Latest / Becker’s Healthcare Podcast / Building Trust in Agentic AI: Data, Governance and the Human in the Loop

Transcript

Building Trust in Agentic AI: Data, Governance and the Human in the Loop · Becker’s Healthcare Podcast

• 0:00
  - Hello, everyone. This is Erica Spicer Mason with
• 0:02
  - Becker's Health Care. Thank you so much for
• 0:04
  - tuning in to the Becker's Health Care podcast
• 0:07
  - series.
• 0:08
  - Today, we're talking about how health care organizations
• 0:11
  - are moving past AI tools that simply summarize
• 0:14
  - information or make recommendations
• 0:16
  - toward agentic workflows that can coordinate tasks, find
• 0:20
  - information, and take approved actions.
• 0:22
  - And as these workflows become more advanced, their
• 0:24
  - success will depend on the quality of the
• 0:26
  - data and the context behind them.
• 0:28
  - So joining me to unpack what that evolution
• 0:31
  - means for health care is Morgan Beshle, the
• 0:34
  - vice president of products at Refspring.
• 0:36
  - Morgan, great to have you with us. Thanks
• 0:38
  - so much for being here today.
• 0:40
  - Absolutely. Thanks so much for having me.
• 0:43
  - Yeah. Thrilled to have you. And before we
• 0:45
  - get into our conversation on AI, wanted to
• 0:47
  - give you the chance to just share a
• 0:49
  - little bit more about yourself, the work that
• 0:51
  - you're doing in the health care space, really
• 0:53
  - whatever you think would be helpful for our
• 0:54
  - listeners to know.
• 0:56
  - Sure. So I am vice president of product
• 0:59
  - at,
• 1:00
  - RevSpring. I'm responsible for the KIRUS Health business
• 1:03
  - unit product suite.
• 1:05
  - We were recently acquired by RevSpring back in
• 1:07
  - late October twenty twenty five.
• 1:10
  - I have spent my career the last twenty
• 1:13
  - two years building
• 1:15
  - health care products across health plans, health systems,
• 1:18
  - providers.
• 1:19
  - I have really specialized
• 1:21
  - in b to b to c
• 1:24
  - products where you're selling to one buyer, building
• 1:26
  - the experience for ultimately the consumer care manager.
• 1:30
  - So some some interesting
• 1:31
  - design incentive tension there.
• 1:34
  - Some of the problems that I've focused my
• 1:36
  - career on range from
• 1:38
  - helping consumers make sure they take their chronic
• 1:41
  - disease medications
• 1:42
  - as prescribed,
• 1:44
  - helping people transition home from the hospital.
• 1:47
  - But here at RevSpring, I'm focused
• 1:49
  - primarily on care access,
• 1:52
  - which is really a two sided problem we
• 1:55
  - think about every day. One is helping consumers
• 1:58
  - find the right care,
• 2:00
  - while at the same time helping providers, clinicians,
• 2:04
  - health systems
• 2:05
  - make the right care visible
• 2:07
  - to the right patients.
• 2:09
  - We also serve the the health plan side
• 2:11
  - of the market too, and they also are
• 2:13
  - focused on on care access from a different
• 2:16
  - angle in terms of how do you help
• 2:18
  - members and consumers understand their benefits,
• 2:21
  - navigate their network, get to those highest value
• 2:24
  - providers based on what they need.
• 2:26
  - Fantastic. Well, it sounds like you've done quite
• 2:29
  - a range of things in health care, Morgan,
• 2:31
  - and so I think your perspective here is
• 2:33
  - going to be really valuable.
• 2:35
  - And I'm interested to hear kind of where
• 2:37
  - AI is fitting into this picture of your
• 2:39
  - focus work on increasing access to care.
• 2:42
  - So let's go ahead and start high level
• 2:44
  - on that note. You You know, a lot
• 2:46
  - of the conversation around AI right now is
• 2:49
  - focused on getting a system access to data.
• 2:52
  - But I know that making data accessible
• 2:54
  - and ensuring that it's trustworthy data, those are
• 2:57
  - two different concepts.
• 2:58
  - So based on your experience in this space,
• 3:01
  - what does that gap look like inside most
• 3:03
  - health most health systems right now? And where
• 3:05
  - do you think teams,
• 3:07
  - tend to underestimate it the most?
• 3:10
  - Yeah. Great question. I I spend a lot
• 3:12
  - of my week talking to health systems about
• 3:14
  - exactly this.
• 3:16
  - I actually don't know that health systems are
• 3:19
  - underestimating
• 3:20
  - this. I I actually think they're painfully aware
• 3:23
  - of it.
• 3:24
  - You know, the technology,
• 3:26
  - at least in health care
• 3:28
  - I think the
• 3:29
  - technology is is ahead, and that's not really
• 3:32
  - the problem. The data foundation
• 3:35
  - has been a challenge
• 3:37
  - for years before AI.
• 3:40
  - I've spent my crib career building products and
• 3:43
  - technology
• 3:44
  - for payers and systems, always in health care.
• 3:47
  - And what I've always found is the same
• 3:49
  - truth that
• 3:51
  - you can build the most beautiful user interface,
• 3:53
  - the most beautiful and innovative mobile app, and
• 3:56
  - now AI is the same story. Right? It's
• 3:58
  - in this incredible technology that can create just
• 4:01
  - incredible experiences and process,
• 4:04
  - immense amount of information really quickly.
• 4:07
  - But the constraint in all of those types
• 4:09
  - of experiences
• 4:10
  - is if it doesn't have the right context,
• 4:13
  - it doesn't have the right curated validated data,
• 4:16
  - it's not necessarily gonna be able to give
• 4:19
  - you the right answer or give you a
• 4:21
  - validated answer. Right?
• 4:23
  - I think sometimes
• 4:25
  - when I talk to our customers, what I
• 4:28
  - find is
• 4:29
  - the stakes are really different in health care.
• 4:32
  - So we're seeing this incredible adoption
• 4:34
  - of AI
• 4:36
  - across all of our personal and professional lives.
• 4:40
  - But in health care, the stakes of something
• 4:43
  - going wrong are really high. Right? Health care
• 4:45
  - is incredibly
• 4:46
  - personal.
• 4:48
  - People are often seeking health care in their
• 4:50
  - most vulnerable moment.
• 4:52
  - If I don't get routed to the right
• 4:54
  - provider, the right clinician
• 4:56
  - for my care, or I get delayed by
• 4:58
  - six months, that could have an implication
• 5:00
  - for
• 5:01
  - missing the chance for an early diagnosis
• 5:03
  - and ultimately my prognosis
• 5:05
  - and and, you know, how long I'm here.
• 5:07
  - Right?
• 5:09
  - So,
• 5:10
  - you know, the way that I think about
• 5:12
  - this and and talk to our customers and
• 5:14
  - and my product
• 5:16
  - team is I think what health care really
• 5:18
  - needs is health care is is AI with
• 5:20
  - a seat belt.
• 5:22
  - AI with a seat belt. Morgan, I love
• 5:24
  - that.
• 5:26
  - Yeah, everything that you're you're saying here definitely
• 5:29
  - tracks with what we're hearing more broadly from
• 5:32
  - leaders across the industry. And
• 5:35
  - I think to your points about risk
• 5:38
  - in this realm, that feels especially,
• 5:41
  - heightened, I think, or the concern around it
• 5:43
  - feels especially heightened when we're talking about this
• 5:45
  - shift into AgenTic.
• 5:46
  - So I wanna go there next. You know,
• 5:49
  - when organizations
• 5:50
  - are ready to move beyond
• 5:53
  - AI tools that are, you know, just summarizing
• 5:55
  - or recommending,
• 5:57
  - toward those AgenTic workflows that actually take action,
• 6:01
  - How do they decide which part of a
• 6:02
  - workflow to hand over to AI first? Can
• 6:05
  - you walk us through that?
• 6:07
  - Yeah. You know, and here's maybe another
• 6:11
  - depressing hot take is I I don't think
• 6:14
  - that health care is really even ready
• 6:17
  - to go you know, we we've all experienced
• 6:19
  - this rapid evolution of AI in the last
• 6:21
  - six to twelve months. I mean, I I
• 6:24
  - do my job differently than I used to.
• 6:26
  - I even,
• 6:27
  - you know, handle my purse it's a new
• 6:29
  - tool in my personal life. Right? We're all
• 6:31
  - experiencing this massive shift. And I think when
• 6:33
  - it first came on the the scene, it
• 6:35
  - was answering questions or doing research in a
• 6:38
  - really impressive way.
• 6:40
  - And then it was starting to be able
• 6:41
  - to build things, and now it's actually able
• 6:44
  - to take this agentic action.
• 6:47
  - That third piece, I'm seeing a lot of
• 6:49
  - hesitation
• 6:50
  - in the health care market.
• 6:52
  - And part of it goes back to what
• 6:54
  - we just talked about. It's because of the
• 6:56
  - data. The root cause that I see is
• 6:59
  - the data is gated by it could be
• 7:01
  - gated by systems. It could also be gated
• 7:03
  - by humans. So
• 7:05
  - the data, for example, in care access
• 7:09
  - of how do you know which clinician
• 7:12
  - is the right clinician to route a patient
• 7:15
  - to who needs a total knee replacement?
• 7:17
  - You've got
• 7:18
  - two clinicians that both have a primary specialty
• 7:21
  - of orthopedic surgery.
• 7:23
  - They technically went to school for this. They
• 7:25
  - can perhaps do this and are fully credentialed
• 7:28
  - to do this type of surgery.
• 7:30
  - But perhaps one is has become a total
• 7:32
  - expert in ankles, and that's what, you know,
• 7:34
  - they fill their schedules with, and that's primarily
• 7:36
  - what they focus on.
• 7:38
  - Where does that information live inside an organization?
• 7:41
  - It could be trapped inside a system that
• 7:44
  - isn't accessible
• 7:45
  - to the agent or the AI
• 7:48
  - due to authentication
• 7:49
  - and firewalls and things like that. It could
• 7:52
  - also
• 7:53
  - be gated by a human,
• 7:55
  - perhaps a practice or a department that still
• 7:57
  - believes that there's just so much nuance
• 8:01
  - in that scheduling experience that they really need
• 8:04
  - a front desk human to ask the right
• 8:06
  - questions. Do they have a referral?
• 8:08
  - Do they have the right insurance?
• 8:11
  - Have they gotten imaging done in advance or
• 8:13
  - the right labs ahead of the visit?
• 8:15
  - Those are all really important questions. And if
• 8:17
  - they don't live
• 8:18
  - in a structured format that the AI and
• 8:21
  - agent can access,
• 8:23
  - it makes it really hard to for a
• 8:26
  - CIO at a health system to take the
• 8:28
  - leap to allow the agent to take action.
• 8:30
  - So it really comes back to those fundamentals,
• 8:33
  - the foundation of
• 8:35
  - the golden record source of truth for in
• 8:38
  - the care access space as an example, for
• 8:40
  - clinicians, locations, the relationships between them, how they
• 8:44
  - practice,
• 8:45
  - what kinds of services they provide, where they
• 8:47
  - work, what their hours are, how you schedule
• 8:50
  - with them, are they accepting new patients, all
• 8:52
  - those questions, what insurances they take. If that
• 8:55
  - isn't
• 8:56
  - in a in a a golden record structured
• 8:59
  - format that the AI can access, like an
• 9:01
  - MCP server,
• 9:04
  - it's gonna be really hard for anyone to
• 9:06
  - be willing
• 9:07
  - to take the risk
• 9:09
  - to move towards that that agentic phase we're
• 9:12
  - in now because they can't they can't trust
• 9:15
  - the agent. They already know from the starting
• 9:17
  - gate that the agent doesn't have the information
• 9:20
  - it needs
• 9:21
  - to do its job.
• 9:24
  - Yeah. Morgan, appreciate the insights. And it really
• 9:27
  - seems to be leading to where I wanted
• 9:30
  - to to go next in this conversation, which
• 9:32
  - is that concept of having a human in
• 9:34
  - the loop. We we hear that as a
• 9:36
  - common refrain when when we're talking about AI
• 9:38
  - in general, but I think in these earlier
• 9:41
  - days of a gem tick, it seems like
• 9:43
  - something that is essential
• 9:44
  - in this space.
• 9:46
  - So how do you determine
• 9:48
  - where
• 9:49
  - humans should be in a workflow
• 9:52
  - to really, you know, kind of have this
• 9:54
  - seat belt effect that you touched on earlier.
• 9:56
  - You know, is determining where human in the
• 9:58
  - loop is at a risk calculation?
• 10:00
  - Is it more of a regulatory requirement
• 10:02
  - or or something that organizations are looking at,
• 10:06
  - on an organizational level?
• 10:08
  - What I see in a lot of these
• 10:10
  - organizations
• 10:11
  - and and the approach I'd recommend myself
• 10:14
  - is many of these organizations set up some
• 10:16
  - sort of AI steering committee
• 10:18
  - that have some sort of risk calculation
• 10:21
  - framework
• 10:22
  - to evaluate these opportunities.
• 10:24
  - I mean, look, in health care, you've seen
• 10:26
  - you've seen AI adoption successful in certain areas.
• 10:29
  - Right? We've seen AI scribes take off, for
• 10:31
  - example, to the to the extent that they're
• 10:33
  - now starting to be built, you know, really
• 10:35
  - directly into the EHR.
• 10:37
  - We've seen
• 10:38
  - progress in rev cycle management, prior authorization,
• 10:42
  - but a lot of hesitation in other areas
• 10:44
  - like access.
• 10:45
  - There's a lot to think about here. So
• 10:47
  - what I see is is is organizations are
• 10:50
  - applying some kind of risk calculation, and they're
• 10:53
  - thinking about things like,
• 10:56
  - what
• 10:57
  - are the implications if something goes wrong?
• 11:01
  - How egregious is that
• 11:03
  - consequence?
• 11:05
  - How many patients would it happen to? Is
• 11:07
  - it reversible? That's a really important question. Right?
• 11:10
  - So if you have AI and your rev
• 11:12
  - cycle management workflow
• 11:14
  - and perhaps, you know, you you don't collect
• 11:17
  - the right payment option and you have to
• 11:19
  - go back and you lose $10, that that's
• 11:21
  - a reversible
• 11:22
  - type of mistake.
• 11:24
  - It's not a reversible mistake if you delay
• 11:27
  - care for someone for six months
• 11:29
  - and you miss an opportunity for early detection
• 11:32
  - and it changes their their prognosis with their
• 11:34
  - disease. Right? Mhmm. So asking developing a a
• 11:38
  - risk calculation framework of things like
• 11:41
  - blast radius,
• 11:43
  - reversibility,
• 11:44
  - etcetera, I think is is really, really important.
• 11:47
  - And the other piece is,
• 11:50
  - is it measurable?
• 11:51
  - Can I can I measure and detect these
• 11:54
  - things I'm concerned about going wrong so that
• 11:56
  - I can react quickly?
• 11:58
  - Or is this still in a stage where
• 12:00
  - it's not really mature enough to even measure
• 12:02
  - it? Those are some of the ways I
• 12:04
  - see people choosing
• 12:06
  - where and how to put a human in
• 12:08
  - the loop. To be honest with you, I'm
• 12:10
  - not seeing that many examples where there isn't
• 12:12
  - a human in the loop
• 12:14
  - at some point in the process
• 12:18
  - today.
• 12:19
  - Because you might have a human in the
• 12:21
  - loop that's there in real time checking on
• 12:23
  - every recommendation.
• 12:25
  - AI scribe's a great example of that. Huge
• 12:28
  - success. The research shows we've
• 12:30
  - realized major returns in investment in clinician efficiency.
• 12:35
  - But those clinicians are still reviewing
• 12:37
  - the output of those scribes every night and
• 12:40
  - cleaning up those notes. So that's an example
• 12:41
  - of a human in the loop for each
• 12:43
  - time, each check. Right?
• 12:45
  - But you could also have a human in
• 12:46
  - the loop that's looking at maybe a daily
• 12:48
  - trend analysis or report or monitoring
• 12:51
  - alerts when something is detected,
• 12:55
  - that's going wrong.
• 12:57
  - So far, at least in my space, I
• 13:00
  - see that in all cases,
• 13:03
  - health system CIOs are still putting a human
• 13:05
  - in the loop pretty much a 100% of
• 13:07
  - the time as they start to introduce
• 13:09
  - AI into their organizations.
• 13:11
  - Mhmm.
• 13:13
  - Yeah. Morgan, I I appreciate too the the
• 13:15
  - risk calculation framework that you touched on here.
• 13:18
  - I think that's going to be really helpful
• 13:19
  - for folks listening in,
• 13:21
  - especially because it seems like unless
• 13:25
  - folks are living and breathing in this AI
• 13:27
  - space, which many are now, but I think
• 13:30
  - the concept of AgenTic can probably seem a
• 13:32
  - bit amorphous to some still.
• 13:34
  - So just wondering if we can take this
• 13:36
  - point just a little bit deeper. You know,
• 13:39
  - walk us through what it actually looks like
• 13:41
  - when an agent moves from recommending an action
• 13:44
  - to being allowed to
• 13:49
  - organizational standpoint before that handoff happens?
• 13:53
  - Yeah. Absolutely. I I think of an example
• 13:55
  - of, you know, something we're working on at
• 13:58
  - Rev Spring right now
• 14:00
  - that,
• 14:01
  - has both a kind of human in the
• 14:03
  - loop option and an agentic option, and I
• 14:06
  - think we'll eventually get more to that agentic
• 14:09
  - side of the spectrum as it matures.
• 14:12
  - But we started at QIRUS
• 14:15
  - with
• 14:17
  - solving the problem for large academic medical centers,
• 14:21
  - big hospitals where we would take all their
• 14:23
  - clinician and care location data in from disparate
• 14:26
  - sources,
• 14:28
  - transform it into a golden record, and serve
• 14:30
  - it back up to them in an administrative
• 14:33
  - user interface that they could see, touch,
• 14:37
  - look at the golden record, and then make
• 14:39
  - edits to it. Right? And this you know,
• 14:40
  - we brought this to market ten, fifteen years
• 14:42
  - ago
• 14:43
  - and have supported, you know, most of the
• 14:46
  - major health systems across the country
• 14:49
  - with their data.
• 14:51
  - That was innovative at the time. They could
• 14:53
  - go into the platform. They could,
• 14:56
  - you know, categorize
• 14:58
  - physicians. They could upload photos. They could,
• 15:01
  - add more deeper, more granular clinical expertise
• 15:05
  - to their profiles.
• 15:07
  - Now with AI, we can do things like
• 15:11
  - take all the historical
• 15:13
  - clinical codes and claims from how that clinician
• 15:17
  - actually practices and run them through
• 15:20
  - a, you know, AI enabled predictive model of
• 15:24
  - sorts to predict
• 15:26
  - how should their profile be set up based
• 15:28
  - on all the data we have, but also
• 15:30
  - the insights and inferences we can take from
• 15:32
  - the data. For example, I might know let's
• 15:35
  - go back to my orthopedic surgery
• 15:37
  - example. I might know the primary specialty
• 15:40
  - from the structured data,
• 15:43
  - but looking at all this other data I
• 15:45
  - have about how that physician actually practices,
• 15:47
  - I can see that they are a total
• 15:50
  - knee replacement expert,
• 15:53
  - and that's what we should market them for.
• 15:55
  - That's what we should make them visible for.
• 15:57
  - That's the kind of those are the kinds
• 15:58
  - of patients we should route them to.
• 16:01
  - But at this time,
• 16:02
  - we are still putting a human in the
• 16:04
  - loop in terms of we are now pushing
• 16:06
  - recommendations
• 16:08
  - into our platform for that individual clinician,
• 16:11
  - the department head, the organization to review.
• 16:15
  - Over time,
• 16:16
  - I think as
• 16:18
  - organizations
• 16:19
  - gain confidence that
• 16:21
  - when they do review it, there's a very
• 16:23
  - small percentage of the time that they're making
• 16:26
  - changes because the changes they did make, the
• 16:28
  - model actually
• 16:30
  - is self learning and self healing. Right? It
• 16:32
  - starts to get smarter over time. I think
• 16:35
  - what we're gonna start to see is our
• 16:36
  - customers wanting us to automate
• 16:39
  - some of that. They're gonna get alert fatigue.
• 16:42
  - They're gonna be ready to start trusting the
• 16:43
  - model. But I think you have to build
• 16:45
  - that trust first, and you have to give
• 16:47
  - the model the time to go through that
• 16:49
  - training period,
• 16:52
  - to learn and and refine. And you have
• 16:54
  - to show the organization
• 16:56
  - with data that there are very few
• 16:58
  - edge cases. AKA, they're really not getting that
• 17:01
  - much value from that human in the loop.
• 17:04
  - And I don't think we're ever gonna get
• 17:05
  - away from
• 17:07
  - once that agentic solution is in place, I
• 17:09
  - think you're going to have to also
• 17:11
  - marry that with really strong
• 17:14
  - reporting,
• 17:15
  - dashboards,
• 17:16
  - measurement
• 17:17
  - to continuously monitor if that agent is behaving.
• 17:20
  - It's getting you the results that you want.
• 17:23
  - Morgan, I I so appreciate you,
• 17:26
  - shedding some light and specificity
• 17:28
  - on what the agent can look like in
• 17:31
  - action and in workflow. It's so helpful.
• 17:34
  - And I feel like this conversation has really
• 17:36
  - been grounded in kind of where we're at
• 17:38
  - in health care right now with agentic, what
• 17:41
  - needs to be resolved,
• 17:42
  - you know, some
• 17:44
  - opportunities and and great possibilities, it sounds like,
• 17:46
  - if this
• 17:47
  - goes well and if there's strong governance. But,
• 17:50
  - just appreciate the overview. I know this is
• 17:52
  - a topic that's pretty meaty to tackle in
• 17:54
  - a short podcast episode. But,
• 17:57
  - before I let you go, I wanted to
• 17:58
  - make sure I got some of your final
• 18:00
  - thoughts.
• 18:01
  - You know, for a health care leader who's
• 18:03
  - currently evaluating
• 18:04
  - vendors in this space, what's the question they
• 18:07
  - should be asking to ensure that their agentic
• 18:10
  - AI efforts will succeed?
• 18:12
  - And are there any other final thoughts that
• 18:13
  - you'd like to leave them with?
• 18:16
  - Yeah. Well, I have first of all, I've
• 18:17
  - really enjoyed this, so thank you so much
• 18:19
  - for having me. I would say,
• 18:22
  - just remember that we're all
• 18:24
  - figuring this out together. AI is new for
• 18:27
  - everyone. There's no one that has some,
• 18:29
  - you know, advantage on the field.
• 18:32
  - I would be asking
• 18:34
  - my strategic partners,
• 18:36
  - my IT team, anyone I'm working with that's
• 18:39
  - thinking about
• 18:40
  - rolling AI out to solve a problem,
• 18:44
  - I would ask them, how will we know
• 18:46
  - it's working at the three month mark, at
• 18:48
  - the six month mark, at the twelve month
• 18:50
  - mark? This is a question
• 18:53
  - that we ask as product leaders, as product
• 18:56
  - managers,
• 18:57
  - all the time. Right? You start with the
• 18:58
  - end in mind.
• 19:00
  - I think that kind of end in mind
• 19:02
  - thinking
• 19:04
  - helps you work backwards to figure out
• 19:07
  - what you're comfortable with and
• 19:09
  - makes you, you know, make decisions based on
• 19:12
  - the outcome you're looking for.
• 19:15
  - Really great advice. Morgan, thank you again for
• 19:17
  - your time and your insights and,
• 19:20
  - just giving a a a really candid look
• 19:22
  - at where we're at in this space right
• 19:24
  - now and what's ahead. So thank you again
• 19:25
  - for the time.
• 19:27
  - Thanks for having me, and thank you to
• 19:28
  - Becker's for spending some time on such an
• 19:31
  - important topic.
• 19:32
  - Absolutely. We're happy to. And, of course, we'd
• 19:35
  - also like to thank our podcast sponsor for
• 19:37
  - today's episode, RevSpring.
• 19:39
  - Listeners, be sure to tune in to more
• 19:41
  - podcasts from Becker's by visiting our podcast page
• 19:44
  - at beckershospitalreview.com.