Latest / Becker’s Payer Issues Podcast / Building the Data and AI Foundation for Health Plans

Transcript

Building the Data and AI Foundation for Health Plans · Becker’s Payer Issues Podcast

• 0:00
  - Hello, everyone. This is Erica Spicer Mason with
• 0:02
  - Becker's Healthcare.
• 0:04
  - Thank you so much for tuning in to
• 0:05
  - the Becker's Healthcare podcast series.
• 0:08
  - Today, we're going to talk about how health
• 0:10
  - plans can modernize their data and AI capabilities
• 0:13
  - to better manage
• 0:14
  - rising costs and utilization.
• 0:17
  - And joining me for this conversation is Mike
• 0:19
  - Sankey, vice president of health care and life
• 0:22
  - sciences go to market at Databricks.
• 0:25
  - Mike, it's so great to have you on
• 0:27
  - the podcast today. Thank you so much for
• 0:28
  - joining us. It's great to be here. Thanks
• 0:31
  - for having me.
• 0:32
  - Oh, we're glad to have you. And before
• 0:34
  - we talk a little bit more about health
• 0:36
  - plan strategy here, wanted to know if you
• 0:38
  - could share just a little bit more about
• 0:40
  - your work, your background,
• 0:42
  - whatever you think would be helpful for listeners
• 0:44
  - to know going into this conversation.
• 0:46
  - Yeah. So I guess briefly on Databricks, Databricks
• 0:49
  - is the data and AI platform. We work
• 0:52
  - across most of the Fortune 500.
• 0:55
  - In health care and life sciences, we have
• 0:58
  - more than 1,500
• 0:59
  - customers
• 1:00
  - who are using Databricks on a monthly basis,
• 1:04
  - and we do quite a bit of work
• 1:06
  - with health plans in The US that'll obviously
• 1:09
  - be the focus of
• 1:10
  - today's conversation. I've been at Databricks five and
• 1:13
  - a half years, and my role there
• 1:15
  - is really to take what we built, which
• 1:17
  - is a horizontal technology platform,
• 1:20
  - and help customers adopt it and really extend
• 1:23
  - it into
• 1:24
  - all of these industry specific use cases and
• 1:27
  - applications.
• 1:28
  - And we do that through a combination of
• 1:30
  - solutions that we build. So, for example, interoperability
• 1:33
  - libraries around FHIR, which are
• 1:36
  - very relevant to clinical data feeds for health
• 1:38
  - plans,
• 1:39
  - as well as through the ecosystem
• 1:42
  - that we cultivate. So there are
• 1:44
  - a lot of partners who are building on
• 1:46
  - Databricks. They may have
• 1:48
  - solutions that are specific to
• 1:52
  - industry around prior authorization or around fraud detection
• 1:55
  - that plug in
• 1:57
  - as well as a number of data providers,
• 1:59
  - and we connect to different
• 2:01
  - MCPs like with CMS,
• 2:04
  - that
• 2:05
  - really help our customers
• 2:07
  - use this technology
• 2:09
  - in industry. So that's a bit about me.
• 2:12
  - Oh, great to learn more about you, Mike,
• 2:14
  - and the work that you're doing. I'm sure
• 2:17
  - it's keeping you very busy because as we
• 2:18
  - know, so many organizations are working to either
• 2:21
  - implement
• 2:22
  - AI tools or or expand them and and
• 2:24
  - make sure that their data is ready to
• 2:26
  - do so. So,
• 2:28
  - so I wanna start high level here. We
• 2:30
  - know health plans are heading into 2027
• 2:33
  - facing some very steep cost trends.
• 2:36
  - You know, there's pressure coming from GLP ones.
• 2:38
  - There's rising behavioral health utilization,
• 2:41
  - shifting acuity.
• 2:43
  - I'm curious to know what you're hearing from
• 2:45
  - health plans about where their current tools or
• 2:47
  - capabilities
• 2:48
  - are currently falling short of some of those
• 2:50
  - pressures.
• 2:51
  - Yeah. It's probably worth just spending a couple
• 2:53
  - minutes on some of those macroeconomic
• 2:55
  - pressures that you mentioned because they're very real.
• 2:58
  - There are health plans operating at a financial
• 3:00
  - loss, including a number of health systems that
• 3:03
  - are acting as insurers
• 3:05
  - and trying to manage their own risk pools.
• 3:07
  - And a lot of the large,
• 3:10
  - you know, diversified health services, health plan organizations,
• 3:13
  - CVS, Humana, United,
• 3:15
  - Cigna, Centene,
• 3:17
  - they're all either exiting or downsizing
• 3:20
  - in the Affordable Care Act market or Medicare
• 3:22
  - Advantage.
• 3:24
  - So, you know, many people maybe outside of
• 3:27
  - the industry think that health plans are just
• 3:29
  - minting money year over year,
• 3:30
  - but the reality is relatively minor changes can
• 3:33
  - swing them from a profit to a loss,
• 3:35
  - and that's magnified in these larger plans
• 3:38
  - that have tens of millions of members.
• 3:41
  - So we take a plan that has 10,000,000
• 3:42
  - members, and let's say they make 1,000 in
• 3:44
  - profit from every member over the course of
• 3:46
  - the year.
• 3:47
  - That's 10,000,000,000.
• 3:49
  - But
• 3:50
  - that $1,000,
• 3:51
  - that could be a single month retail cost
• 3:53
  - of a GLP one.
• 3:55
  - It could be just a portion of the
• 3:57
  - facility fee for an ER visit, which wouldn't
• 3:59
  - include the physician's fee and the diagnostic test.
• 4:03
  - So the point being that, again, these utilization
• 4:05
  - changes across a large member population, they can
• 4:08
  - really make an impact.
• 4:09
  - And then we're combining that with the shifting
• 4:11
  - regulatory landscape in areas like Medicaid redetermination
• 4:15
  - and, more recently, the administration
• 4:17
  - ending, a nearly $10,000,000,000
• 4:20
  - subsidy for Medicare Part d.
• 4:22
  - So
• 4:23
  - all this means that there's a lot to
• 4:25
  - react to and a lot of these decisions
• 4:28
  - that health plans need to get right to
• 4:30
  - create benefit designs, to accurately forecast utilization,
• 4:35
  - and to help to manage these dynamic conditions
• 4:37
  - and changing member health over the course of
• 4:39
  - a year, and it's not trivial.
• 4:41
  - And many of these organizations that we work
• 4:43
  - with, I think, self admittedly are data rich
• 4:46
  - but insights poor.
• 4:48
  - In other words, they have massive amounts of
• 4:49
  - claim and member data, but they struggle to
• 4:51
  - process it quickly enough.
• 4:53
  - They struggle to apply additional intelligence and logic.
• 4:57
  - So an example of this could be somewhat
• 4:59
  - basic things like grouping all claims associated with
• 5:02
  - the knee replacement as a single unit.
• 5:04
  - And, fundamentally, they struggle to empower business leaders
• 5:07
  - that need to make these decisions against this
• 5:10
  - landscape that I talked about
• 5:11
  - with the right data and the right time.
• 5:14
  - So all of those, you know, contribute to
• 5:17
  - some of the challenges that they're facing.
• 5:19
  - Yeah.
• 5:20
  - I really appreciate how you broke down some
• 5:23
  - of these costs, Mike, and and what it
• 5:24
  - means on a larger scale. And you mentioned
• 5:28
  - some of these challenges around
• 5:30
  - being data rich and info poor. So I
• 5:32
  - wanna take that a little bit further.
• 5:35
  - Most plans already have claims data, pharmacy data,
• 5:38
  - authorization data. There's no shortage of it. But
• 5:41
  - I'm curious what you see
• 5:43
  - break
• 5:44
  - between having that data
• 5:46
  - and then catching a utilization shift within the
• 5:49
  - month that it happens
• 5:51
  - early enough for the health plan to act.
• 5:53
  - What is that looking like?
• 5:55
  - Yeah. I think a a lot of organizations
• 5:58
  - are stuck in this back and forth between
• 6:01
  - the technical teams that are building the pipelines
• 6:03
  - or building the data products or building the
• 6:05
  - dashboards and the business leaders that are actually
• 6:08
  - making decisions.
• 6:09
  - And people largely are trying to self serve
• 6:12
  - out of dashboards,
• 6:14
  - but they're not really built for deep analytics
• 6:16
  - and
• 6:18
  - reasoning and trying to
• 6:20
  - understand some of the root cause and dynamics
• 6:22
  - and actions that that an organization can take.
• 6:25
  - And
• 6:26
  - oftentimes, you'll see these leaders. They'll get frustrated.
• 6:29
  - They'll be clicking around in the dashboard. They
• 6:30
  - still don't get the answer they want without
• 6:32
  - needing a team to go and a technical
• 6:34
  - team that requires
• 6:36
  - rebuilding a
• 6:37
  - pipeline or data table or a dashboard.
• 6:41
  - And then what's happened is many of these
• 6:43
  - large organizations,
• 6:44
  - they've
• 6:45
  - grown to have shadow IT organizations, and that
• 6:49
  - exacerbates some of these challenges in identifying the
• 6:51
  - source of truth.
• 6:53
  - So I think a large part of it
• 6:54
  - is in the fundamental challenges these organizations have
• 6:58
  - had in actually democratizing
• 7:01
  - the the use of those data.
• 7:04
  - Yeah. That concept of democratizing the use of
• 7:06
  - data, Mike, I think is a really important
• 7:08
  - one.
• 7:09
  - And that leads me to ask, you know,
• 7:11
  - if nontechnical
• 7:12
  - teams are are accessing
• 7:15
  - important data,
• 7:16
  - how can they safely query live claims data
• 7:19
  - without,
• 7:20
  - you know, five people getting five different answers
• 7:23
  - at any given moment?
• 7:25
  - Yeah. That... That's one of the biggest areas
• 7:28
  - we've been spending time as an organization. It's
• 7:31
  - building the infrastructure
• 7:33
  - around that context layer that both humans and
• 7:36
  - AI agents need to succeed.
• 7:38
  - And if you think historically
• 7:40
  - of
• 7:41
  - where that context or that business logic has
• 7:43
  - lived, it's been hard coded in data pipelines
• 7:47
  - or living in disparate business intelligence dashboards or
• 7:50
  - it's in the head of Sally or Joe
• 7:51
  - or someone who's worked in the claims department
• 7:54
  - for twenty years.
• 7:56
  - And AI agents, in some ways, are even
• 7:58
  - worse than humans in that they're likely to
• 8:00
  - just make something up if they don't have
• 8:02
  - the appropriate context.
• 8:04
  - So what we've tried to do and help,
• 8:06
  - our health plan customers with is build the
• 8:09
  - capabilities to store that context at the data
• 8:11
  - layer
• 8:12
  - so that anyone or anything interacting with the
• 8:14
  - data, whether it's a human SQL programmer or
• 8:17
  - business intelligence dashboard
• 8:19
  - or an AI agent, they get the same
• 8:21
  - information
• 8:22
  - with the same shared definitions.
• 8:25
  - And this really requires both modeled semantics as
• 8:28
  - well as learned knowledge.
• 8:30
  - So let's say that someone wants to understand
• 8:33
  - the PMPM for commercial members who've had a
• 8:36
  - knee replacement.
• 8:37
  - We need a business definition of how to
• 8:39
  - calculate
• 8:40
  - per member per month.
• 8:42
  - We may need to store that acronym as
• 8:45
  - logic.
• 8:46
  - We also need to define the time period
• 8:48
  - of how we calculate that in a consistent
• 8:50
  - way, whether that's the most recent fiscal year
• 8:53
  - to date or a rolling twelve months.
• 8:55
  - We also would need a definition of commercial
• 8:57
  - line of business. That's probably a grouping of
• 8:59
  - many different segments
• 9:01
  - across a health plan, in terms of their
• 9:04
  - employer business and administrative services.
• 9:07
  - And all these definitions need to be stored,
• 9:09
  - accessible, and governed.
• 9:11
  - And we have features called business semantics and
• 9:13
  - metric views that support these.
• 9:16
  - The other thing that we need to get
• 9:18
  - right is we need to know the correct
• 9:19
  - tables to pull from because you can rest
• 9:22
  - assured that a health plan has dozens of
• 9:24
  - fields across many tables that are just labeled
• 9:26
  - claim,
• 9:27
  - And we've created capabilities to understand
• 9:30
  - the most relevant data sources and rank them
• 9:32
  - and help teams certify them.
• 9:34
  - So relevance, for example, could be based on
• 9:36
  - how frequently a given table is used to
• 9:38
  - populate reports for the CFO.
• 9:41
  - And all these things together, they create that
• 9:43
  - governed data intelligence substrate to answer the questions
• 9:46
  - more consistently and correctly.
• 9:48
  - So in your question of if five people
• 9:50
  - ask the same question,
• 9:52
  - if they're able to refer to the same
• 9:55
  - business definitions,
• 9:56
  - logic, and underlying
• 9:58
  - data, then they should get the same answer
• 10:00
  - time and time again.
• 10:03
  - It makes a lot of sense, Mike, and
• 10:05
  - I really appreciate you walking our listeners through
• 10:07
  - that.
• 10:08
  - And I wanna back up a little bit
• 10:10
  - too and talk a bit more about adopting
• 10:12
  - AI in the first place.
• 10:14
  - Adoption of AI tools is essentially universal. I
• 10:18
  - think it's safe to say at this point,
• 10:19
  - especially in the health plan space,
• 10:21
  - but readiness
• 10:23
  - for
• 10:24
  - various AI tools is another story. So where
• 10:26
  - are you seeing health plans get stuck when
• 10:29
  - it comes to adoption?
• 10:31
  - I think that context layer that I described
• 10:33
  - is probably the single biggest rate limiting factor
• 10:36
  - right now.
• 10:37
  - And we believe organizations need four things to
• 10:40
  - succeed with AI.
• 10:42
  - We refer to them as the four c's
• 10:44
  - if you need a memory device.
• 10:47
  - So context is one. We've talked quite a
• 10:49
  - bit about that.
• 10:50
  - And the other three are control, cost, and
• 10:53
  - choice.
• 10:54
  - So let's spend a bit of time on
• 10:56
  - the governance challenges facing health plans and health
• 10:58
  - care in general because those really get to
• 11:00
  - the concepts around
• 11:02
  - control, costs, and choice.
• 11:04
  - There's not a single unifying standard for AI
• 11:07
  - in health care, and states continue to introduce
• 11:09
  - their own legislation, which really complicates things for
• 11:12
  - plans
• 11:13
  - that operate across multiple states.
• 11:15
  - So organizations
• 11:17
  - need a robust framework for evaluations, including around
• 11:20
  - bias
• 11:21
  - and observability of these models.
• 11:24
  - And just like with their human resources, they
• 11:26
  - need to control what agents access and what
• 11:28
  - they do with the data.
• 11:30
  - Add all of this together, and many health
• 11:32
  - plans, they can't keep pace with the changes
• 11:33
  - among the Frontier AI labs.
• 11:36
  - So they really need to focus on building
• 11:39
  - an infrastructure that preserves model choice and gives
• 11:42
  - them the ability to work across models, which
• 11:44
  - also helps them manage cost because it enables
• 11:47
  - them to send trivial tasks to lower cost
• 11:49
  - models.
• 11:50
  - So beyond getting the context and this data
• 11:53
  - substrate right, I would say most organizations are
• 11:55
  - struggling with governance at scale,
• 11:58
  - as it pertains to controlling,
• 12:01
  - what agents are accessing, how they're accessing it,
• 12:04
  - the cost around it, and then model choice
• 12:07
  - and choice and cost are
• 12:09
  - very closely intertwined at this point.
• 12:11
  - Mhmm.
• 12:13
  - Yeah. And to your point too, Mike, about
• 12:15
  - there not being one universal standard,
• 12:17
  - how so many of these regulations
• 12:19
  - around AI and governance can vary state to
• 12:21
  - state
• 12:21
  - makes it very complicated and complex for plans.
• 12:25
  - I'm curious to know a little bit more
• 12:26
  - about the plans that you've seen get the
• 12:29
  - governance piece right.
• 12:31
  - What did they have to put in place
• 12:32
  - before people could trust the data they were
• 12:35
  - working from? What does that look like?
• 12:39
  - So,
• 12:40
  - for this, maybe I can just share a
• 12:42
  - couple of customer examples.
• 12:44
  - We hosted our big data and AI summit
• 12:46
  - earlier this summer. We had so many health
• 12:48
  - plan customers attend and share what they're doing.
• 12:51
  - So one of those was Humana, and they've
• 12:53
  - been a Databricks customer for a while, including
• 12:55
  - in support of their value based care initiatives.
• 12:58
  - And they recently expanded their use of Databricks
• 13:01
  - to the office of the CFO.
• 13:03
  - And the entire focus of their presentation was
• 13:05
  - about rebuilding finance on a trusted governed data
• 13:08
  - foundation
• 13:09
  - and operationalizing
• 13:10
  - AI across these different functions.
• 13:12
  - And for them, that meant, including domain product
• 13:15
  - owners from underwriting, from pricing, from risk assessment,
• 13:18
  - from revenue management,
• 13:20
  - from areas like cost and expense management,
• 13:24
  - and having those domain owners define the business
• 13:26
  - semantics and the metrics to feed a unified
• 13:29
  - semantic model
• 13:31
  - that could be used across,
• 13:33
  - the office of the CFO. So it really
• 13:35
  - was this combination of
• 13:36
  - the technical foundation, but also the intelligence
• 13:41
  - from their team and from these different functions
• 13:43
  - to define
• 13:44
  - that context layer.
• 13:46
  - The second example I'll share is from Centene,
• 13:49
  - and they presented on
• 13:51
  - the gifts of machine agentic governance intelligence, which
• 13:54
  - they shorthand refer to as the gifts of
• 13:56
  - the Magi.
• 13:58
  - And it was all about delivering secure, repeatable,
• 14:01
  - and explainable AI.
• 14:03
  - And this included a lot of different components
• 14:05
  - for security. It included things like usage tracking
• 14:08
  - and monitoring,
• 14:10
  - a single interface to all external models so
• 14:12
  - they could track
• 14:13
  - how the models were hitting their data effectively,
• 14:18
  - as well as comprehensive guardrails.
• 14:20
  - For explainability, and this gets into some of
• 14:22
  - the governance and the regulations, they create an
• 14:25
  - immutable record behind every AI decision.
• 14:27
  - So they capture and lock everything, including the
• 14:30
  - prompts, the responses, the tool calls in these
• 14:32
  - multi agent systems,
• 14:34
  - and they have an evaluation process that's actually
• 14:36
  - a combination of automated scores
• 14:39
  - and human review.
• 14:40
  - And they're able to continuously monitor this life
• 14:42
  - cycle for things like model drift and policy
• 14:45
  - infractions.
• 14:47
  - And finally, in terms of repeatability, they've deployed
• 14:50
  - the scaffolding and created a library of reusable
• 14:52
  - templates
• 14:54
  - to ensure consistency and correctness.
• 14:56
  - So this includes the context layer, which we've
• 14:59
  - talked about quite a bit
• 15:00
  - to support how agents retrieve information.
• 15:03
  - And what this has done is it's created
• 15:05
  - an operating velocity that enables them to rapidly
• 15:08
  - deploy across different areas of their business.
• 15:11
  - A recent example of this was a new
• 15:13
  - solution for care managers to support interactions with
• 15:15
  - members
• 15:16
  - and help to close gaps in care.
• 15:19
  - And, Mike, I I really appreciate the detailed
• 15:22
  - examples from both of those health plans. I
• 15:24
  - think that's going to be really helpful,
• 15:26
  - for folks listening in who are, you know,
• 15:28
  - just in the early stages of this journey
• 15:29
  - and to have that concrete view of what
• 15:31
  - this looks like,
• 15:33
  - and also appreciate the details you shared around
• 15:35
  - the four c's as well. And, again, for
• 15:37
  - listeners, when we talk about AI adoption, readiness
• 15:39
  - looks like those four c's, context, control, cost,
• 15:42
  - choice.
• 15:44
  - Mike, a lot of valuable insight you've given
• 15:46
  - us today. So before I let you go,
• 15:48
  - I wanted to kind of get your look
• 15:49
  - ahead to 2030.
• 15:52
  - What capabilities around data, AI, and cost visibility
• 15:55
  - will become table stakes for health plans, and
• 15:58
  - what should leaders be building toward right now?
• 16:01
  - Yeah. So I would definitely encourage leaders to
• 16:03
  - think about building that proper foundation so they
• 16:06
  - can achieve that scale velocity.
• 16:09
  - And if we think about 2030, we're not
• 16:11
  - that far away. We're looking a few years
• 16:13
  - out. What will be table stakes for health
• 16:15
  - plans?
• 16:16
  - And I'll throw out a few ideas here
• 16:19
  - that I don't think are that far fetched.
• 16:20
  - I think many organizations are already on this
• 16:23
  - journey.
• 16:24
  - The first is around fraud detection, and I
• 16:25
  - think we're gonna move to a world where
• 16:27
  - fraud is detected before the claim is paid
• 16:30
  - rather than the catch and chase model that,
• 16:33
  - you know, many organizations operate in today.
• 16:36
  - And that would mean models are scoring every
• 16:38
  - claim prepayment. They're stopping improper payments.
• 16:42
  - They're embedding AI to do things like automatically
• 16:44
  - generate letters if they have to request more
• 16:46
  - information.
• 16:47
  - And, ultimately, the goal is to reduce fraud,
• 16:49
  - waste, and abuse to low single digits. Right
• 16:52
  - now, for many organizations, it's actually in the
• 16:54
  - high single digits, and that has a major
• 16:57
  - impact on their operating margins.
• 16:59
  - The second is around prior authorization. Obviously, has
• 17:02
  - has,
• 17:03
  - gotten a lot of attention.
• 17:06
  - Recently, there are a lot of regulations
• 17:08
  - around standards coming out from CMS,
• 17:12
  - and I'd like to see a world where
• 17:14
  - prior authorization is 80% plus automated.
• 17:17
  - Today, it's probably about half of that. It
• 17:19
  - depends, of course, on the organization and the
• 17:21
  - book of business.
• 17:23
  - And this would mean gold carding,
• 17:25
  - more top quality providers,
• 17:27
  - having agents auto adjudicate most requests,
• 17:29
  - and shortening that median turnaround time to under
• 17:32
  - twenty four hours, even under twelve hours in
• 17:34
  - instead of days.
• 17:37
  - The third area is around the payment models,
• 17:39
  - and
• 17:40
  - I'd like to see a world where all
• 17:42
  - payments or nearly all payments are value based
• 17:44
  - and every dollar is tied to outcomes with
• 17:46
  - those quality metrics aligned
• 17:49
  - across plans and providers.
• 17:51
  - And the fourth is around using AI and
• 17:54
  - supporting a virtual concierge for every member.
• 17:57
  - A personalized assistant that stays with the member,
• 18:00
  - understands their history,
• 18:01
  - and can help resolve payment, benefit, and care
• 18:04
  - questions end to end across every channel. And,
• 18:06
  - again, I don't think that's far fetched. We're
• 18:08
  - seeing this embedded
• 18:10
  - in many different functions already,
• 18:12
  - within health plans, and I think having that
• 18:15
  - personalized concierge
• 18:17
  - will really improve the member experience.
• 18:20
  - Mike, I wanna thank you again for all
• 18:22
  - of the insight and the practical guidance you've
• 18:25
  - you've shared with our audience today and and
• 18:27
  - for the time that you spent with Becker's.
• 18:29
  - I've learned a lot from our conversation, and
• 18:31
  - I'm sure our listeners from health plans will
• 18:33
  - feel the same. So thank you again for
• 18:35
  - your time today.
• 18:36
  - Yeah. Appreciate it. Take care. And we'd also
• 18:39
  - like to thank our podcast sponsor for today,
• 18:42
  - Databricks.
• 18:43
  - Listeners, be sure to tune in to more
• 18:45
  - podcasts from Becker's by visiting our podcast page
• 18:48
  - at beckershospitalreview.com.