Trascrizione completa di Inside Darwinbox’s Cortex: Building an AI-Native HCM

help us understand what that is? 81 00:03:47.460 --> 00:03:50.599 Yeah. So when we went on this journey, we realized 82 00:03:50.659 --> 00:03:56.389 that the LLMs were increasingly getting smarter. So we were 83 00:03:56.409 --> 00:03:59.990 looking at these really smart agents of sorts and we 84 00:04:00.030 --> 00:04:02.550 were pointing them to the product and we were asking 85 00:04:02.569 --> 00:04:05.389 them to do things. Now imagine if you were an 86 00:04:05.509 --> 00:04:08.830 HR tech operator and you were asked to bulk upload 87 00:04:08.909 --> 00:04:10.229 some employee information. 88 00:04:10.310 --> 00:04:10.830 What would you do? 89 00:04:11.229 --> 00:04:12.650 You would click on a folder and then you would 90 00:04:12.669 --> 00:04:14.719 click on another subfolder and then go to a certain 91 00:04:15.009 --> 00:04:17.220 special place in your app. By the way, there are 92 00:04:17.240 --> 00:04:19.790 like thousands of these special places in the app where 93 00:04:19.810 --> 00:04:21.870 you kind of make these little transactions. 94 00:04:21.949 --> 00:04:22.149 Yeah. 95 00:04:22.410 --> 00:04:24.370 And we were just expecting that the agents would understand 96 00:04:24.410 --> 00:04:27.209 what to do with an intent. But turns out that 97 00:04:27.250 --> 00:04:30.970 wasn't the case. These were very, very smart models. 98 00:04:31.069 --> 00:04:31.310 Yeah. 99 00:04:31.589 --> 00:04:33.230 But they couldn't figure out what the hell is going 100 00:04:33.290 --> 00:04:34.129 on with your organization. 101 00:04:34.389 --> 00:04:34.629 Yeah. 102 00:04:34.829 --> 00:04:36.990 Or with the instance that's set up for you. So 103 00:04:37.009 --> 00:04:37.990 we realized that it was. 104 00:04:38.000 --> 00:04:41.839 Akin to hiring a really smart management trainee. They're very 105 00:04:41.879 --> 00:04:43.800 keen to work. They're very eager to work. But they 106 00:04:43.839 --> 00:04:45.759 don't know where to go to make a certain change. 107 00:04:45.819 --> 00:04:49.120 Yeah. So we essentially created. 108 00:04:48.480 --> 00:04:51.779 A construct that we call context graph, which is essentially 109 00:04:51.819 --> 00:04:56.560 the system describing itself in a meticulous detail to any 110 00:04:56.699 --> 00:05:00.199 AI that wants to interact with Davenport's HCM. Which means 111 00:05:00.240 --> 00:05:03.000 when you now ask a question as to, oh, can 112 00:05:03.040 --> 00:05:06.160 you create a new hiring workflow for me? It knows 113 00:05:06.319 --> 00:05:09.720 how to do that because the entirety of the context, 114 00:05:09.779 --> 00:05:14.399 the past configurations, you know, good HR knowledge is all 115 00:05:14.560 --> 00:05:16.139 encoded into this context graph. 116 00:05:16.600 --> 00:05:18.120 So one quick consult, it. 117 00:05:18.120 --> 00:05:22.060 Knows how to work the product, how your organization is 118 00:05:22.100 --> 00:05:25.319 set up, and what good HR practice looks like. So 119 00:05:25.350 --> 00:05:29.029 you get an output that's a lot more accurate. And that, 120 00:05:29.209 --> 00:05:32.750 we realized, is a missing piece in the industry. And 121 00:05:32.790 --> 00:05:35.490 we wanted first to kind of come up with this construct. 122 00:05:36.810 --> 00:05:38.930 And this allowed us to sort of take the oven 123 00:05:38.949 --> 00:05:43.439 box and make it available in Cloud and Copilot and 124 00:05:43.540 --> 00:05:46.939 Slack and essentially make it headless because all you needed 125 00:05:46.980 --> 00:05:49.370 to now do was speak to the system. 126 00:05:49.649 --> 00:05:49.829 Right. 127 00:05:50.029 --> 00:05:51.410 And the content staff would figure the rest out. 128 00:05:51.449 --> 00:05:57.009 Yeah. So you mentioned outputs, right? So like the outputs 129 00:05:57.029 --> 00:05:57.970 that the AI is producing. 130 00:05:58.149 --> 00:05:59.970 And one of the outputs that it produces is something 131 00:05:59.990 --> 00:06:03.040 what you all are calling labels for managers based on 132 00:06:03.079 --> 00:06:07.029 the signals that the system is identifying. So can you 133 00:06:07.209 --> 00:06:09.670 share a little bit about what those playbooks are and 134 00:06:09.730 --> 00:06:12.629 what do they, at a very basic level, enable managers 135 00:06:12.670 --> 00:06:12.910 to do? 136 00:06:13.750 --> 00:06:14.350 It's quite simple. 137 00:06:15.449 --> 00:06:20.110 There's constantly stuff going on and changing in your team's context, right? 138 00:06:20.649 --> 00:06:23.689 One of your best performers is leaving. What do you do? 139 00:06:24.310 --> 00:06:27.589 I'm now running around helter-skelter trying to figure out how 140 00:06:27.610 --> 00:06:29.329 to talk to my comp guy, how to talk to 141 00:06:29.360 --> 00:06:31.800 my recruitment manager, how to talk to my talent management 142 00:06:31.819 --> 00:06:34.379 guy and figure out what went wrong. 143 00:06:35.470 --> 00:06:37.649 Figure out how to mitigate this problem statement. 144 00:06:38.949 --> 00:06:43.110 Essentially, what we did with Playbooks was we allowed Cortex 145 00:06:43.170 --> 00:06:46.209 the ability to understand your context and literally give you 146 00:06:46.329 --> 00:06:49.379 a series of steps that you could do to mitigate 147 00:06:49.399 --> 00:06:50.800 a critical workforce situation. 148 00:06:51.480 --> 00:06:54.699 Something like attrition, for instance, right? What do you do? 149 00:06:54.980 --> 00:06:59.860 Oh, turns out that they're being paid under market. Turns 150 00:07:00.000 --> 00:07:03.180 out that they recently had a manager change that is 151 00:07:03.220 --> 00:07:05.620 kind of dipping that sentiment a little bit. Turns out 152 00:07:05.660 --> 00:07:08.620 that they've been overworking themselves and have not sort of 153 00:07:08.660 --> 00:07:10.540 had a good balance in the last few years. 154 00:07:10.660 --> 00:07:10.959 Yeah. 155 00:07:11.259 --> 00:07:13.560 So Cortex then understands

all of this context and gives 156 00:07:13.579 --> 00:07:15.839 you a quick playbook. Hey, you've got to assign a 157 00:07:15.860 --> 00:07:19.199 mandatory leave in the next one month. You've got to 158 00:07:19.339 --> 00:07:22.259 sort of do a quick check-in with the HRBP because 159 00:07:22.300 --> 00:07:24.500 there's a new manager change and that has led to 160 00:07:24.519 --> 00:07:28.420 some sort of sentiment. Right. The next step, send an 161 00:07:28.459 --> 00:07:31.560 alert to the compensation planning committee and sort of get 162 00:07:31.579 --> 00:07:32.490 a mitigation in place. 163 00:07:32.519 --> 00:07:32.850 Right. 164 00:07:32.879 --> 00:07:33.000 Yeah. 165 00:07:33.399 --> 00:07:36.879 So Playbook literally is just a series of steps that 166 00:07:36.899 --> 00:07:41.029 a smart HR consultant would give to mitigate any situation 167 00:07:41.069 --> 00:07:43.589 to your managers or your HR VPs. Except the smart 168 00:07:43.629 --> 00:07:47.410 HR consultant is now available 24-7 to anyone who wants 169 00:07:47.430 --> 00:07:48.550 to solve a workforce problem. 170 00:07:48.769 --> 00:07:49.750 And that's what Cortex does. 171 00:07:50.089 --> 00:07:51.870 And of course, there's like a lot more nuance that's 172 00:07:51.910 --> 00:07:52.829 going into producing. 173 00:07:52.850 --> 00:07:54.149 Absolutely, yeah. 174 00:07:54.240 --> 00:07:58.279 Okay, so you all launched Cortex back in August. So 175 00:07:58.560 --> 00:08:01.399 what have you seen customers do differently with it now 176 00:08:01.459 --> 00:08:02.399 that they weren't doing it before? 177 00:08:02.970 --> 00:08:06.069 First of all, we launched Cortex on the 4th of August. 178 00:08:07.240 --> 00:08:11.899 In about three weeks, 50% of our customer base wanted 179 00:08:11.920 --> 00:08:12.420 to get on Cortex. 180 00:08:12.480 --> 00:08:12.779 Oh, wow. 181 00:08:13.019 --> 00:08:18.120 Yeah, we registered an insane amount of interest. So, we're overwhelmed. Yeah. 182 00:08:18.540 --> 00:08:21.980 I'll be honest. There's a lot going on. What turns out, 183 00:08:23.459 --> 00:08:25.250 the value proposition wasn't hard to buy. 184 00:08:25.589 --> 00:08:25.750 Right. 185 00:08:25.870 --> 00:08:28.529 It just makes sense, you know. Why should building take 186 00:08:28.569 --> 00:08:29.170 so much time? 187 00:08:29.329 --> 00:08:29.550 Right. 188 00:08:29.899 --> 00:08:33.019 I should just be able to wish things into existence. 189 00:08:33.220 --> 00:08:35.240 You know, I want a new workflow. I shouldn't be 190 00:08:35.259 --> 00:08:37.379 waiting on HR IT teams. I shouldn't be waiting on 191 00:08:37.759 --> 00:08:40.210 a consultant to come configure it for me. So the 192 00:08:40.250 --> 00:08:44.730 value proposition was easily sold. Everyone wanted this utopia. But 193 00:08:44.750 --> 00:08:47.149 the real challenge is getting it on the tracks. You know, 194 00:08:47.750 --> 00:08:51.090 the logistics around it. How do you model what your 195

00:08:51.129 --> 00:08:52.129 cost is going to look like? 196 00:08:52.389 --> 00:08:52.629 Right? 197 00:08:52.990 --> 00:08:55.590 How do you choose which workflow to play around with? 198 00:08:56.409 --> 00:08:58.789 How do you kind of be judicious? How do you 199 00:08:59.220 --> 00:09:06.070 be responsible about deploying some of these things. So unsurprisingly, 200 00:09:07.090 --> 00:09:11.350 I know that was not your question, but unsurprisingly, service 201 00:09:11.370 --> 00:09:14.450 delivery was one of the first avenues that all of 202 00:09:15.049 --> 00:09:17.850 our design partners and launch partners tackled. 203 00:09:18.330 --> 00:09:21.659 It was the easiest to demonstrate efficiency gains on. 204 00:09:22.620 --> 00:09:24.559 I have a thousand requests coming in every day. Can 205 00:09:24.580 --> 00:09:27.779 you boil them down to just a hundred that my 206 00:09:27.799 --> 00:09:30.309 people have to handle? Can 900 be handed off to 207 00:09:30.370 --> 00:09:30.789 an Asian? 208 00:09:31.350 --> 00:09:33.570 Beautiful. That was a quick efficiency gain. 209 00:09:34.029 --> 00:09:36.549 And I think that allowed HR a seat at the 210 00:09:36.590 --> 00:09:41.350 table to now negotiate a better AI strategy. And I 211 00:09:41.370 --> 00:09:44.580 think now we see a lot of them expanding into 212 00:09:44.620 --> 00:09:47.659 newer use cases, right? And I would argue that AI 213 00:09:47.700 --> 00:09:49.980 and HR is not just an efficiency gains play. 214 00:09:50.659 --> 00:09:52.580 It's actually a very strategic invention play. 215 00:09:52.600 --> 00:09:56.120 If you think about it, you can't quite put a 216 00:09:56.179 --> 00:09:59.929 cost to making the right internal mobility happen. 217 00:09:59.990 --> 00:10:00.730 At the right time. 218 00:10:01.090 --> 00:10:04.389 Or there's a special project running in some corner of 219 00:10:04.409 --> 00:10:06.879 the organization and I know you have a perfect fit 220 00:10:06.929 --> 00:10:09.659 right here. How do you connect the dots? So we've 221 00:10:09.700 --> 00:10:13.500 seen a lot of these experiments run and people drawing 222 00:10:13.539 --> 00:10:16.440 a lot of benefits from it. Like I said, the 223 00:10:16.519 --> 00:10:20.519 use cases range... you know, wide. We've seen people do 224 00:10:20.590 --> 00:10:29.070 financial modeling. We've seen people evaluate merging multiple teams. The 225 00:10:29.289 --> 00:10:32.669 market is volatile. There's a lot going on. So they're 226 00:10:32.690 --> 00:10:35.179 kind of evaluating what would it look like if I 227 00:10:35.259 --> 00:10:38.279 kind of, you know, normalized certain grades or what would 228 00:10:38.340 --> 00:10:41.340 happen if I merged these two teams and what would 229 00:10:41.379 --> 00:10:44.659 be the outcome to business if these two KPIs would 230 00:10:44.740 --> 00:10:45.460 sort of speak to. 231 00:10:45.399 --> 00:10:48.350 Each other, things like that. So very interesting use cases.

232 00:10:49.710 --> 00:10:52.289 But like I said, the first step, almost sort of 233 00:10:52.309 --> 00:10:55.480 a rite of passage is figure out your efficiency gains. 234 00:10:56.039 --> 00:10:58.159 It sort of gets you a seat at the table. 235 00:10:58.340 --> 00:11:01.100 You know you can save 20% cost. That kind of 236 00:11:01.120 --> 00:11:03.080 gets you in the game, right? Gives you a nice 237 00:11:03.240 --> 00:11:05.460 incentive to kind of whip up new use cases. 238 00:11:05.519 --> 00:11:08.230 Yeah, that makes sense. Thank you so much, Samar. This 239 00:11:08.289 --> 00:11:09.149 has been a pleasure. 240 00:11:09.169 --> 00:11:11.549 Thank you so much for joining us for this episode 241 00:11:11.629 --> 00:11:12.950 and we'll see you in the next one. 242 00:11:13.190 --> 00:11:13.570 Absolutely. 243 00:11:14.279 --> 00:11:16.980 So that was my conversation with Samrat Krishna, the head 244 00:11:17.320 --> 00:11:20.360 of marketing for US at Darwin Box. I want to 245 00:11:20.399 --> 00:11:22.639 spend a few minutes talking about what I took away 246 00:11:22.679 --> 00:11:24.899 from it and three things that stood out to me. 247 00:11:25.360 --> 00:11:27.700 So the first was that Samrat gave a clear account 248 00:11:27.759 --> 00:11:31.259 of what AI native means for them. Darwin Box first 249 00:11:31.519 --> 00:11:35.220 added AI features across its product, but it realized that 250 00:11:35.279 --> 00:11:36.919 to make AI truly useful. 251 00:11:37.110 --> 00:11:38.659 It needed to change the foundation. 252 00:11:39.120 --> 00:11:42.090 So Samrat said that the company had to return or 253 00:11:42.470 --> 00:11:45.750 unlearn 10 years of product design. Now, the key word 254 00:11:45.789 --> 00:11:49.620 here is product design. The decade of HR logic that 255 00:11:49.669 --> 00:11:53.100 the HCM product had still sits at the center. The 256 00:11:53.159 --> 00:11:57.000 rebuild adds layers of AI around that HCM core, that 257 00:11:57.200 --> 00:11:59.559 HR core, signals, context, and action. 258 00:12:01.019 --> 00:12:02.159 That is a sound approach. 259 00:12:02.440 --> 00:12:06.460 Many vendors now are calling themselves AI native HCM products. 260 00:12:06.539 --> 00:12:09.070 So if you're assessing a tech or looking to buy, 261 00:12:09.090 --> 00:12:11.929 I would encourage you to ask what has stayed and 262 00:12:12.009 --> 00:12:15.019 what has changed. The second thing I want to talk 263 00:12:15.080 --> 00:12:18.600 about is context. So Samrat said that smart AI agents 264 00:12:18.679 --> 00:12:21.320 did not know how a customer's system was set up, 265 00:12:21.720 --> 00:12:24.870 and he compared them to an eager management trainee who 266 00:12:24.889 --> 00:12:28.070 does not know where to go. So Darwin Box essentially 267 00:12:28.110 --> 00:12:31.850 built a context graph. The system describes itself to the agent, 268 00:12:31.950 --> 00:12:34.389 including how the customer is set up and what good 269 00:12:34.600 --> 00:12:38.879 HR practice looks like. The same graph lets its platform, Cortex, 270 00:12:39.159 --> 00:12:41.039 run inside Cloud, Copilot. 271 00:12:41.100 --> 00:12:43.360 And Slag. But here's my flag. 272 00:12:43.850 --> 00:12:47.820 Samrat said that an AI-native system needs the end-to-end context 273 00:12:47.909 --> 00:12:51.899 of the organization. At launch, the signals behind Cortex come 274 00:12:52.000 --> 00:12:55.860 from HR system data only. That works for workers like 275 00:12:55.960 --> 00:12:59.700 hourly workers or frontline employees whose working life the system records. 276 00:13:00.179 --> 00:13:02.980 But when it comes to knowledge workers, the context can 277 00:13:03.019 --> 00:13:06.509 be thin. An engineer's real insights can live in many 278 00:13:06.570 --> 00:13:10.980 systems like code repositories, calendars, and project management tools. So 279 00:13:11.019 --> 00:13:14.059 buyers should ask what signals and what systems are covered 280 00:13:14.159 --> 00:13:18.279 and how those signals are brought in. Data connectors can 281 00:13:18.360 --> 00:13:22.120 bring in more data, however, that doesn't automatically align the 282 00:13:22.139 --> 00:13:26.370 data or determine which signals feed into which decisions. That 283 00:13:26.450 --> 00:13:28.830 work needs its own lead and that is actually the 284 00:13:28.870 --> 00:13:31.629 critical piece of it. The third thing I want to 285 00:13:31.669 --> 00:13:35.809 talk about is how customers are using these AI native solutions. 286 00:13:36.169 --> 00:13:39.240 So Darwin Box launched Cortex in August And through an 287 00:13:39.340 --> 00:13:43.320 early access program, they saw increasing levels of customer interest. 288 00:13:43.419 --> 00:13:46.820 So Samraj said about 50% of its customers wanted it 289 00:13:46.860 --> 00:13:50.659 within three weeks and customers began with service delivery. In 290 00:13:50.700 --> 00:13:54.840 his example, a thousand daily requests dropped to about 100 291 00:13:54.840 --> 00:13:58.620 for people to handle. He said that gain earns HR 292 00:13:58.700 --> 00:14:02.100 a strategic seat to help shape AI strategy for the 293 00:14:02.179 --> 00:14:04.639 organization because now a lot of its work that it 294 00:14:04.679 --> 00:14:08.360 was spending time doing is now being automated. But that 295 00:14:08.460 --> 00:14:10.919 work right now sits at the simple end of things. 296 00:14:11.440 --> 00:14:15.240 In our 2026 People Analytics technology study, we found that 297 00:14:15.240 --> 00:14:19.980 70% of vendors have built co-pilots and 49%, so almost 50%, 298 00:14:19.980 --> 00:14:23.659 have built agents. But most agent work handles simple to 299 00:14:23.720 --> 00:14:28.970 moderate tasks. As companies increasingly scrutinize their AI spend and investments, 300 00:14:29.450 --> 00:14:32.090 I'd be more interested to see what the adoption of 301 00:14:32.110 --> 00:14:35.700 AI agents looks like for those that handle more complex 302 00:14:35.759 --> 00:14:39.659 tasks and automate end-to-end workflows. The other thing that I 303 00:14:39.720 --> 00:14:43.259 want to flag is that Cortex builds playbooks for managers, 304 00:14:43.480 --> 00:14:46.279 and in the example that Samrat gave, a top performer 305 00:14:46.340 --> 00:14:49.080 that was likely to leave, probably because of factors such 306 00:14:49.159 --> 00:14:52.539 as below-market pay, a manager change, and a heavy workload, 307 00:14:52.860 --> 00:14:55.840 is then connected to a suggestion steps like mandatory leave. 308 00:14:56.730 --> 00:14:59.889 The link from signals to a predictive problem is Darwin 309 00:15:00.049 --> 00:15:04.909 Box's own hypothesis, which is built on past customer data. However, 310 00:15:05.190 --> 00:15:08.710 a manager who accepts a playbook must also be able 311 00:15:08.730 --> 00:15:11.610 to explain that to the leadership. So if a manager 312 00:15:11.870 --> 00:15:14.000 is being suggested to take an action and does take 313 00:15:14.029 --> 00:15:17.600 that action, he or she should be able to explain. 314 00:15:17.620 --> 00:15:21.379 Why they did that. Darwin Box logs what the agent did. 315 00:15:21.919 --> 00:15:23.360 What I'd also want to look. 316 00:15:23.259 --> 00:15:26.059 At as a user is whether the record explains why 317 00:15:26.120 --> 00:15:30.440 that action was recommended. So as users, ask which signals 318 00:15:30.559 --> 00:15:34.350 drove the advice and how confident the system was. So 319 00:15:34.370 --> 00:15:36.590 that's what stood out to me from my conversation with 320 00:15:36.620 --> 00:15:39.409 Darwin Box. If the conversation was interesting to you and 321 00:15:39.429 --> 00:15:41.690 you'd like to learn more about our research and what 322 00:15:41.730 --> 00:15:44.909 we are looking at in HR technology, follow our research, 323 00:15:44.970 --> 00:15:47.450 join our membership, check it out. And if you found 324 00:15:47.490 --> 00:15:50.029 this useful, share it with a friend, hit like and 325 00:15:50.070 --> 00:15:51.429 share your thoughts wherever you're listening. 326 00:15:51.779 --> 00:15:52.080 Thank you.