Latest / Becker’s Payer Issues Podcast / Speed vs. Judgment: Rethinking AI’s Role in Underwriting

Transcript

Speed vs. Judgment: Rethinking AI’s Role in Underwriting · Becker’s Payer Issues Podcast

• 0:00
  - Hi, everyone. This is Lucas Voss with Becker's
• 0:02
  - Healthcare. Thanks so much for tuning in to
• 0:04
  - the Becker's Healthcare podcast series. It's fantastic
• 0:07
  - to have you. Today, we're talking about
• 0:10
  - speed
• 0:11
  - versus
• 0:12
  - judgment,
• 0:13
  - rethinking AI's role
• 0:15
  - in
• 0:16
  - underwriting. And joining me to discuss just that.
• 0:19
  - It's so great to have you is Mark
• 0:21
  - Jeffries. He's the general manager health at Gradient
• 0:23
  - AI. Mark, it's so great to have you.
• 0:25
  - Thanks for being here. Yeah. Appreciate it, Lucas.
• 0:28
  - Appreciate the opportunity.
• 0:30
  - Before we head into our conversation here, I
• 0:32
  - just wanna give, our listeners a sense of
• 0:34
  - what you do for work. If you could
• 0:36
  - just introduce yourself a little bit, talk a
• 0:38
  - little bit about your work in health care.
• 0:40
  - Yeah. Yeah. Happy to. So again, Mark Jeffries.
• 0:43
  - I'm the general manager for the health practice
• 0:45
  - here at Gradient AI. I've been with Gradient,
• 0:48
  - almost two years now. I'm accountable for strategy,
• 0:51
  - execution, and growth of the health business,
• 0:54
  - and I have a keen focus on developing
• 0:56
  - our portfolio of solutions to meet customer needs.
• 0:59
  - Prior to joining Gradient, I was at Cigna
• 1:02
  - for the last ten years where I helped
• 1:03
  - build and then ultimately led their supplemental health
• 1:05
  - business, which was part of the US health
• 1:08
  - care organization.
• 1:09
  - I've also, in my career, spent time at
• 1:11
  - Aetna and the Medicare Advantage business, MetLife, and
• 1:13
  - the Hartford in group benefits and financial services.
• 1:15
  - So I have a
• 1:17
  - a broad and varied background.
• 1:19
  - Again, it's so great to have you. I
• 1:21
  - wanna talk about the title of our episode
• 1:24
  - today or part of it, speed versus judgment,
• 1:26
  - because this is something that's extremely prevalent in
• 1:29
  - health care right now, especially when it comes
• 1:31
  - to
• 1:32
  - AI and where the humans are involved still
• 1:35
  - and the robots take over. Again, especially in
• 1:38
  - underwriting, as AI takes more of that group
• 1:42
  - health underwriting process over, which is certainly this
• 1:44
  - case,
• 1:46
  - from your perspective, which decisions
• 1:49
  - are still
• 1:50
  - important where you need an underwriter to have
• 1:53
  - that judgment in?
• 1:55
  - Yeah. That's... Frankly, it's an interesting question. Right?
• 1:58
  - Personally, I'd argue that all decisions still call
• 2:01
  - for an underwriter's judgment. Yeah. That probably sounds
• 2:04
  - kind of odd coming from a guy whose
• 2:05
  - company develops AI powered tools for underwriting,
• 2:08
  - but the underwriter still plays a fundamental role
• 2:11
  - in determining the context
• 2:13
  - in which an AI generated result or output
• 2:15
  - should be relied on for a decision.
• 2:18
  - You know, the underwriter uses their judgment to
• 2:20
  - decide about the the level of risk they're
• 2:22
  - willing to assume when using that AI solution
• 2:25
  - to help manage their overall book of business.
• 2:27
  - You know, it's also important to recognize that
• 2:30
  - the use of AI tools
• 2:31
  - isn't an all or nothing proposition.
• 2:34
  - Right? And I think a lot of the
• 2:35
  - conversation that's happening in the industry today is
• 2:38
  - kind of looking at it through that binary
• 2:40
  - lens. You either use it or you don't.
• 2:43
  - I don't think that's really the reality or
• 2:45
  - where the opportunity lies.
• 2:47
  - Using various features in our tools, underwriters can
• 2:50
  - enable multiple tiers of assessment and adjudication
• 2:53
  - that... To drive efficiencies in their business processes.
• 2:56
  - You know, for example, using an integrated rules
• 2:58
  - engine, an underwriter can enable
• 3:00
  - sophisticated workflow processes based on the presence of
• 3:03
  - certain conditions or drugs. If the conditions or
• 3:06
  - drugs are present, the quote can be routed
• 3:08
  - to different levels of review within the process.
• 3:11
  - Right? So that type of approach helps the
• 3:13
  - underwriter allocate their valuable time and expertise to
• 3:16
  - quotes that need a deeper level of review.
• 3:19
  - I wanna tap just on your experience a
• 3:21
  - little bit here specifically. Again, you you spent
• 3:23
  - so much time in this space too.
• 3:25
  - Is there an example that you can think
• 3:26
  - of where this was... The process that you've
• 3:29
  - just described
• 3:30
  - happened where an AI generated assessment looked really
• 3:33
  - straightforward, right, but an underwriter took... Then took
• 3:36
  - a closer look and said, okay. Hey. I
• 3:38
  - I I need to take a look here.
• 3:39
  - We might need a review. And and then
• 3:41
  - from your perspective again, what prompts those reviews?
• 3:43
  - What prompts the underwriter
• 3:45
  - to say, hey, I I need to take
• 3:46
  - another look?
• 3:47
  - Yeah. Yeah. That's a that's a great question
• 3:49
  - as well.
• 3:51
  - First first, I'd say it's important to keep
• 3:52
  - in mind that, you know, AI tools,
• 3:55
  - are providing a point in time assessment
• 3:58
  - based on, you know, an employee census,
• 4:00
  - that's provided at the time of that initial
• 4:03
  - assessment. Right?
• 4:04
  - So I've seen situations
• 4:06
  - where,
• 4:06
  - you know, significant changes have occurred within a
• 4:09
  - census population between the time of the initial
• 4:12
  - assessment and the time of the final quote
• 4:14
  - being released.
• 4:15
  - So in one case, it was simply an
• 4:17
  - error. The employees from a single site in
• 4:20
  - a multi site organization were left off the
• 4:22
  - original census.
• 4:23
  - This stuff happens. Right? And in another case,
• 4:26
  - though, there was a material change in the
• 4:28
  - health of an employee that emerged
• 4:30
  - after the initial assessment. Right? So if you
• 4:32
  - think about, you know, how care is delivered
• 4:35
  - and how claims
• 4:37
  - appear within, you know, an ecosystem.
• 4:40
  - They happen, you know,
• 4:41
  - can happen between, you know, in that thirty
• 4:44
  - to sixty day window from an initial assessment
• 4:46
  - to a final quote. But in both cases,
• 4:49
  - the underwriter
• 4:50
  - prompted the review
• 4:52
  - because they were seeking the most current information
• 4:55
  - prior to releasing the final quote.
• 4:57
  - Right? That that quest for the most recent
• 4:59
  - information
• 5:00
  - is a key role that the underwriter plays
• 5:02
  - in the process.
• 5:04
  - Right? So they are,
• 5:05
  - you know, sitting in the driver's seat. They're
• 5:08
  - the ones, who are making a determination
• 5:11
  - to look for more information before they make
• 5:13
  - a decision to put a quote in market.
• 5:16
  - So you mentioned a a lot of it
• 5:17
  - is dependent on,
• 5:19
  - actually, the situation, obviously, where where they're at
• 5:21
  - at the point of time that will determine,
• 5:23
  - okay, I need to look at it again.
• 5:24
  - I I need to verify something. But when
• 5:26
  - we're going into that review, when we're going
• 5:28
  - into that that personal underwriter review,
• 5:31
  - what does it mean to have a meaningful
• 5:33
  - review of a model's recommendation, and what does
• 5:36
  - that look like in practice? And, again, what
• 5:38
  - are that in... What's that information that you're
• 5:40
  - looking for that an underwriter needs
• 5:43
  - to make that independent decision?
• 5:45
  - Yeah. Yeah. That's a... That's also a really
• 5:48
  - good good question here. So
• 5:51
  - in the case of our products, you know,
• 5:52
  - we provide multiple levels of information to help
• 5:55
  - the underwriter assess the risk of a group.
• 5:58
  - So at the highest level, we're providing
• 6:00
  - a risk score that's made up of a
• 6:02
  - demographic and morbid... Or morbidity score. The score
• 6:05
  - is expressed as a number where one is
• 6:07
  - viewed as an expected risk of groups like
• 6:09
  - the one that's being assessed.
• 6:12
  - A score of 1.2 then says that particular
• 6:15
  - group is 20%
• 6:16
  - more risky than expected.
• 6:18
  - That scores a... Is a quick way to
• 6:21
  - initially determine the depth of review need... Needed
• 6:24
  - for a group. The next level of information
• 6:26
  - within our tools highlights specific conditions, drugs,
• 6:29
  - and whether they're present, within that particular group.
• 6:32
  - You know, these can be,
• 6:34
  - ultimately customized by the underwriter to to use
• 6:37
  - flags so that the things they care about,
• 6:39
  - or concerned about can always be elevated and
• 6:42
  - highlighted,
• 6:43
  - so that it speeds up their initial review
• 6:45
  - process.
• 6:46
  - And then from there, you know, depending on
• 6:48
  - the level of information,
• 6:49
  - that they need to inform their decision, they
• 6:51
  - can drill down into
• 6:53
  - further details to understand prevalence of conditions and
• 6:56
  - drugs and the predicted costs,
• 6:58
  - expectations for that group on a forward looking
• 7:01
  - basis.
• 7:01
  - All that information is is great. Right? And
• 7:04
  - is super powerful,
• 7:05
  - to enable the underwriter.
• 7:07
  - However, what's not included
• 7:09
  - are the subjective inputs
• 7:11
  - that are also important components in making a
• 7:14
  - decision. For example,
• 7:16
  - tenure of client,
• 7:17
  - multiple product ownership, broker relationship,
• 7:20
  - size of the group, etcetera. Right?
• 7:23
  - Also are factors that that may
• 7:26
  - influence a decision.
• 7:27
  - So it's the underwriter's role to balance the
• 7:30
  - art, the subjective, and the science, the objective
• 7:32
  - to effectively manage their book of business.
• 7:36
  - Now this, might come as a surprise to
• 7:38
  - some, but we can disagree with AI.
• 7:42
  - It is it is it is allowed. Again,
• 7:44
  - when that disagreement happens, right, you you just
• 7:47
  - talk through the process. When an underwriters disagrees
• 7:50
  - with the model's recommendations, says, hey. There's some
• 7:52
  - factors I wanna consider here. I wanna change
• 7:53
  - that. How should that decision then be resolved
• 7:56
  - and documented?
• 7:57
  - Yeah. So the use of the word recommendation
• 8:00
  - here is interesting. You know, it almost feels
• 8:02
  - like the underwriter and the model are, you
• 8:04
  - know, on this... On a same level playing
• 8:06
  - field.
• 8:07
  - They're they're really not. Right?
• 8:09
  - You know, the model is providing information
• 8:11
  - which has been developed and curated to provide
• 8:14
  - inputs to the underwriter.
• 8:16
  - The underwriter makes the ultimate recommendation.
• 8:19
  - Right? It does vary by organization, but but
• 8:22
  - most organizations
• 8:23
  - have a workbench solution that their, their underwriters
• 8:27
  - are working in. And and typically, right, if
• 8:30
  - there is,
• 8:31
  - you know, disagreement
• 8:32
  - or additional information that they're factoring in that's
• 8:35
  - influencing their decision, they're documenting,
• 8:38
  - those types of things within that workbench so
• 8:40
  - that there's a kind of a longitudinal history
• 8:42
  - of the relationship with that particular prospect or
• 8:45
  - customer.
• 8:47
  - And I would expect, you know, they would
• 8:48
  - make note of, of a model, you know,
• 8:51
  - a a quote unquote, a model disagreement
• 8:53
  - like this. It's actually a a really important
• 8:56
  - part of the process because tracking
• 8:59
  - that type of disagreement
• 9:00
  - can help highlight if there's, if there's drift
• 9:03
  - in the model output or the underwriting parameters
• 9:06
  - that they need to address. Right? And so
• 9:08
  - it's an important feedback loop.
• 9:10
  - And when we engage with our customers and
• 9:12
  - and deliver our solutions, we're working with them,
• 9:16
  - right, to to gather that feedback over time
• 9:18
  - to ensure that, you know, we're calibrating
• 9:20
  - our output from our solutions that... To specifically
• 9:23
  - meet their needs.
• 9:25
  - Mark, it's so great to have you. I
• 9:27
  - I wanna hit on... Before we close our
• 9:29
  - conversation, I wanna hit on on one more
• 9:30
  - important thing here because you have a very
• 9:32
  - unique perspective on this. There is a lot
• 9:34
  - of vendor fatigue around AI right now, specifically
• 9:37
  - in health care. There are so many solutions.
• 9:41
  - People are bombarded
• 9:42
  - with
• 9:43
  - options when it comes to artificial
• 9:45
  - intelligence
• 9:46
  - for group health plans specifically.
• 9:49
  - When they are evaluating these AI underwriting tools
• 9:52
  - that can clearly help them as we've outlined,
• 9:54
  - what should they be looking for
• 9:57
  - in a vendor, in a partner, and what
• 9:59
  - are some of the red flags
• 10:01
  - that you're looking for specifically that a platform
• 10:03
  - isn't really optimized for speed over judgment?
• 10:07
  - Sure. Sure. I think,
• 10:09
  - it it really starts with, do the vendors
• 10:12
  - understand the market? Right? And so if you
• 10:14
  - truly understand the underwriting of the health care
• 10:17
  - market, you know it's a balance between art
• 10:19
  - and science.
• 10:20
  - Right? It's not a clear cut binary
• 10:23
  - yes, no decision in most cases.
• 10:26
  - A vendor that's solely focused on straight through
• 10:28
  - processing,
• 10:29
  - you know, to drive kind of OPEX savings
• 10:32
  - doesn't fully appreciate the subjective realities
• 10:35
  - of the underwriting within this world. So I'm
• 10:37
  - not saying there isn't a need or a
• 10:39
  - place for automation in health care underwriting. There
• 10:42
  - absolutely is. But the recognition that the underwriter
• 10:45
  - sits at the center of the ecosystem is
• 10:47
  - really important.
• 10:49
  - AI powered models and automation are ultimately
• 10:52
  - empowering them to focus their time and experience
• 10:55
  - on the quotes that need their expertise,
• 10:57
  - which is the art in the process.
• 11:00
  - So my advice, right, partnering with a vendor
• 11:02
  - who understands this dynamic will ensure
• 11:05
  - that you find that kind of that right
• 11:06
  - balance to maximize the value of the relationship
• 11:09
  - between the underwriter and then the technology that
• 11:12
  - you you use to empower them.
• 11:14
  - Mark, I feel like we struck that balance
• 11:16
  - here today between speed and judgment in our
• 11:18
  - episode.
• 11:20
  - It's it's so fantastic to have you. Thanks
• 11:22
  - so much for taking some time for us.
• 11:23
  - Oh, absolutely. I really appreciate it. Thank you.
• 11:26
  - And we also want to thank our podcast
• 11:28
  - sponsor, Gradient AI. You can tune in to
• 11:29
  - more podcasts from Becker's Health Care by visiting
• 11:31
  - our podcast page at beggershospitalreview.com.