in five Americans lives in a rural area, but only one in ten clinicians does, and 80% of Essentia's population is rural. "It's not really a workforce problem that we're gonna be able to hire our way out of." The goal is to extend the reach of care teams, not replace them.

[00:09:37] Why foundational knowledge becomes more important as AI makes building easier — the hard question shifts from "can we build it" to what should and shouldn't be built. The bottleneck moves to judgment, clinical reasoning, and workflow design. "If you can build faster, you can also build mistakes faster."

[00:11:28] The two risks Dr. Kunstman watches for as tools get easier to use — over-automation, where clinicians accept the AI answer and stop applying judgment, and skill atrophy. His own example: after using ambient with nearly every patient, his dictation has degraded into run-on sentences with missing punctuation, a skill he'd had for 20 years. He also warns that poor tools cost clinician confidence permanently.

[00:13:03] Why AI doesn't devalue analysts — it changes where their value shows up. Analysts know the system triggers, the expected system behavior, where the output belongs, and where clinicians will actually see it. Less time in repetitive build, more in designing, validating, and improving.

[00:14:56] What ambient looks like once it moves past the note — it already surfaces orders to act on, and Dr. Kunstman expects it to start reading context: you're seeing a hypertension patient whose kidney function hasn't been checked in six months, so here's the order.

[00:17:57] How Essentia is approaching voice agents with rural patients — piloting internally with the IS group, nothing patient-facing yet. Because the voices are realistic enough that you'd swear you're talking to a person, the team wants to be upfront that it's AI: "We wanna build that trust so that we don't lose them down the road."

[00:19:09] How CMIOs should prepare for the constraint moving from build to governance — start with the workflow, identify friction points, target the high-volume and high-friction ones, and set clear ownership and measurable outcomes up front. Dr. Kunstman's build-a-thon lesson: this needs a fusion team of four — an operations owner, an analyst, informatics to bridge workflow and build, and a data scientist or AI partner to evaluate performance, safety, and token cost.

[00:21:06] The AI spend conversation health systems haven't had yet — usage today is largely an all-you-can-eat package, but that will change. Epic dashboards show individual AI tool use and effectiveness; Dr. Kunstman wants a roll-up view, and expects finance to look closely once consumption cost arrives.

[00:22:35] Why his ambient question changed from "does it work" to "how do I live with this" — after helping run one of the country's more rigorous ambient evaluations at UW Health, the work is now operational: workflow fit, note style, exam room habits, ease of editing, training. A colleague uncomfortable saying exam findings aloud had shelved the tool entirely; a small change — finish the visit, leave the room, then speak the exam — brought him back.

[00:25:40] How ambient works better as a toolset than a prescribed workflow — Dr. Kunstman had students write their own note and compared it against what ambient produced, a useful learning exercise even though he's clear ambient isn't the gold standard. Adapting the tool to how clinicians actually work opens uses no one designed for.

[00:26:39] Where health systems get ambient wrong — treating it as a documentation project rather than a redesign of care delivery. Faster notes, more presence with the patient, and lower cognitive load are real, but the larger opportunity is care support: ambient orders, reminders, and care opportunities served up in real time. "Faster notes are nice, but better care is really the goal."

[00:28:03] The unresolved question about AI recommendations a clinician declines — if a nudge is surfaced, documented, and not acted on, could it be used against the clinician later? Dr. Kunstman says the conversation has started without a clear answer, and notes ambient recordings are retained roughly 72 hours and locked down from him, the patient, and administrators. His own worry is more immediate: he now spends the visit talking to the patient instead of looking at the screen, and doesn't know how a real-time nudge would even reach him.

[00:30:40] The four-question filter for separating real projects from impressive demos — find the pain before choosing the tool. Is it solving a high-friction problem? Is someone operationally accountable? Do we understand the workflow well enough to improve it? Can we measure whether it actually makes care better? The fourth is the hardest. Without an operational home, it waits.

[00:32:24] How Dr. Kunstman stays close to the work — visiting clinics, shadowing clinicians, observing how work actually happens rather than building off assumptions, then whiteboarding or diagramming current state, finding the pain points and desired state, and only then looking for technology to plug the hole. "AI is a means. It's really not the goal."

[00:33:26] Why the rural context raises the stakes on speed, trust, and safety at once — access barriers are real, patients want to feel known and not processed, and care is harder to reach, so every touchpoint matters. Essentia's footprint takes ten hours to drive across on a good day, so a single appointment can cost a patient hours. One AI ambition is a "patient portrait" capturing needs, communication preferences, digital readiness, and access barriers, so the next visit fits the patient — virtual, in person, or outreach.

[00:36:04] The signal that makes Dr. Kunstman optimistic about AI in rural health — senior leadership at Essentia is actively promoting use, and a recent open AI office hours call drew more than 110 people from across the organization.

‍

‍Fast 5 Lightning Round:

1. What is your favorite book or book you’ve gifted the most?
   Start With Why by Simon Sinek
2. If you could instantly master any skill, what would it be?
   Photographic pattern recognition.
3. Would you rather have Super strength, super speed, or the ability to read people’s minds?
   Mind reading.
4. What is something in healthcare you believe others might find insane?
   Easily July 1st... we take a complex, high-stakes healthcare system and we insert new learners all on the same day.
5. What is the last movie or TV show you saw, and what did you think of it?
   "The Pitt. I really liked how they introduced technology without judgment in it. Think about the ambient message... It was spot on. It was, "try it, understand it, proofread it, and use your good judgment." And boy, that downtime episode, I mean, that really hit home..."

The Digital Patient has been recognized as Feedspot's #1 Patient Engagement Podcast of 2025. Thank you to our listeners for making this happen!