Latest / Becker’s Payer Issues Podcast / Balancing Innovation and Ethics: The Role of Agentic AI in the Payer Landscape

Transcript

Balancing Innovation and Ethics: The Role of Agentic AI in the Payer Landscape · Becker’s Payer Issues Podcast

0:00
Hello, everyone. This is Erica Spicer Mason with
0:02
Becker's Healthcare. Thank you so much for tuning
0:04
in to the Becker's Healthcare podcast series.
0:07
So today, we're going to discuss AI and
0:09
the payer landscape and how to balance innovation,
0:12
ethics, and regulatory
0:14
change.
0:14
And joining me for this discussion is Nitu
0:17
Rajpal,
0:18
the CEO of Lilac Software.
0:20
Nitu, welcome to the podcast. Thank you so
0:23
much for making time for us today. Thank
0:25
you, Erica. It's really great to be here.
0:27
Oh, we're so thrilled to have you. And
0:29
before we get into our discussion, wanted to
0:32
see if you'd like to share a little
0:33
bit more about yourself, your work in health
0:35
care, your organization,
0:37
whatever feels pertinent to listeners.
0:39
Oh, thank you. So I'm Neetu Rajpal, and
0:43
I am the
0:44
CEO and cofounder of Lilac Software.
0:47
With Lilac Software, we're very focused on making
0:51
sure good data gets to all the practitioners
0:54
that are currently in health care. We're building
0:57
a fully managed
0:59
solution for data analytics,
1:01
insights, AI
1:03
for payers and providers.
1:05
Before my work at Lilac,
1:08
I was the CTO of Oscar Health. I
1:11
appreciated and loved being at Oscar Health.
1:15
I oversaw work for all of the tech
1:17
stack that's used to run the insurance company,
1:20
everything from
1:21
from claims processing to enrollment, billing, brokers, member
1:25
engagement. And going through that journey really, like,
1:29
inspired what I am doing at EyeLac now.
1:32
And and overall, my career has all been
1:34
building software. So I've been building software for
1:36
multiple decades,
1:37
and enterprise data analytics is is my passion
1:41
to get make sure the right data gets
1:42
to the right people to make good decisions
1:45
and to help them with, with agents.
1:49
Yeah. Me too. I really appreciate the background
1:52
information. And it it sounds like you have
1:54
the perfect blend of experience, you know, coming
1:56
from a a major payer organization and all
1:58
of your decades of experience in software development
2:00
and building.
2:01
Sounds like you are the perfect person to
2:03
talk to us about
2:04
AI and the payer landscape
2:06
today. So
2:08
I wanted to start us off by
2:10
just kind of acknowledging a trend that I
2:12
know I've seen and our audience has seen
2:14
in health care coverage lately, and that's the
2:17
fact that Agentic AI is really gaining momentum.
2:20
So can you provide a brief overview of
2:23
what this technology is and maybe even share
2:25
a practical example of how it's being used
2:27
right now?
2:28
Yeah. I think Adjentic AI is one of
2:31
the most exciting advancements in in in AI.
2:35
I
2:35
I know I say that in addition to
2:38
the fact that there was already chat TPT
2:40
and there's machine learning.
2:42
But just to give a bit of a
2:43
primer, machine learning in is
2:45
is colloquially
2:46
more about, like, finding the right set of
2:48
patterns and data and understanding data.
2:51
Large language models use that information
2:54
to then understand and get context and and
2:57
to forecast
2:59
build new documents or build new conversations or
3:01
build new things.
3:03
Agentic
3:05
AI is this exciting
3:07
advancement
3:08
where
3:08
Agentic AI systems can, like, really be in
3:12
the world and understand it and observe it
3:14
and learn from it,
3:16
and then reason about it and build a
3:18
plan to take action, and then go ahead
3:20
and take an action. So that all sounds
3:23
very much like like what a human AI
3:26
assistant would do, and and I think that's
3:29
a pretty good way of thinking about agentic
3:31
systems and agentic AI,
3:34
assistants
3:35
or independently
3:36
operating folks who can start doing more and
3:38
more things.
3:39
And as the assistants get more familiar with
3:43
the work that they're doing, they get more
3:45
and more independent, and then I think that's
3:46
also a good corollary.
3:48
In in terms of, like, health care world,
3:52
a a simple simple example that I can
3:54
come up with is something like, think about,
3:56
like, appeals and grievances or grievances against claims.
4:00
And it's somebody files a grievance about a
4:03
denied claim.
4:05
AI, in this case, can step in and
4:06
streamline the process of, like, working through that
4:10
grievance as in they can cross reference all
4:12
of the information that's available about the member,
4:15
about the policy and the plan, about identifying
4:18
what the discrepancy is, if there's any missing
4:20
information.
4:21
AI can gather all of this information behind
4:23
the scenes without a person being involved, And
4:26
then they can, like, understand the context and
4:28
then
4:29
generate
4:30
a personalized response for this particular
4:33
claim grievance.
4:34
And that, I think, is the most exciting
4:36
part. Then they bring a plan and an
4:38
execution plan perhaps
4:40
to to a human who can approve it,
4:43
but with all of the supporting information included,
4:46
and then go execute on that plan. And
4:49
this this, like, super dedicated problem solver
4:52
doesn't actually have a scale requirement.
4:54
A person can totally do it, and a
4:57
person will likely do it in most complicated
4:59
scenarios for a long time. But the simpler
5:03
scenarios can actually be handled by an agentic
5:06
system in this case. So that I think
5:08
is a pretty practical example, and I know
5:10
of places that are already starting to implement
5:12
this.
5:14
Fascinating. And I think you hit on an
5:16
area that certainly is kind of the center
5:18
of a lot of conversations when it comes
5:20
to efficiency in health care organizations,
5:22
whether it's on the provider or payer side.
5:25
You know, claims is is such an area
5:27
where folks are really trying to identify those
5:29
opportunities to streamline things. So
5:31
really appreciate that example. And I was actually
5:34
planning to ask you also how how payers
5:36
can leverage Genentech AI maybe to even help
5:39
address quality or care delivery challenges. Could you
5:42
shed some light on that and maybe elaborate
5:44
on any improvements you've seen in those use
5:46
cases?
5:47
Absolutely.
5:49
As of right now, like, the the STARS
5:51
program and the and the HEDIS gap closure
5:54
and just generally performing well on STARS is
5:57
very near and dear to my heart.
6:00
So I'll speak a pick an example from
6:02
there. So
6:03
most
6:04
folks in in pair organizations that I've talked
6:06
to try very, very hard to, quote, unquote,
6:10
close every gap as in identifying all of
6:12
the members that need a particular service
6:16
and help them get that service. It could
6:19
be it could be as simple as ensuring
6:21
that all the right members who should be
6:23
getting a breast cancer screening are, in fact,
6:25
getting a breast cancer screening. The way this
6:27
happens today
6:29
is is
6:30
some person generates some list, and then some
6:33
other person takes that list, and then some
6:36
other person or the same person makes a
6:38
phone call and and and then tries to
6:40
make another phone call to encourage the member
6:43
to to go schedule an appointment with a
6:45
provider
6:46
and then go to the appointment and and
6:49
and, like, have have a mammogram.
6:52
All of this is ripe for an agentic
6:55
system to do everything behind the scenes as
6:57
in identify the members that need a gap
6:59
closed,
7:01
figure out the right time and way that
7:03
that gap needs to be closed,
7:05
making a phone call to the member,
7:08
identifying that the barrier is actually lack of
7:10
scheduling, then maybe just, like, going ahead and
7:13
scheduling an appointment.
7:14
And then also maybe recognizing there are some,
7:17
like, nonmedical
7:18
determinants of,
7:20
detractors of health here. So maybe the member
7:22
can't get to the to their appointment on
7:25
time. So, like, booking an Uber as well.
7:27
So everything to do from
7:29
knowing who needs help,
7:32
identifying
7:33
how they need help, identifying the the blockers
7:36
for them to close that gap, and then
7:39
executing on all of those pieces
7:41
behind the scenes to just make it easy
7:44
for the member to actually go get their
7:45
mammogram.
7:47
What I think is, like, super exciting and
7:49
where it, like, really transforms things
7:52
is it is not necessarily that you can
7:54
do this for a group of members or
7:56
a subset of members. You can, in fact,
7:58
do this for every member because you're no
8:00
longer limited. Your scale is no longer limited
8:03
by the number of people who can make
8:05
phone calls or the amount of time a
8:07
person
8:08
you can scale horizontally
8:10
to compute.
8:11
You can have lots of agents doing lots
8:13
of things all at the same time, and
8:16
you can personalize care. You can personalize outreach.
8:19
You can personalize
8:20
the the the plan the member needs to
8:23
go close the care gap. So I'm, like,
8:25
super excited about super excited about the possibilities
8:29
here where
8:30
personalization
8:31
of health care is is that is is
8:34
available now.
8:37
Yeah. Me too. That that example
8:39
of the care gap list really hits home
8:41
for me personally. I had a past life
8:43
working as a health educator at a health
8:45
plan, and I remember manually working off of
8:47
those lists and doing the phone call outreach
8:50
and
8:51
talk about time consuming. So to your point
8:52
about being able to scale that, not just
8:54
to a small subset of members, but potentially
8:57
all members at a health plan, it's it's
8:59
a really exciting idea.
9:01
Yeah. I'm very excited for our future here.
9:04
Yeah.
9:05
Yeah. And
9:06
so a follow-up question for you too. So
9:09
it sounds like how health plans are are
9:11
certainly
9:12
experimenting with AI, piloting it, and even using
9:14
it in these ways as you've mentioned.
9:16
And as they're as they're innovating here, how
9:19
can they also ensure that these AI tools
9:21
are ethical and unbiased and also patient focused,
9:25
you know, especially when we're talking about Medicare
9:27
or Medicare Advantage population?
9:30
Yeah. Really, really excellent question. We we all
9:33
know that your AI models are only going
9:35
to be as good as the data that
9:37
they've been used to train get trained on,
9:40
and any bias in that data will come
9:42
through in the models themselves. So I think
9:46
the very least, it it needs to it
9:48
needs to be monitored
9:50
as to which model you're using to implement
9:53
your solution and maybe trying a few different
9:55
ones and and having lots of spot checks
9:58
in the middle.
9:59
At least for a while until these systems,
10:02
like, you we can we can eliminate,
10:06
known biases from these models. Or
10:09
so I think the the mitigating bias
10:12
based on the data that the models have
10:14
been trained on and mitigating
10:16
that risk by using multiple models is is
10:20
is the first one.
10:22
Secondly, I I would encourage that, like, when
10:24
we try things, we actually keep the member
10:27
at the center of all of it and,
10:29
ensure that, like, the real world needs are
10:32
met.
10:33
So we're not we're not building tech ahead
10:36
of its ability to be used in productive
10:38
ways.
10:39
And that may not always mean just everything
10:43
all tech. It may mean a blended version
10:45
of
10:46
tech and process. And and one of the
10:49
last things I'd say here is
10:52
picking
10:53
the amount of automation you want here.
10:56
I think there
10:58
tends to be a lot
11:00
of intention to
11:02
automate everything.
11:04
I have been doing tech for a very,
11:06
very long time, and I think
11:08
the first
11:10
seventy percent is usually the most exciting and
11:12
the fastest piece to get to. The next
11:14
10 is a little bit harder, and then
11:16
the last 20 is really, really hard.
11:19
And preciseness
11:20
and and accuracy
11:22
in health care specifically matters a lot. So
11:25
I actually think it's perfectly fine to go
11:28
with, like, aim for 80% as long as
11:32
the accuracy and and completeness are really high
11:35
on the required list of things.
11:38
And given how, like, stretched thin almost everybody
11:42
in a health plan is, I think 80%
11:45
is pretty awesome.
11:47
So mitigating bias in the data that's used
11:50
by using multiple models,
11:52
embedding
11:53
patient
11:55
or member at the center of whatever value
11:57
that we're bringing. And finally, like,
12:00
I think it's it's worth worth the question
12:02
to ask if %
12:04
is really should be the goal. Maybe it
12:06
should just be 80% for for a while
12:08
until we're sure.
12:10
I really appreciate how practical those tips are.
12:13
Me too. So thank you for that. And
12:15
I think that 80%
12:16
metric that you mentioned, it's definitely something interesting
12:19
for leaders to keep in mind as they
12:21
move forward with innovation.
12:23
And similarly,
12:24
you know, we're talking about ethics here with
12:27
AI and technology implementation.
12:29
I I know that we're also seeing in
12:31
the news some headlines about
12:33
AI regulations
12:35
kind of going away or at least being
12:37
much less restrictive,
12:38
and that's, you know, rapidly changing and evolving
12:41
in 2025.
12:42
So how do you foresee the role of
12:45
AI in health care changing, especially shaping the
12:47
future of value based care and Medicare programs?
12:51
Yeah. I hope
12:53
that we will build systems that are built
12:56
on top of systems that are currently subject
12:59
to HIPAA and
13:01
security and compliance and privacy
13:03
and
13:04
security
13:05
rules around there. And I think a really
13:08
good way would be to assume your own
13:11
AI system is also subject to the same
13:13
rules. So data privacy, especially when it comes
13:17
to health care, seems really, really valuable data
13:20
to protect
13:21
even if the regulations
13:23
lag in time. Because I think where there
13:25
is a there is a
13:27
really
13:28
different
13:31
regulations and the advancements in AI are moving
13:34
at. And and I think
13:36
the people's privacy and people's privacy on their
13:39
health care data is is paramount, and I
13:41
and I really encourage the
13:44
plans using AI to impose the same requirement
13:47
on themselves and while the regulation is still
13:50
being worked out.
13:52
Yeah.
13:53
I I really appreciate how you put that
13:55
need to, you know, how there's kind of
13:56
this discrepancy in the velocity of the speed
13:59
of innovation and and the regulations
14:01
behind it. So
14:03
really appreciate all of these insights that you've
14:05
shared with me and our listeners today. Is
14:08
there anything else that you wanted to highlight
14:10
or maybe anything we didn't touch on but
14:11
you wanna mention?
14:14
Yeah. Thank you for the opportunity. I,
14:17
actually wanna just talk. Like, I've I am
14:19
a technologist, and I've been a technologist pretty
14:21
much my whole career.
14:23
One of the
14:25
things that I find the most fascinating about
14:28
doing things in health care is it is
14:30
a very integrated environment,
14:32
I think.
14:33
So
14:35
being able
14:36
to bring technology
14:38
to a plan, but fully recognizing
14:41
the current state of the world is going
14:43
to be really helpful. It's a it's a
14:45
bit of a a weird way to say
14:47
it, but I think bringing everybody along is
14:49
actually going to be the fastest way to
14:52
get this technology adopted. It's like you how
14:54
you need trust,
14:55
you need adoption,
14:57
and you need tech that works. And you
14:59
only know if your agentic system works and
15:02
is actually solving the problem of the workflow
15:04
and the practitioners
15:06
if those practitioners are using it and are
15:08
telling you what is working and what is
15:10
not working, which requires a lot of trust
15:12
and it requires a lot of adoption. So
15:15
just as much investment in tech, I think,
15:17
also needs investment in just change management with
15:21
the organization.
15:22
I'm very excited about our future here. I
15:25
I think there is real opportunity here to
15:27
personalize care and outreach and and help lots
15:31
of people in the ways they like to
15:33
be helped.
15:35
Yeah. No.
15:36
I can hear the excitement in your voice,
15:38
Me too, and it's always it's a pleasure
15:40
to talk to people who are really passionate
15:42
about the work that they're doing. And and,
15:44
I really appreciate also those last three pillars
15:47
you shared with our listeners. You know, what
15:48
will it take looking ahead? It'll take trust,
15:50
adoption, and tech that works.
15:52
Such a fantastic discussion. Thank you so much
15:54
again, Mitu, for making the time for Becker's
15:56
today.
15:57
Thank you very much, and thank you very
15:59
much for the opportunity. Really excited to be
16:01
here.
16:02
And we also wanna thank our podcast sponsor
16:04
for today, Lilac Software.
16:06
You can tune into more podcasts from Becker's
16:08
Healthcare by visiting our podcast page at beckershospitalreview.com.