Latest / Becker’s Healthcare Podcast / AI Powered Revenue Cycle Transformation with Doug Proctor of Candid Health

Transcript

AI Powered Revenue Cycle Transformation with Doug Proctor of Candid Health · Becker’s Healthcare Podcast

• 0:00
  Hi, everyone. This is Brian Zimmerman with Becker's
• 0:02
  Healthcare. Thank you so much for tuning into
• 0:04
  the Becker's Healthcare podcast.
• 0:06
  Joining me for today's discussion on AI, the
• 0:09
  revenue cycle, is Doug Proctor, COO and cofounder
• 0:12
  of Candid Health. Doug, thank you for being
• 0:14
  here today. Thanks for having me on. To
• 0:17
  kick us off here, go ahead and and
• 0:19
  and share a bit about yourself and and
• 0:20
  and and the work you're doing so so
• 0:22
  listeners can sort of appreciate where you're coming
• 0:24
  from as we move through the conversation today.
• 0:27
  Yeah. For sure. So first off, thanks again
• 0:29
  for having me on.
• 0:30
  So I am our chief operating officer. I'm
• 0:32
  one of the cofounders here at Candid Health.
• 0:34
  We are a sort of end to end
• 0:36
  revenue cycle automation platform. We have hundreds of
• 0:39
  provider customers across dozens of medical specialties
• 0:42
  that we support with everything from claim generation
• 0:45
  to claim submission, remit processing, financial ledgering,
• 0:49
  sort of the full suite of reporting and
• 0:50
  AR workflows.
• 0:52
  And really what powers what we do is
• 0:54
  sort of a
• 0:55
  a fully configurable
• 0:57
  rules engine for users to be able to
• 0:59
  manage the logic
• 1:00
  for all the different constituent pieces of claim
• 1:02
  generation and remit processing. And then we deploy
• 1:05
  AI and machine learning in very targeted and
• 1:06
  specific places where we think it creates a
• 1:08
  lot of value for
• 1:10
  really helping to do one of three things.
• 1:11
  One, increase net collections for our provider customers
• 1:15
  through reduction of denials and and so forth.
• 1:18
  Number two,
• 1:19
  reduce their cost to collect, automating away a
• 1:21
  lot of manual human work. And number three,
• 1:24
  giving, like, a scalable technical foundation where your
• 1:26
  organization can just get better and better. In
• 1:28
  terms of my background, sort of how how
• 1:30
  I fit in. So, I work day to
• 1:32
  day supporting our product teams, our our customer
• 1:35
  teams, and our people teams. Also sit on
• 1:38
  our board of directors with Nick Perry, who's
• 1:39
  my cofounder,
• 1:40
  and our CEO. And prior to Candid, Nick
• 1:43
  and I actually first met at Palantir. Nick
• 1:45
  led a lot of Palantir's commercial health care
• 1:47
  work, and I led a lot of Palantir's,
• 1:49
  work in the military and defense space, which
• 1:51
  actually has lots of fun, interesting similarities, and
• 1:54
  I'm sure we could talk about it some
• 1:55
  other time. But that's a bit about me
• 1:56
  and what we're up
• 1:57
  up to at Candid. Yeah. I I appreciate
• 1:58
  it. And if if any of those fun
• 2:00
  similarities come to mind, please, by all means,
• 2:02
  Doug, share share them with our listeners. But
• 2:05
  wanna wanna dive in here first and think
• 2:06
  about, you know, as, you know, clear from
• 2:10
  your background, the work you're doing, there's so
• 2:12
  much opportunity
• 2:13
  in applying AI to the revenue cycle right
• 2:15
  now. So curious to get, I guess, your
• 2:17
  high level thoughts of where the technology stands
• 2:20
  today. I know so many leaders are investing
• 2:22
  heavily in this space.
• 2:23
  Where is it driving meaningful value? You talked
• 2:26
  about sort of the,
• 2:27
  you know, helping reduce cost to collect, yielding
• 2:29
  some other benefits. But curious where what's the
• 2:31
  state of this technology, AI, and RCM today,
• 2:34
  and where is the value really coming through?
• 2:37
  Yeah. It's a it's a good question.
• 2:39
  I guess, first off, I would just say,
• 2:40
  at Candid, we are incredibly bullish about
• 2:43
  AI's sort of opportunity
• 2:45
  for driving really significant productivity
• 2:48
  improvements in in health care and specifically in
• 2:51
  RCM. I guess before answering the question, I
• 2:53
  would just kinda define some terms because I
• 2:54
  feel like in today's world, people say I
• 2:57
  AI. They might mean a lot of different
• 2:58
  things.
• 2:59
  When we talk about AI,
• 3:01
  I think we're specifically thinking about sort of
• 3:03
  generative AI, large language models, agents, stuff like
• 3:07
  that.
• 3:08
  But in sort of looser terms, we definitely
• 3:09
  are doing some interesting stuff there as well
• 3:11
  as just in the context of
• 3:14
  classic and sort of conventional machine learning,
• 3:17
  like clustering algorithms,
• 3:19
  just looking at patterns in very large volumes
• 3:21
  of structured data, as well as just some
• 3:24
  classic kind of automation of, like, defining business
• 3:27
  logic and letting machines do what machines do
• 3:29
  really well. In terms of where I think
• 3:31
  we're seeing AI, sort of focusing in particular
• 3:34
  on that agentic and gen AI side,
• 3:36
  I think the places where we're seeing a
• 3:38
  lot of momentum and a lot of traction
• 3:40
  are,
• 3:41
  especially around places where there are relatively low
• 3:44
  complexity
• 3:45
  routine operations that human users are having to
• 3:47
  do a lot of times over and over.
• 3:50
  And I think it's even a little bit
• 3:51
  less about, like, fully
• 3:53
  sort of automating away the workflow for the
• 3:56
  human, but actually just letting
• 3:59
  AI agents or large language models
• 4:01
  operate in a way where they can really
• 4:03
  accelerate the human workflow. So so so to
• 4:05
  give you a concrete example of something that's
• 4:06
  deployed across
• 4:07
  all of our customers at Candid, whether or
• 4:09
  not our customers know it, because it's sort
• 4:11
  of baked in and pretty seamless,
• 4:13
  eligibility response.
• 4:15
  Whenever you're running an eligibility check,
• 4:17
  you get this sort of very messy and
• 4:19
  noisy response that comes back. And a lot
• 4:21
  of times when you have coordinate bay coordination
• 4:23
  of benefit issues, figuring out which insurance company
• 4:25
  to actually send the claim to can be
• 4:27
  a pretty laborious process. There are many, many
• 4:30
  different instantiations of this shape of problem. But
• 4:33
  But essentially, what we're doing is look at
• 4:34
  the first eligibility response, have a large language
• 4:37
  model, automatically check that response for a variety
• 4:40
  of other data elements, which if present, maybe
• 4:42
  you wanna go ahead and run another eligibility
• 4:44
  check based on some updated information. And then
• 4:47
  based on that response, maybe you wanna actually
• 4:48
  update which insurance company you're gonna go ahead
• 4:50
  and send that claim to. I think it's
• 4:52
  a good example of a place where a
• 4:54
  human user could easily do that. It just
• 4:56
  takes time. You're just, like, detail oriented, pouring
• 4:58
  through a bunch of data that comes back,
• 5:00
  and it's sort of just a bunch of
• 5:02
  sort
• 5:03
  of messy if then logic. I think that's
• 5:05
  a really great example of places where AI
• 5:07
  can be super helpful.
• 5:09
  It's relatively low consequence as distinct from, say,
• 5:12
  like, determining which ICD 10 code or CPT
• 5:15
  code to assign based on medical records, which
• 5:17
  sort of higher consequence. But it ends up
• 5:20
  actually taking a lot of the time of
• 5:22
  human users to actually, like, go through and
• 5:24
  do those types of workflows. So when we
• 5:26
  think about sort of in our sort of
• 5:28
  product where we're seeing the greatest opportunity right
• 5:30
  now, I think it's examples of things like
• 5:32
  that, but then also just accelerating the feedback
• 5:35
  loop. So, you know, if you're a user
• 5:37
  in Candid and you're pressing buttons to do
• 5:39
  a workflow,
• 5:40
  behind the scenes, Candid's also, first off, gonna
• 5:43
  be logging sort of all of the transactional
• 5:45
  work that's happening,
• 5:46
  running it through sort of a classic machine
• 5:49
  learning clustering algorithm,
• 5:50
  and then translating that into a large language
• 5:53
  model,
• 5:54
  which is actually generating recommendations,
• 5:57
  for rules that should be deployed within the
• 5:59
  system
• 6:00
  to automate away the need for that human
• 6:02
  work to be done in the first place.
• 6:03
  So I think those are two different examples.
• 6:05
  One's kind of that high volume,
• 6:07
  relatively low complexity operational work, and then, secondary
• 6:11
  is kind of around accelerating the feedback loops
• 6:13
  through insights for users. I would say I
• 6:15
  think there's enormous, enormous opportunities.
• 6:18
  So there's there's a lot of ground to
• 6:19
  be taken.
• 6:21
  Yeah. So so really some really great specific
• 6:23
  callouts for from you there around those opportunities.
• 6:26
  Curious here, you you mentioned, you know, sort
• 6:28
  of workflows here and and infrastructure,
• 6:31
  I guess, is one way to think about
• 6:33
  it. I I I've heard a lot of
• 6:34
  times when an AI solution,
• 6:35
  an an an adoption
• 6:37
  falls flat. It's not necessarily because of,
• 6:40
  the AI itself, but because the the workflows,
• 6:43
  the infrastructure
• 6:44
  of healthcare was not built for AI. Correct?
• 6:47
  So so how can folks start to do
• 6:49
  that work in terms of building the right
• 6:51
  foundation
• 6:52
  for to to support AI success because, you
• 6:54
  know, so many of these these workflows, these
• 6:57
  the infrastructure here
• 6:58
  predates AI by a lot. You know,
• 7:01
  a hospital, for example, can have a high-tech
• 7:03
  robot AI in a fax machine. Right? So
• 7:05
  there's just stuff that's just sort of a
• 7:07
  a a mix of advancements. So I I
• 7:10
  curious to get your thoughts on that.
• 7:12
  Yeah. I mean, I think there's it's a
• 7:13
  it's a good question. Like, how do you
• 7:14
  actually make sure that this stuff is valuable,
• 7:16
  and that it's not just some expensive pilot
• 7:19
  or a free pilot that's expensive because it's
• 7:21
  a lot of your time or your team's
• 7:22
  time? I think it starts with having a
• 7:24
  really clear understanding of what is the problem
• 7:27
  you're trying to solve, what is the outcome
• 7:28
  you're going for, and sort of what are
• 7:30
  the steps that you're gonna pursue to get
• 7:32
  there. So, you know,
• 7:34
  AI is not gonna be some magic box
• 7:36
  that figures out exactly what you want it
• 7:38
  to do and comes back and tells you
• 7:40
  and does it for you without needing any
• 7:42
  prompting. I think anyone who's played with, you
• 7:44
  know, OpenAI, ChatGPT,
• 7:46
  anything like that, like, it can go a
• 7:48
  lot of directions, and it might be a
• 7:49
  really compelling kind of first approximation of a
• 7:53
  thing. But when you're doing these fairly complex
• 7:55
  and high consequence
• 7:56
  end to end operational workflows in health care,
• 7:59
  knowing what problem you're solving is really important.
• 8:01
  That's the first thing. And so
• 8:02
  having focus and clarity, like, we are a
• 8:05
  very metrics oriented bunch at Candid. And so
• 8:08
  anytime we're thinking about any sort of, like,
• 8:10
  machine learning or AI or just classic automation
• 8:13
  or software development project, what's the metric you're
• 8:16
  trying you're trying to move? That's point one.
• 8:18
  Point two is understanding, like, the prompting or
• 8:21
  configuration or sort of data foundation that it's
• 8:23
  working with. And so, you know,
• 8:26
  from day one at Candid, we've really been
• 8:27
  building this platform from the ground up with
• 8:30
  an eye towards how do we deploy machine
• 8:32
  learning, AI, and automation
• 8:34
  in a way that's really impactful for our
• 8:36
  organizations.
• 8:37
  And the way we started was, let's rebuild
• 8:39
  the entirety of the EDI layer. Let's rebuild
• 8:41
  the API layer, the data models,
• 8:44
  instrument the end to end workflows.
• 8:46
  And what that's let us do is really
• 8:48
  nicely constrain the problem space
• 8:50
  for AI agents and and ML as well
• 8:53
  to be able to plug in in very
• 8:54
  specific ways. It's as much about saying,
• 8:56
  like, what data do you provide and what
• 8:59
  instructions do you provide is as important as
• 9:01
  saying, like, what is the constraint that you're
• 9:03
  imposing,
• 9:04
  for this thing to not be able to
• 9:06
  do? What actions can it perform? What actions
• 9:08
  can't it perform? What data can it access?
• 9:11
  What data is it not accessing
• 9:13
  inside of the system, outside of system? And
• 9:15
  I think that's the thing where people kinda
• 9:17
  if if you're trying to enter this problem
• 9:19
  space as a very AI native top down,
• 9:21
  like, let's just deploy a large language model,
• 9:24
  but you haven't done the hard plumbing work
• 9:26
  to actually understand
• 9:28
  kind of the data integration
• 9:29
  and, like, the actual unit operations of the
• 9:31
  workflow. I think it's really easy to end
• 9:33
  up in a situation where you're running an
• 9:35
  LLM pilot, and it looked great on the
• 9:38
  surface. But when you get into the day
• 9:39
  to day, it's just not yielding. And I
• 9:41
  I think usually that falls down because there's
• 9:43
  either lack of clarity and purpose
• 9:45
  or some lack of, like, actually understanding the
• 9:48
  fundamental
• 9:49
  sort of workflow and kind of the data
• 9:50
  that needs to go in. Because if it's
• 9:52
  again, I think everyone would have heard this
• 9:54
  saying at this point, but if it's garbage
• 9:55
  in, it's garbage out. So if if there's
• 9:57
  not good prompting or configuration layer or any
• 9:59
  sort of orchestration engine behind it, you're not
• 10:02
  gonna be able to scale the thing, which
• 10:03
  is I think really where the promise lies
• 10:05
  is the ability to use new technologies to
• 10:07
  scale human operations so that you can be
• 10:09
  more efficient and
• 10:11
  have people focus in kind of more valuable
• 10:13
  places.
• 10:14
  Yeah. And what you're describing there to me
• 10:16
  also strikes me as sort of it you
• 10:18
  you've got the technology. You've got sort of
• 10:20
  the the workflows, the infrastructure shift. But doing
• 10:22
  that work to to get that adoption, right,
• 10:25
  requires just a fundamental
• 10:27
  mindset shift, right, of the of the human
• 10:29
  beings involved in making some of these decisions.
• 10:31
  So
• 10:32
  with that in mind, I I guess, what
• 10:34
  is a a piece of guidance or advice
• 10:36
  you would offer to
• 10:37
  RCM leaders, health IT leaders trying to really
• 10:41
  get true value out of out of AI,
• 10:43
  practical value.
• 10:44
  How how can they pursue that? What what
• 10:46
  advice would you give them?
• 10:48
  Yeah. I mean, honestly, the first would be
• 10:52
  trust your intuitions. Like, if you've been doing
• 10:54
  this work for five years, ten years, twenty
• 10:56
  years, thirty years,
• 10:57
  and you're talking about using AI and you
• 11:00
  don't understand,
• 11:01
  like, how it's actually gonna do what they're
• 11:04
  saying it's gonna do, like, you should make
• 11:06
  sure the conversation doesn't move forward. I think
• 11:08
  my experience might take, and I think our
• 11:11
  view at Candid is it's really important to
• 11:13
  be able to show your work.
• 11:15
  And if you're talking to an organization that's
• 11:17
  planning to use AI and you're a biller
• 11:19
  and they haven't explained to you how it
• 11:20
  works, like, I would hit the pause button
• 11:22
  because
• 11:23
  it's actually not that complicated, and you should
• 11:25
  be able to walk through the unit operations.
• 11:27
  Even if the algorithm itself is complex, like,
• 11:29
  what it's doing and how, I think, is
• 11:30
  actually very explainable.
• 11:32
  So that'd be number one is just, like,
• 11:33
  trust your intuitions,
• 11:35
  and make sure that it's not some sort
• 11:37
  of, like, magic box that you expect
• 11:40
  huge outcomes to come out of without, like,
• 11:42
  really understanding how it works. The second is,
• 11:44
  I think, making sure that there's kind of
• 11:46
  both near term and long term thinking and
• 11:48
  strategy.
• 11:49
  I think,
• 11:51
  one of the things that we see from
• 11:52
  time to time is
• 11:54
  and and and, honestly, this is probably less
• 11:55
  true today than it was maybe a year,
• 11:57
  year and a half, two years ago.
• 12:00
  You know, a year year or two ago,
• 12:01
  I think you would see a lot of
• 12:03
  people just like deploying, you know, random AI
• 12:05
  use cases without really any forethought about where
• 12:07
  it leads.
• 12:08
  But in the cases where I think we
• 12:10
  see the greatest impact, it's really saying, okay.
• 12:12
  What's the initial proof point or pilot, and
• 12:14
  what does it look like at scale when
• 12:16
  we can roll this thing out? And I
• 12:18
  think where organizations are being thoughtful about kind
• 12:21
  of preparing the ground the ground and the
• 12:23
  foundation
• 12:24
  targeted deployments against specific opportunities, whether that's a
• 12:28
  metric you wanna move in a certain direction
• 12:30
  or a set of people doing work that
• 12:31
  you wanna kinda reduce and reorient.
• 12:34
  I think those are the two big things
• 12:35
  is, like, number one, trust your instincts. And
• 12:37
  number two, be thoughtful about kind of, are
• 12:40
  you in a position does your current technology
• 12:42
  stack actually allow you to be in a
• 12:44
  position to deploy some near term AI use
• 12:47
  cases and then actually scale it up? Because
• 12:49
  it kinda sucks if you do a lot
• 12:50
  of work, pilot it, people are pumped, but
• 12:52
  then you don't actually have the ability to,
• 12:54
  like, scale the thing for whatever reason that
• 12:56
  might be, whether it's technical or operational.
• 12:59
  Right. And that that failure to scale then
• 13:01
  will hinder future buy in for future projects
• 13:03
  too. Right? If you fall flat on your
• 13:05
  face on one, you've got it's gonna be
• 13:07
  harder to get that trust a a second
• 13:09
  time. Correct? Yeah. And these like, the other
• 13:11
  thing is these problems are just so interconnected.
• 13:14
  I think one of the the sort of
• 13:16
  real artistry components to this whole thing is
• 13:19
  picking the problem. So how do you make
• 13:21
  sure that you're picking a problem that's well
• 13:23
  suited
• 13:24
  for this technology,
• 13:26
  and really being problem oriented. I think sometimes
• 13:28
  there's also a tendency to just sort of,
• 13:30
  like, take a technology and say, okay. Cool.
• 13:32
  Like, we're gonna lead with tech. But our
• 13:34
  our philosophy is very much, like, start with
• 13:36
  the problem,
• 13:37
  go get, you know, on the ground in
• 13:39
  person with the user,
• 13:41
  help them really understand, like, how the technology
• 13:44
  might kind of shape their day to day,
• 13:46
  and just be smart about, like, the order
• 13:47
  of operations in which you're choosing to kinda
• 13:49
  go after different problems.
• 13:51
  Yeah. Thank you. Thank you so much, Doug.
• 13:53
  It's it's been a pleasure having you on
• 13:54
  the podcast. The way I guess, I'll phrase
• 13:56
  our our final question. Is there anything
• 13:59
  we didn't get to that you wanna be
• 14:01
  sure to share for listeners? Or conversely, is
• 14:04
  there something you'd just like to reemphasize before
• 14:06
  before we let you go?
• 14:08
  Yeah. I mean, I think as I reflect
• 14:09
  on you know, we're six years in at
• 14:11
  this point at Candid.
• 14:13
  And I think one of the things that
• 14:16
  I would just say as a reflection is
• 14:18
  we're kinda over that initial hump of AI.
• 14:20
  And I think a lot of people have
• 14:21
  run some pilots and maybe it's proved out
• 14:24
  value, maybe a lot of value, in some
• 14:25
  cases, maybe less.
• 14:27
  And I I think the thing I would
• 14:28
  just say is this,
• 14:30
  revenue cycle management,
• 14:32
  US health care system, like, it is a
• 14:33
  foundational
• 14:34
  societal level problem. It's like a crazy amount
• 14:37
  of
• 14:38
  inefficiency
• 14:39
  and waste.
• 14:40
  These are solvable problems.
• 14:42
  And I personally would say that I'm incredibly
• 14:45
  bullish on
• 14:46
  AI's ability to really move the needle. I
• 14:49
  think the key is actually thinking about AI
• 14:52
  in the context of the broader technology landscape.
• 14:55
  And so a lot of what we've learned
• 14:56
  thus far is that AI is not like
• 14:58
  a silver bullet. It's not some panacea.
• 15:01
  There are a lot of other components to
• 15:03
  the problem space that you need to do
• 15:04
  really well. But in a if you think
• 15:06
  about the problem in a kinda holistic way,
• 15:09
  there actually is a huge amount of opportunity
• 15:11
  to drive really significant significant
• 15:14
  impact in ROI where AI is playing a
• 15:16
  huge part of it. But it needs to
• 15:17
  go hand in hand with other pieces
• 15:19
  of sort of innovation and technology. And so
• 15:22
  if you're kinda stuck on a legacy architecture,
• 15:24
  just thinking about, like, where is your organization
• 15:26
  today? Where do you wanna be three years,
• 15:28
  five years from today?
• 15:30
  I think the thing we see is just
• 15:32
  there's a lot of pressure in The US
• 15:33
  health care system. Like, wages are going up,
• 15:36
  inflation, reimbursement rates are going down. And so,
• 15:39
  you know, I would, again, just say I'm
• 15:40
  incredibly bullish on the role that AI has
• 15:42
  to play.
• 15:44
  I think especially where it's paired with a
• 15:45
  thoughtful, broader kind of technical strategy.
• 15:48
  For sure.
• 15:50
  Some optimistic stuff here for for some challenges
• 15:53
  that have long exist and sort of are
• 15:55
  deeply entrenched, but sort of a path forward
• 15:57
  emerging. Doug, I appreciate you closing us out
• 15:59
  on that note. Thanks for having me.
• 16:02
  Yep. Great to have you on, Doug. We
• 16:03
  also wanna thank our podcast sponsor, Candid Health.
• 16:06
  You can tune to more podcasts from Becker's
• 16:09
  Healthcare by visiting our podcast page at beckershospitalreview.com.