WEBVTT

1
00:00:00.080 --> 00:00:02.799
- Hello, everyone. This is Erica Spicer Mason with

2
00:00:02.799 --> 00:00:03.699
- Becker's Healthcare.

3
00:00:04.000 --> 00:00:05.440
- Thank you so much for tuning in to

4
00:00:05.440 --> 00:00:07.540
- the Becker's Healthcare podcast series.

5
00:00:08.240 --> 00:00:10.160
- Today, we're going to talk about how health

6
00:00:10.160 --> 00:00:13.380
- plans can modernize their data and AI capabilities

7
00:00:13.679 --> 00:00:14.660
- to better manage

8
00:00:14.984 --> 00:00:16.684
- rising costs and utilization.

9
00:00:17.625 --> 00:00:19.785
- And joining me for this conversation is Mike

10
00:00:19.785 --> 00:00:22.425
- Sankey, vice president of health care and life

11
00:00:22.425 --> 00:00:24.844
- sciences go to market at Databricks.

12
00:00:25.625 --> 00:00:27.144
- Mike, it's so great to have you on

13
00:00:27.144 --> 00:00:28.744
- the podcast today. Thank you so much for

14
00:00:28.744 --> 00:00:31.010
- joining us. It's great to be here. Thanks

15
00:00:31.010 --> 00:00:31.829
- for having me.

16
00:00:32.289 --> 00:00:34.609
- Oh, we're glad to have you. And before

17
00:00:34.609 --> 00:00:36.369
- we talk a little bit more about health

18
00:00:36.369 --> 00:00:38.530
- plan strategy here, wanted to know if you

19
00:00:38.530 --> 00:00:40.049
- could share just a little bit more about

20
00:00:40.049 --> 00:00:41.670
- your work, your background,

21
00:00:42.289 --> 00:00:44.210
- whatever you think would be helpful for listeners

22
00:00:44.210 --> 00:00:45.774
- to know going into this conversation.

23
00:00:46.715 --> 00:00:49.774
- Yeah. So I guess briefly on Databricks, Databricks

24
00:00:49.994 --> 00:00:52.475
- is the data and AI platform. We work

25
00:00:52.475 --> 00:00:55.054
- across most of the Fortune 500.

26
00:00:55.434 --> 00:00:57.934
- In health care and life sciences, we have

27
00:00:58.075 --> 00:00:59.434
- more than 1,500

28
00:00:59.434 --> 00:00:59.934
- customers

29
00:01:00.929 --> 00:01:03.829
- who are using Databricks on a monthly basis,

30
00:01:04.769 --> 00:01:06.689
- and we do quite a bit of work

31
00:01:06.689 --> 00:01:09.090
- with health plans in The US that'll obviously

32
00:01:09.090 --> 00:01:10.150
- be the focus of

33
00:01:10.769 --> 00:01:13.729
- today's conversation. I've been at Databricks five and

34
00:01:13.729 --> 00:01:15.269
- a half years, and my role there

35
00:01:15.734 --> 00:01:17.655
- is really to take what we built, which

36
00:01:17.655 --> 00:01:19.755
- is a horizontal technology platform,

37
00:01:20.295 --> 00:01:23.255
- and help customers adopt it and really extend

38
00:01:23.255 --> 00:01:23.995
- it into

39
00:01:24.694 --> 00:01:27.174
- all of these industry specific use cases and

40
00:01:27.174 --> 00:01:27.674
- applications.

41
00:01:28.430 --> 00:01:30.189
- And we do that through a combination of

42
00:01:30.189 --> 00:01:33.009
- solutions that we build. So, for example, interoperability

43
00:01:33.469 --> 00:01:35.650
- libraries around FHIR, which are

44
00:01:36.109 --> 00:01:38.509
- very relevant to clinical data feeds for health

45
00:01:38.509 --> 00:01:39.009
- plans,

46
00:01:39.629 --> 00:01:41.810
- as well as through the ecosystem

47
00:01:42.109 --> 00:01:43.890
- that we cultivate. So there are

48
00:01:44.375 --> 00:01:46.775
- a lot of partners who are building on

49
00:01:46.775 --> 00:01:48.234
- Databricks. They may have

50
00:01:48.775 --> 00:01:51.275
- solutions that are specific to

51
00:01:52.055 --> 00:01:55.734
- industry around prior authorization or around fraud detection

52
00:01:55.734 --> 00:01:56.795
- that plug in

53
00:01:57.280 --> 00:01:59.299
- as well as a number of data providers,

54
00:01:59.359 --> 00:02:01.060
- and we connect to different

55
00:02:01.600 --> 00:02:03.380
- MCPs like with CMS,

56
00:02:04.240 --> 00:02:04.740
- that

57
00:02:05.359 --> 00:02:07.060
- really help our customers

58
00:02:07.680 --> 00:02:08.900
- use this technology

59
00:02:09.760 --> 00:02:12.314
- in industry. So that's a bit about me.

60
00:02:12.955 --> 00:02:14.635
- Oh, great to learn more about you, Mike,

61
00:02:14.635 --> 00:02:17.034
- and the work that you're doing. I'm sure

62
00:02:17.034 --> 00:02:18.715
- it's keeping you very busy because as we

63
00:02:18.715 --> 00:02:21.514
- know, so many organizations are working to either

64
00:02:21.514 --> 00:02:22.014
- implement

65
00:02:22.474 --> 00:02:24.875
- AI tools or or expand them and and

66
00:02:24.875 --> 00:02:26.474
- make sure that their data is ready to

67
00:02:26.474 --> 00:02:27.194
- do so. So,

68
00:02:28.269 --> 00:02:30.509
- so I wanna start high level here. We

69
00:02:30.509 --> 00:02:33.150
- know health plans are heading into 2027

70
00:02:33.150 --> 00:02:35.969
- facing some very steep cost trends.

71
00:02:36.430 --> 00:02:38.689
- You know, there's pressure coming from GLP ones.

72
00:02:38.989 --> 00:02:40.930
- There's rising behavioral health utilization,

73
00:02:41.549 --> 00:02:42.610
- shifting acuity.

74
00:02:43.324 --> 00:02:45.324
- I'm curious to know what you're hearing from

75
00:02:45.324 --> 00:02:47.724
- health plans about where their current tools or

76
00:02:47.724 --> 00:02:48.224
- capabilities

77
00:02:48.685 --> 00:02:50.444
- are currently falling short of some of those

78
00:02:50.444 --> 00:02:50.944
- pressures.

79
00:02:51.805 --> 00:02:53.644
- Yeah. It's probably worth just spending a couple

80
00:02:53.644 --> 00:02:55.504
- minutes on some of those macroeconomic

81
00:02:55.805 --> 00:02:57.960
- pressures that you mentioned because they're very real.

82
00:02:58.520 --> 00:03:00.920
- There are health plans operating at a financial

83
00:03:00.920 --> 00:03:03.240
- loss, including a number of health systems that

84
00:03:03.240 --> 00:03:04.540
- are acting as insurers

85
00:03:05.240 --> 00:03:07.340
- and trying to manage their own risk pools.

86
00:03:07.879 --> 00:03:09.260
- And a lot of the large,

87
00:03:10.200 --> 00:03:13.384
- you know, diversified health services, health plan organizations,

88
00:03:13.604 --> 00:03:15.145
- CVS, Humana, United,

89
00:03:15.925 --> 00:03:16.905
- Cigna, Centene,

90
00:03:17.365 --> 00:03:19.305
- they're all either exiting or downsizing

91
00:03:20.645 --> 00:03:22.965
- in the Affordable Care Act market or Medicare

92
00:03:22.965 --> 00:03:23.465
- Advantage.

93
00:03:24.724 --> 00:03:27.259
- So, you know, many people maybe outside of

94
00:03:27.259 --> 00:03:29.099
- the industry think that health plans are just

95
00:03:29.099 --> 00:03:30.560
- minting money year over year,

96
00:03:30.939 --> 00:03:33.500
- but the reality is relatively minor changes can

97
00:03:33.500 --> 00:03:35.340
- swing them from a profit to a loss,

98
00:03:35.340 --> 00:03:37.759
- and that's magnified in these larger plans

99
00:03:38.379 --> 00:03:40.159
- that have tens of millions of members.

100
00:03:41.144 --> 00:03:42.745
- So we take a plan that has 10,000,000

101
00:03:42.745 --> 00:03:44.984
- members, and let's say they make 1,000 in

102
00:03:44.984 --> 00:03:46.905
- profit from every member over the course of

103
00:03:46.905 --> 00:03:47.484
- the year.

104
00:03:47.944 --> 00:03:48.924
- That's 10,000,000,000.

105
00:03:49.305 --> 00:03:49.805
- But

106
00:03:50.104 --> 00:03:51.465
- that $1,000,

107
00:03:51.465 --> 00:03:53.864
- that could be a single month retail cost

108
00:03:53.864 --> 00:03:55.004
- of a GLP one.

109
00:03:55.400 --> 00:03:57.080
- It could be just a portion of the

110
00:03:57.080 --> 00:03:59.560
- facility fee for an ER visit, which wouldn't

111
00:03:59.560 --> 00:04:02.140
- include the physician's fee and the diagnostic test.

112
00:04:03.080 --> 00:04:05.260
- So the point being that, again, these utilization

113
00:04:05.400 --> 00:04:08.040
- changes across a large member population, they can

114
00:04:08.040 --> 00:04:09.180
- really make an impact.

115
00:04:09.775 --> 00:04:11.775
- And then we're combining that with the shifting

116
00:04:11.775 --> 00:04:14.835
- regulatory landscape in areas like Medicaid redetermination

117
00:04:15.935 --> 00:04:17.555
- and, more recently, the administration

118
00:04:17.855 --> 00:04:20.014
- ending, a nearly $10,000,000,000

119
00:04:20.014 --> 00:04:22.035
- subsidy for Medicare Part d.

120
00:04:22.810 --> 00:04:23.310
- So

121
00:04:23.610 --> 00:04:25.689
- all this means that there's a lot to

122
00:04:25.689 --> 00:04:27.389
- react to and a lot of these decisions

123
00:04:28.330 --> 00:04:30.970
- that health plans need to get right to

124
00:04:30.970 --> 00:04:34.110
- create benefit designs, to accurately forecast utilization,

125
00:04:35.014 --> 00:04:37.814
- and to help to manage these dynamic conditions

126
00:04:37.814 --> 00:04:39.574
- and changing member health over the course of

127
00:04:39.574 --> 00:04:41.194
- a year, and it's not trivial.

128
00:04:41.814 --> 00:04:43.975
- And many of these organizations that we work

129
00:04:43.975 --> 00:04:46.214
- with, I think, self admittedly are data rich

130
00:04:46.214 --> 00:04:47.435
- but insights poor.

131
00:04:48.089 --> 00:04:49.850
- In other words, they have massive amounts of

132
00:04:49.850 --> 00:04:51.850
- claim and member data, but they struggle to

133
00:04:51.850 --> 00:04:53.230
- process it quickly enough.

134
00:04:53.610 --> 00:04:56.430
- They struggle to apply additional intelligence and logic.

135
00:04:57.129 --> 00:04:59.610
- So an example of this could be somewhat

136
00:04:59.610 --> 00:05:02.415
- basic things like grouping all claims associated with

137
00:05:02.415 --> 00:05:04.355
- the knee replacement as a single unit.

138
00:05:04.894 --> 00:05:07.774
- And, fundamentally, they struggle to empower business leaders

139
00:05:07.774 --> 00:05:10.014
- that need to make these decisions against this

140
00:05:10.014 --> 00:05:11.555
- landscape that I talked about

141
00:05:11.935 --> 00:05:13.769
- with the right data and the right time.

142
00:05:14.649 --> 00:05:17.129
- So all of those, you know, contribute to

143
00:05:17.129 --> 00:05:18.910
- some of the challenges that they're facing.

144
00:05:19.930 --> 00:05:20.430
- Yeah.

145
00:05:20.889 --> 00:05:23.129
- I really appreciate how you broke down some

146
00:05:23.129 --> 00:05:24.970
- of these costs, Mike, and and what it

147
00:05:24.970 --> 00:05:28.074
- means on a larger scale. And you mentioned

148
00:05:28.074 --> 00:05:29.615
- some of these challenges around

149
00:05:30.395 --> 00:05:32.634
- being data rich and info poor. So I

150
00:05:32.634 --> 00:05:34.415
- wanna take that a little bit further.

151
00:05:35.274 --> 00:05:38.634
- Most plans already have claims data, pharmacy data,

152
00:05:38.634 --> 00:05:41.430
- authorization data. There's no shortage of it. But

153
00:05:41.589 --> 00:05:43.289
- I'm curious what you see

154
00:05:43.750 --> 00:05:44.250
- break

155
00:05:44.629 --> 00:05:46.329
- between having that data

156
00:05:46.629 --> 00:05:49.589
- and then catching a utilization shift within the

157
00:05:49.589 --> 00:05:50.649
- month that it happens

158
00:05:51.110 --> 00:05:53.350
- early enough for the health plan to act.

159
00:05:53.350 --> 00:05:54.649
- What is that looking like?

160
00:05:55.375 --> 00:05:57.634
- Yeah. I think a a lot of organizations

161
00:05:58.014 --> 00:06:00.194
- are stuck in this back and forth between

162
00:06:01.134 --> 00:06:03.454
- the technical teams that are building the pipelines

163
00:06:03.454 --> 00:06:05.615
- or building the data products or building the

164
00:06:05.615 --> 00:06:08.014
- dashboards and the business leaders that are actually

165
00:06:08.014 --> 00:06:08.834
- making decisions.

166
00:06:09.930 --> 00:06:12.490
- And people largely are trying to self serve

167
00:06:12.490 --> 00:06:13.310
- out of dashboards,

168
00:06:14.090 --> 00:06:16.750
- but they're not really built for deep analytics

169
00:06:16.970 --> 00:06:17.470
- and

170
00:06:18.009 --> 00:06:19.550
- reasoning and trying to

171
00:06:20.330 --> 00:06:22.750
- understand some of the root cause and dynamics

172
00:06:22.810 --> 00:06:25.164
- and actions that that an organization can take.

173
00:06:25.164 --> 00:06:25.664
- And

174
00:06:26.044 --> 00:06:29.004
- oftentimes, you'll see these leaders. They'll get frustrated.

175
00:06:29.004 --> 00:06:30.685
- They'll be clicking around in the dashboard. They

176
00:06:30.685 --> 00:06:32.604
- still don't get the answer they want without

177
00:06:32.604 --> 00:06:34.845
- needing a team to go and a technical

178
00:06:34.845 --> 00:06:35.824
- team that requires

179
00:06:36.599 --> 00:06:37.259
- rebuilding a

180
00:06:37.800 --> 00:06:39.899
- pipeline or data table or a dashboard.

181
00:06:41.159 --> 00:06:43.240
- And then what's happened is many of these

182
00:06:43.240 --> 00:06:44.060
- large organizations,

183
00:06:44.919 --> 00:06:45.419
- they've

184
00:06:45.800 --> 00:06:48.699
- grown to have shadow IT organizations, and that

185
00:06:49.254 --> 00:06:51.735
- exacerbates some of these challenges in identifying the

186
00:06:51.735 --> 00:06:52.634
- source of truth.

187
00:06:53.495 --> 00:06:54.855
- So I think a large part of it

188
00:06:54.855 --> 00:06:58.694
- is in the fundamental challenges these organizations have

189
00:06:58.694 --> 00:07:00.314
- had in actually democratizing

190
00:07:01.495 --> 00:07:03.514
- the the use of those data.

191
00:07:04.379 --> 00:07:06.860
- Yeah. That concept of democratizing the use of

192
00:07:06.860 --> 00:07:08.720
- data, Mike, I think is a really important

193
00:07:08.779 --> 00:07:09.279
- one.

194
00:07:09.819 --> 00:07:11.740
- And that leads me to ask, you know,

195
00:07:11.740 --> 00:07:12.560
- if nontechnical

196
00:07:12.860 --> 00:07:14.639
- teams are are accessing

197
00:07:15.339 --> 00:07:16.240
- important data,

198
00:07:16.664 --> 00:07:19.784
- how can they safely query live claims data

199
00:07:19.784 --> 00:07:20.284
- without,

200
00:07:20.584 --> 00:07:23.324
- you know, five people getting five different answers

201
00:07:23.384 --> 00:07:24.604
- at any given moment?

202
00:07:25.384 --> 00:07:28.425
- Yeah. That... That's one of the biggest areas

203
00:07:28.425 --> 00:07:31.164
- we've been spending time as an organization. It's

204
00:07:31.330 --> 00:07:32.470
- building the infrastructure

205
00:07:33.490 --> 00:07:36.770
- around that context layer that both humans and

206
00:07:36.770 --> 00:07:38.389
- AI agents need to succeed.

207
00:07:38.930 --> 00:07:40.150
- And if you think historically

208
00:07:40.529 --> 00:07:41.029
- of

209
00:07:41.810 --> 00:07:43.970
- where that context or that business logic has

210
00:07:43.970 --> 00:07:46.710
- lived, it's been hard coded in data pipelines

211
00:07:47.435 --> 00:07:50.555
- or living in disparate business intelligence dashboards or

212
00:07:50.555 --> 00:07:51.995
- it's in the head of Sally or Joe

213
00:07:51.995 --> 00:07:53.854
- or someone who's worked in the claims department

214
00:07:54.314 --> 00:07:55.375
- for twenty years.

215
00:07:56.074 --> 00:07:58.314
- And AI agents, in some ways, are even

216
00:07:58.314 --> 00:08:00.689
- worse than humans in that they're likely to

217
00:08:00.689 --> 00:08:02.370
- just make something up if they don't have

218
00:08:02.370 --> 00:08:03.589
- the appropriate context.

219
00:08:04.290 --> 00:08:06.470
- So what we've tried to do and help,

220
00:08:06.850 --> 00:08:09.009
- our health plan customers with is build the

221
00:08:09.009 --> 00:08:11.730
- capabilities to store that context at the data

222
00:08:11.730 --> 00:08:12.230
- layer

223
00:08:12.615 --> 00:08:14.935
- so that anyone or anything interacting with the

224
00:08:14.935 --> 00:08:17.574
- data, whether it's a human SQL programmer or

225
00:08:17.574 --> 00:08:19.035
- business intelligence dashboard

226
00:08:19.574 --> 00:08:21.495
- or an AI agent, they get the same

227
00:08:21.495 --> 00:08:21.995
- information

228
00:08:22.774 --> 00:08:24.235
- with the same shared definitions.

229
00:08:25.319 --> 00:08:28.279
- And this really requires both modeled semantics as

230
00:08:28.279 --> 00:08:29.579
- well as learned knowledge.

231
00:08:30.360 --> 00:08:32.860
- So let's say that someone wants to understand

232
00:08:33.720 --> 00:08:36.279
- the PMPM for commercial members who've had a

233
00:08:36.279 --> 00:08:37.019
- knee replacement.

234
00:08:37.720 --> 00:08:39.799
- We need a business definition of how to

235
00:08:39.799 --> 00:08:40.299
- calculate

236
00:08:40.754 --> 00:08:41.975
- per member per month.

237
00:08:42.835 --> 00:08:45.315
- We may need to store that acronym as

238
00:08:45.315 --> 00:08:45.815
- logic.

239
00:08:46.274 --> 00:08:48.514
- We also need to define the time period

240
00:08:48.514 --> 00:08:50.695
- of how we calculate that in a consistent

241
00:08:50.754 --> 00:08:53.475
- way, whether that's the most recent fiscal year

242
00:08:53.475 --> 00:08:55.174
- to date or a rolling twelve months.

243
00:08:55.740 --> 00:08:57.659
- We also would need a definition of commercial

244
00:08:57.659 --> 00:08:59.659
- line of business. That's probably a grouping of

245
00:08:59.659 --> 00:09:00.799
- many different segments

246
00:09:01.580 --> 00:09:04.379
- across a health plan, in terms of their

247
00:09:04.379 --> 00:09:06.720
- employer business and administrative services.

248
00:09:07.179 --> 00:09:09.580
- And all these definitions need to be stored,

249
00:09:09.580 --> 00:09:10.799
- accessible, and governed.

250
00:09:11.465 --> 00:09:13.945
- And we have features called business semantics and

251
00:09:13.945 --> 00:09:15.644
- metric views that support these.

252
00:09:16.504 --> 00:09:18.264
- The other thing that we need to get

253
00:09:18.264 --> 00:09:19.945
- right is we need to know the correct

254
00:09:19.945 --> 00:09:22.345
- tables to pull from because you can rest

255
00:09:22.345 --> 00:09:24.185
- assured that a health plan has dozens of

256
00:09:24.185 --> 00:09:26.425
- fields across many tables that are just labeled

257
00:09:26.425 --> 00:09:26.879
- claim,

258
00:09:27.440 --> 00:09:29.779
- And we've created capabilities to understand

259
00:09:30.480 --> 00:09:32.639
- the most relevant data sources and rank them

260
00:09:32.639 --> 00:09:34.259
- and help teams certify them.

261
00:09:34.639 --> 00:09:36.720
- So relevance, for example, could be based on

262
00:09:36.720 --> 00:09:38.720
- how frequently a given table is used to

263
00:09:38.720 --> 00:09:40.740
- populate reports for the CFO.

264
00:09:41.414 --> 00:09:43.495
- And all these things together, they create that

265
00:09:43.495 --> 00:09:46.615
- governed data intelligence substrate to answer the questions

266
00:09:46.615 --> 00:09:48.315
- more consistently and correctly.

267
00:09:48.855 --> 00:09:50.774
- So in your question of if five people

268
00:09:50.774 --> 00:09:51.834
- ask the same question,

269
00:09:52.294 --> 00:09:54.394
- if they're able to refer to the same

270
00:09:55.080 --> 00:09:55.820
- business definitions,

271
00:09:56.200 --> 00:09:57.820
- logic, and underlying

272
00:09:58.680 --> 00:10:00.519
- data, then they should get the same answer

273
00:10:00.519 --> 00:10:01.580
- time and time again.

274
00:10:03.800 --> 00:10:05.240
- It makes a lot of sense, Mike, and

275
00:10:05.240 --> 00:10:07.639
- I really appreciate you walking our listeners through

276
00:10:07.639 --> 00:10:08.085
- that.

277
00:10:08.644 --> 00:10:10.004
- And I wanna back up a little bit

278
00:10:10.004 --> 00:10:12.825
- too and talk a bit more about adopting

279
00:10:12.965 --> 00:10:14.264
- AI in the first place.

280
00:10:14.965 --> 00:10:18.165
- Adoption of AI tools is essentially universal. I

281
00:10:18.165 --> 00:10:19.605
- think it's safe to say at this point,

282
00:10:19.605 --> 00:10:21.304
- especially in the health plan space,

283
00:10:21.925 --> 00:10:22.665
- but readiness

284
00:10:23.045 --> 00:10:23.545
- for

285
00:10:24.059 --> 00:10:26.700
- various AI tools is another story. So where

286
00:10:26.700 --> 00:10:29.019
- are you seeing health plans get stuck when

287
00:10:29.019 --> 00:10:30.080
- it comes to adoption?

288
00:10:31.259 --> 00:10:33.580
- I think that context layer that I described

289
00:10:33.580 --> 00:10:36.379
- is probably the single biggest rate limiting factor

290
00:10:36.379 --> 00:10:37.039
- right now.

291
00:10:37.894 --> 00:10:40.774
- And we believe organizations need four things to

292
00:10:40.774 --> 00:10:41.835
- succeed with AI.

293
00:10:42.855 --> 00:10:44.535
- We refer to them as the four c's

294
00:10:44.535 --> 00:10:46.315
- if you need a memory device.

295
00:10:47.095 --> 00:10:49.335
- So context is one. We've talked quite a

296
00:10:49.335 --> 00:10:50.394
- bit about that.

297
00:10:50.774 --> 00:10:53.014
- And the other three are control, cost, and

298
00:10:53.014 --> 00:10:53.514
- choice.

299
00:10:54.269 --> 00:10:56.190
- So let's spend a bit of time on

300
00:10:56.190 --> 00:10:58.509
- the governance challenges facing health plans and health

301
00:10:58.509 --> 00:11:00.830
- care in general because those really get to

302
00:11:00.830 --> 00:11:02.049
- the concepts around

303
00:11:02.750 --> 00:11:04.209
- control, costs, and choice.

304
00:11:04.669 --> 00:11:07.070
- There's not a single unifying standard for AI

305
00:11:07.070 --> 00:11:09.684
- in health care, and states continue to introduce

306
00:11:09.684 --> 00:11:12.404
- their own legislation, which really complicates things for

307
00:11:12.404 --> 00:11:12.904
- plans

308
00:11:13.445 --> 00:11:15.384
- that operate across multiple states.

309
00:11:15.924 --> 00:11:16.904
- So organizations

310
00:11:17.205 --> 00:11:20.644
- need a robust framework for evaluations, including around

311
00:11:20.644 --> 00:11:21.144
- bias

312
00:11:21.639 --> 00:11:23.579
- and observability of these models.

313
00:11:24.199 --> 00:11:26.519
- And just like with their human resources, they

314
00:11:26.519 --> 00:11:28.679
- need to control what agents access and what

315
00:11:28.679 --> 00:11:29.740
- they do with the data.

316
00:11:30.199 --> 00:11:32.120
- Add all of this together, and many health

317
00:11:32.120 --> 00:11:33.959
- plans, they can't keep pace with the changes

318
00:11:33.959 --> 00:11:35.899
- among the Frontier AI labs.

319
00:11:36.475 --> 00:11:38.654
- So they really need to focus on building

320
00:11:39.274 --> 00:11:42.475
- an infrastructure that preserves model choice and gives

321
00:11:42.475 --> 00:11:44.554
- them the ability to work across models, which

322
00:11:44.554 --> 00:11:47.434
- also helps them manage cost because it enables

323
00:11:47.434 --> 00:11:49.355
- them to send trivial tasks to lower cost

324
00:11:49.355 --> 00:11:49.855
- models.

325
00:11:50.639 --> 00:11:53.440
- So beyond getting the context and this data

326
00:11:53.440 --> 00:11:55.759
- substrate right, I would say most organizations are

327
00:11:55.759 --> 00:11:57.700
- struggling with governance at scale,

328
00:11:58.480 --> 00:12:00.340
- as it pertains to controlling,

329
00:12:01.440 --> 00:12:04.065
- what agents are accessing, how they're accessing it,

330
00:12:04.225 --> 00:12:07.504
- the cost around it, and then model choice

331
00:12:07.504 --> 00:12:09.205
- and choice and cost are

332
00:12:09.585 --> 00:12:11.445
- very closely intertwined at this point.

333
00:12:11.825 --> 00:12:12.325
- Mhmm.

334
00:12:13.024 --> 00:12:15.024
- Yeah. And to your point too, Mike, about

335
00:12:15.024 --> 00:12:17.149
- there not being one universal standard,

336
00:12:17.610 --> 00:12:18.909
- how so many of these regulations

337
00:12:19.289 --> 00:12:21.129
- around AI and governance can vary state to

338
00:12:21.129 --> 00:12:21.629
- state

339
00:12:21.929 --> 00:12:24.909
- makes it very complicated and complex for plans.

340
00:12:25.449 --> 00:12:26.970
- I'm curious to know a little bit more

341
00:12:26.970 --> 00:12:29.289
- about the plans that you've seen get the

342
00:12:29.289 --> 00:12:30.829
- governance piece right.

343
00:12:31.424 --> 00:12:32.865
- What did they have to put in place

344
00:12:32.865 --> 00:12:35.445
- before people could trust the data they were

345
00:12:35.504 --> 00:12:37.764
- working from? What does that look like?

346
00:12:39.184 --> 00:12:39.684
- So,

347
00:12:40.304 --> 00:12:42.065
- for this, maybe I can just share a

348
00:12:42.065 --> 00:12:43.524
- couple of customer examples.

349
00:12:44.120 --> 00:12:46.360
- We hosted our big data and AI summit

350
00:12:46.360 --> 00:12:48.279
- earlier this summer. We had so many health

351
00:12:48.279 --> 00:12:50.700
- plan customers attend and share what they're doing.

352
00:12:51.559 --> 00:12:53.480
- So one of those was Humana, and they've

353
00:12:53.480 --> 00:12:55.959
- been a Databricks customer for a while, including

354
00:12:55.959 --> 00:12:57.980
- in support of their value based care initiatives.

355
00:12:58.615 --> 00:13:00.875
- And they recently expanded their use of Databricks

356
00:13:01.014 --> 00:13:02.394
- to the office of the CFO.

357
00:13:03.014 --> 00:13:05.095
- And the entire focus of their presentation was

358
00:13:05.095 --> 00:13:08.054
- about rebuilding finance on a trusted governed data

359
00:13:08.054 --> 00:13:08.554
- foundation

360
00:13:09.095 --> 00:13:09.835
- and operationalizing

361
00:13:10.295 --> 00:13:11.915
- AI across these different functions.

362
00:13:12.559 --> 00:13:15.519
- And for them, that meant, including domain product

363
00:13:15.519 --> 00:13:18.799
- owners from underwriting, from pricing, from risk assessment,

364
00:13:18.799 --> 00:13:20.100
- from revenue management,

365
00:13:20.799 --> 00:13:23.459
- from areas like cost and expense management,

366
00:13:24.079 --> 00:13:26.799
- and having those domain owners define the business

367
00:13:26.799 --> 00:13:29.664
- semantics and the metrics to feed a unified

368
00:13:29.664 --> 00:13:30.565
- semantic model

369
00:13:31.024 --> 00:13:32.485
- that could be used across,

370
00:13:33.264 --> 00:13:35.024
- the office of the CFO. So it really

371
00:13:35.024 --> 00:13:36.485
- was this combination of

372
00:13:36.945 --> 00:13:40.164
- the technical foundation, but also the intelligence

373
00:13:41.279 --> 00:13:43.459
- from their team and from these different functions

374
00:13:43.519 --> 00:13:44.179
- to define

375
00:13:44.799 --> 00:13:46.100
- that context layer.

376
00:13:46.959 --> 00:13:49.519
- The second example I'll share is from Centene,

377
00:13:49.519 --> 00:13:51.220
- and they presented on

378
00:13:51.759 --> 00:13:54.754
- the gifts of machine agentic governance intelligence, which

379
00:13:54.754 --> 00:13:56.914
- they shorthand refer to as the gifts of

380
00:13:56.914 --> 00:13:57.575
- the Magi.

381
00:13:58.274 --> 00:14:00.934
- And it was all about delivering secure, repeatable,

382
00:14:01.075 --> 00:14:02.455
- and explainable AI.

383
00:14:03.075 --> 00:14:05.735
- And this included a lot of different components

384
00:14:05.794 --> 00:14:08.754
- for security. It included things like usage tracking

385
00:14:08.754 --> 00:14:09.414
- and monitoring,

386
00:14:10.040 --> 00:14:12.200
- a single interface to all external models so

387
00:14:12.200 --> 00:14:13.019
- they could track

388
00:14:13.559 --> 00:14:17.179
- how the models were hitting their data effectively,

389
00:14:18.120 --> 00:14:19.740
- as well as comprehensive guardrails.

390
00:14:20.519 --> 00:14:22.855
- For explainability, and this gets into some of

391
00:14:22.855 --> 00:14:25.174
- the governance and the regulations, they create an

392
00:14:25.174 --> 00:14:27.595
- immutable record behind every AI decision.

393
00:14:27.975 --> 00:14:30.214
- So they capture and lock everything, including the

394
00:14:30.214 --> 00:14:32.774
- prompts, the responses, the tool calls in these

395
00:14:32.774 --> 00:14:33.995
- multi agent systems,

396
00:14:34.519 --> 00:14:36.840
- and they have an evaluation process that's actually

397
00:14:36.840 --> 00:14:38.779
- a combination of automated scores

398
00:14:39.159 --> 00:14:40.220
- and human review.

399
00:14:40.600 --> 00:14:42.919
- And they're able to continuously monitor this life

400
00:14:42.919 --> 00:14:45.879
- cycle for things like model drift and policy

401
00:14:45.879 --> 00:14:46.379
- infractions.

402
00:14:47.554 --> 00:14:50.434
- And finally, in terms of repeatability, they've deployed

403
00:14:50.434 --> 00:14:52.995
- the scaffolding and created a library of reusable

404
00:14:52.995 --> 00:14:53.495
- templates

405
00:14:54.034 --> 00:14:56.054
- to ensure consistency and correctness.

406
00:14:56.595 --> 00:14:59.074
- So this includes the context layer, which we've

407
00:14:59.074 --> 00:15:00.375
- talked about quite a bit

408
00:15:00.959 --> 00:15:03.059
- to support how agents retrieve information.

409
00:15:03.759 --> 00:15:05.519
- And what this has done is it's created

410
00:15:05.519 --> 00:15:08.319
- an operating velocity that enables them to rapidly

411
00:15:08.319 --> 00:15:10.899
- deploy across different areas of their business.

412
00:15:11.360 --> 00:15:13.039
- A recent example of this was a new

413
00:15:13.039 --> 00:15:15.759
- solution for care managers to support interactions with

414
00:15:15.759 --> 00:15:16.259
- members

415
00:15:16.774 --> 00:15:18.554
- and help to close gaps in care.

416
00:15:19.335 --> 00:15:21.914
- And, Mike, I I really appreciate the detailed

417
00:15:22.054 --> 00:15:24.375
- examples from both of those health plans. I

418
00:15:24.375 --> 00:15:25.914
- think that's going to be really helpful,

419
00:15:26.375 --> 00:15:28.054
- for folks listening in who are, you know,

420
00:15:28.054 --> 00:15:29.894
- just in the early stages of this journey

421
00:15:29.894 --> 00:15:31.940
- and to have that concrete view of what

422
00:15:31.940 --> 00:15:32.839
- this looks like,

423
00:15:33.459 --> 00:15:35.620
- and also appreciate the details you shared around

424
00:15:35.620 --> 00:15:37.299
- the four c's as well. And, again, for

425
00:15:37.299 --> 00:15:39.879
- listeners, when we talk about AI adoption, readiness

426
00:15:39.940 --> 00:15:42.980
- looks like those four c's, context, control, cost,

427
00:15:42.980 --> 00:15:43.480
- choice.

428
00:15:44.565 --> 00:15:46.404
- Mike, a lot of valuable insight you've given

429
00:15:46.404 --> 00:15:48.245
- us today. So before I let you go,

430
00:15:48.245 --> 00:15:49.845
- I wanted to kind of get your look

431
00:15:49.845 --> 00:15:51.225
- ahead to 2030.

432
00:15:52.004 --> 00:15:55.225
- What capabilities around data, AI, and cost visibility

433
00:15:55.860 --> 00:15:58.019
- will become table stakes for health plans, and

434
00:15:58.019 --> 00:16:00.360
- what should leaders be building toward right now?

435
00:16:01.299 --> 00:16:03.220
- Yeah. So I would definitely encourage leaders to

436
00:16:03.220 --> 00:16:06.580
- think about building that proper foundation so they

437
00:16:06.580 --> 00:16:08.759
- can achieve that scale velocity.

438
00:16:09.394 --> 00:16:11.794
- And if we think about 2030, we're not

439
00:16:11.794 --> 00:16:13.394
- that far away. We're looking a few years

440
00:16:13.394 --> 00:16:15.715
- out. What will be table stakes for health

441
00:16:15.715 --> 00:16:16.215
- plans?

442
00:16:16.995 --> 00:16:19.014
- And I'll throw out a few ideas here

443
00:16:19.075 --> 00:16:20.674
- that I don't think are that far fetched.

444
00:16:20.674 --> 00:16:23.154
- I think many organizations are already on this

445
00:16:23.154 --> 00:16:23.654
- journey.

446
00:16:24.110 --> 00:16:25.950
- The first is around fraud detection, and I

447
00:16:25.950 --> 00:16:27.629
- think we're gonna move to a world where

448
00:16:27.629 --> 00:16:30.370
- fraud is detected before the claim is paid

449
00:16:30.990 --> 00:16:33.470
- rather than the catch and chase model that,

450
00:16:33.470 --> 00:16:35.889
- you know, many organizations operate in today.

451
00:16:36.595 --> 00:16:38.914
- And that would mean models are scoring every

452
00:16:38.914 --> 00:16:41.575
- claim prepayment. They're stopping improper payments.

453
00:16:42.035 --> 00:16:44.855
- They're embedding AI to do things like automatically

454
00:16:44.995 --> 00:16:46.995
- generate letters if they have to request more

455
00:16:46.995 --> 00:16:47.495
- information.

456
00:16:47.955 --> 00:16:49.955
- And, ultimately, the goal is to reduce fraud,

457
00:16:49.955 --> 00:16:52.389
- waste, and abuse to low single digits. Right

458
00:16:52.389 --> 00:16:54.470
- now, for many organizations, it's actually in the

459
00:16:54.470 --> 00:16:57.029
- high single digits, and that has a major

460
00:16:57.029 --> 00:16:59.050
- impact on their operating margins.

461
00:16:59.990 --> 00:17:02.550
- The second is around prior authorization. Obviously, has

462
00:17:02.550 --> 00:17:03.050
- has,

463
00:17:03.590 --> 00:17:05.289
- gotten a lot of attention.

464
00:17:06.105 --> 00:17:07.804
- Recently, there are a lot of regulations

465
00:17:08.265 --> 00:17:11.484
- around standards coming out from CMS,

466
00:17:12.505 --> 00:17:14.345
- and I'd like to see a world where

467
00:17:14.345 --> 00:17:16.924
- prior authorization is 80% plus automated.

468
00:17:17.545 --> 00:17:19.750
- Today, it's probably about half of that. It

469
00:17:19.750 --> 00:17:21.829
- depends, of course, on the organization and the

470
00:17:21.829 --> 00:17:22.730
- book of business.

471
00:17:23.190 --> 00:17:24.730
- And this would mean gold carding,

472
00:17:25.190 --> 00:17:26.730
- more top quality providers,

473
00:17:27.109 --> 00:17:29.609
- having agents auto adjudicate most requests,

474
00:17:29.990 --> 00:17:32.630
- and shortening that median turnaround time to under

475
00:17:32.630 --> 00:17:34.634
- twenty four hours, even under twelve hours in

476
00:17:34.714 --> 00:17:35.615
- instead of days.

477
00:17:37.275 --> 00:17:39.994
- The third area is around the payment models,

478
00:17:39.994 --> 00:17:40.494
- and

479
00:17:40.875 --> 00:17:42.315
- I'd like to see a world where all

480
00:17:42.315 --> 00:17:44.795
- payments or nearly all payments are value based

481
00:17:44.795 --> 00:17:46.875
- and every dollar is tied to outcomes with

482
00:17:46.875 --> 00:17:48.575
- those quality metrics aligned

483
00:17:49.230 --> 00:17:50.690
- across plans and providers.

484
00:17:51.630 --> 00:17:54.190
- And the fourth is around using AI and

485
00:17:54.190 --> 00:17:57.089
- supporting a virtual concierge for every member.

486
00:17:57.470 --> 00:18:00.049
- A personalized assistant that stays with the member,

487
00:18:00.109 --> 00:18:01.569
- understands their history,

488
00:18:01.974 --> 00:18:04.214
- and can help resolve payment, benefit, and care

489
00:18:04.214 --> 00:18:06.934
- questions end to end across every channel. And,

490
00:18:06.934 --> 00:18:08.615
- again, I don't think that's far fetched. We're

491
00:18:08.615 --> 00:18:09.595
- seeing this embedded

492
00:18:10.134 --> 00:18:11.914
- in many different functions already,

493
00:18:12.294 --> 00:18:14.875
- within health plans, and I think having that

494
00:18:15.335 --> 00:18:16.554
- personalized concierge

495
00:18:17.299 --> 00:18:19.159
- will really improve the member experience.

496
00:18:20.419 --> 00:18:22.659
- Mike, I wanna thank you again for all

497
00:18:22.659 --> 00:18:25.319
- of the insight and the practical guidance you've

498
00:18:25.380 --> 00:18:27.380
- you've shared with our audience today and and

499
00:18:27.380 --> 00:18:29.079
- for the time that you spent with Becker's.

500
00:18:29.619 --> 00:18:31.619
- I've learned a lot from our conversation, and

501
00:18:31.619 --> 00:18:33.884
- I'm sure our listeners from health plans will

502
00:18:33.884 --> 00:18:35.565
- feel the same. So thank you again for

503
00:18:35.565 --> 00:18:36.384
- your time today.

504
00:18:36.765 --> 00:18:39.724
- Yeah. Appreciate it. Take care. And we'd also

505
00:18:39.724 --> 00:18:42.044
- like to thank our podcast sponsor for today,

506
00:18:42.044 --> 00:18:42.544
- Databricks.

507
00:18:43.415 --> 00:18:45.095
- Listeners, be sure to tune in to more

508
00:18:45.095 --> 00:18:48.055
- podcasts from Becker's by visiting our podcast page

509
00:18:48.055 --> 00:18:50.635
- at beckershospitalreview.com.