1
00:00:04,080 --> 00:00:06,800
Good afternoon, everyone. We're going to give people

2
00:00:06,800 --> 00:00:27,040
a minute or two to log on, and then we're going to get started.

3
00:00:51,920 --> 00:00:58,320
Hi everyone. Thank you for joining today's webinar, AI and HR: Top five risks and opportunities. We're

4
00:00:58,320 --> 00:01:01,600
recording today's session and we'll share it with you following the webinar. If you have

5
00:01:01,600 --> 00:01:06,480
any questions, please use the Q&A function on your Zoom panel. Now, I'd like to hand it over to our

6
00:01:06,480 --> 00:01:13,040
moderator for today's session, HR Acuity CEO, Deb Muller. Thanks, Lexi. Hi everyone. I see everybody's

7
00:01:13,040 --> 00:01:16,640
sort of telling us where they're calling in from or dialing in from. I don't know what

8
00:01:16,640 --> 00:01:22,400
we do. We zoom in from, I guess, is the... the language we should be using now. But it's

9
00:01:22,400 --> 00:01:27,440
like everybody's from all over, and I know this is a great topic of interest. A lot of people

10
00:01:27,440 --> 00:01:32,320
are joining, but we're going to get started because we have a lot of content to get through as well.

11
00:01:32,320 --> 00:01:36,720
So, I'm going to do a little bit of an overview, some welcomes and introductions, and then hand

12
00:01:36,720 --> 00:01:41,600
it over to our two facilitators today that we're really excited we have with us. I'll get to that

13
00:01:41,600 --> 00:01:49,760
in a second. So, let's see, why don't you just run to the next slide. Okay, so maybe you

14
00:01:49,760 --> 00:01:54,400
don't know about HR Acuity, but we are a technology company. We are the leading provider of employee

15
00:01:54,400 --> 00:01:59,600
relations and investigations case management, and we like to say that we help organizations manage

16
00:01:59,600 --> 00:02:04,720
things from intake all the way through aftercare with really deep analytics. So instead of just

17
00:02:04,720 --> 00:02:10,400
tracking issues, you can help use data to inform what happens next in your organization. So if you

18
00:02:10,400 --> 00:02:13,920
want to learn more about our platform—we're not going to talk about it today—but if you

19
00:02:13,920 --> 00:02:17,920
want to learn more about what we're doing and the exciting things that we're doing with AI, which is

20
00:02:17,920 --> 00:02:22,800
relevant to this conversation, please reach out to us, and we'd love to take you on a curiosity

21
00:02:22,800 --> 00:02:27,360
tour. And we welcome all of our customers. I see some of them logging in here today as well.

22
00:02:27,360 --> 00:02:35,680
So thanks for joining us. Next slide. Okay, a little bit more about HR Acuity. Yes, we are a technology company,

23
00:02:35,680 --> 00:02:41,440
but we're really proud of the employee relations community that we've built over the past decade-plus

24
00:02:41,440 --> 00:02:46,880
together. The first thing is our benchmark study, and I'm excited to say that we actually

25
00:02:46,880 --> 00:02:53,040
just launched today our ninth annual employee relations benchmark study. Lexi is going to put

26
00:02:53,040 --> 00:02:57,680
all the links in the chat, but if you have not gone to our website and downloaded it, or actually this

27
00:02:57,680 --> 00:03:03,760
year you can read it all right online—go, I mean, don't go now—after the hour is over, you'll go and

28
00:03:03,760 --> 00:03:09,920
look through it, download it. It is chock-full of great information, some new insights, some things

29
00:03:09,920 --> 00:03:14,720
about AI I'm going to go through in just a second. So please do that. And if you're interested in

30
00:03:14,720 --> 00:03:18,720
participating next year, our 10th annual, we want to make it bigger than ever. Although this year

31
00:03:18,720 --> 00:03:23,840
it does cover about 9 million employees. Let us know if your organization has more than a thousand

32
00:03:23,840 --> 00:03:30,400
employees. We also have our empower network. It's our employee relations online community. It is free.

33
00:03:30,400 --> 00:03:35,440
There's an app, you can do it online. If you're in employee relations, if you do anything related to

34
00:03:35,440 --> 00:03:40,400
employee relations, you should join. Have your team join. It's a great community. I believe there are

35
00:03:40,400 --> 00:03:49,040
close to 7,500 members right now, all levels from all over the world on live feeds, talking about

36
00:03:49,040 --> 00:03:53,200
what's going on, what questions they have, collaborating. There are small groups to talk about

37
00:03:53,200 --> 00:03:59,040
specific topics like labor, to talk about global issues. It's really... you have to be there. So make

38
00:03:59,040 --> 00:04:05,120
sure you go there. We also have our ERQ, which is our Employee Relations Quotient. It is our ER

39
00:04:05,120 --> 00:04:09,840
maturity model. Go on our website, take it, see where you can go, get some insights as to how you can

40
00:04:09,840 --> 00:04:15,440
move along the maturity for your organization in ER. And finally, if you are an employee relations

41
00:04:15,440 --> 00:04:20,480
leader at an enterprise organization, we'd love for you to join us at our roundtable. I think it's

42
00:04:20,480 --> 00:04:27,200
our ninth or 10th year doing that as well. It's an annual in-person event of global ER leaders. We get

43
00:04:27,200 --> 00:04:32,640
together, we collaborate, we really start defining the future of employee relations. So if you want to

44
00:04:32,640 --> 00:04:38,080
be part of that, it's a very unique opportunity that is going to be in October. Our website

45
00:04:38,080 --> 00:04:42,960
gives you all the information. It's in Boston this year. We'd love to have you join us. All right, next

46
00:04:42,960 --> 00:04:49,120
slide. Important dates. I talked about the benchmark. It is out now. Other things happening this month:

47
00:04:49,120 --> 00:04:54,640
we have another webinar coming out about the benchmark. So I'm going to go into more detail with

48
00:04:54,640 --> 00:04:58,800
our partner, our research partner, Isos. We're going to talk about the results, so mark your

49
00:04:58,800 --> 00:05:04,880
calendar for that on June 24th. June 25th, we're going to have an Empower discussion. So different

50
00:05:04,880 --> 00:05:12,000
from the webinars, if you're in Empower, we have ER professionals that lead these discussions.

51
00:05:12,000 --> 00:05:16,160
So, you know, now we're going to talk about, we're going to hear about AI and ER, what to look out for,

52
00:05:16,160 --> 00:05:20,480
the risks. This is when we talk about it. What are we doing? How are we experimenting? What's really

53
00:05:20,480 --> 00:05:26,720
happening in our companies and helping each other through this new phase of how we do work? And

54
00:05:26,720 --> 00:05:30,800
then again, I mentioned the roundtable. We hope that you can join us, and

55
00:05:30,800 --> 00:05:39,520
that's October 28th and 29th. Again, we will make sure you have all these links. Next slide. Okay, my

56
00:05:39,520 --> 00:05:44,560
favorite slide. For those of you that have attended a webinar, as a thank you to our presenters, we like

57
00:05:44,560 --> 00:05:49,280
to make a donation for their time and for all of you for joining us. And we ask them to,

58
00:05:49,280 --> 00:05:52,720
you know, choose if there's an organization that's meaningful for them or one that's in line with

59
00:05:52,720 --> 00:05:57,440
our mission and vision. And so, we're so excited today to be making a donation to the Just in

60
00:05:57,440 --> 00:06:04,160
Time for Youth, foster youth nonprofits in San Diego. I mean, I can't think of an organization that

61
00:06:04,160 --> 00:06:08,800
is more important to help children that are in the system. And this looks like they do it when they

62
00:06:08,800 --> 00:06:14,240
get out of the system, transitioning out. And I mean, a wonderful organization, and we're so proud to be

63
00:06:14,240 --> 00:06:19,360
able to make a donation. So thank you for coming and thank you to Eli and Oliver for sharing your

64
00:06:19,360 --> 00:06:25,520
expertise so that we can go ahead and make this donation. Okay, so I'm going to do a little starting,

65
00:06:25,520 --> 00:06:29,280
and this is a little sneak preview of what's in that benchmark study because, of course, we

66
00:06:29,280 --> 00:06:37,200
asked about AI in there. And maybe not a big shocker, it's the buzz, you know, when it comes to

67
00:06:37,200 --> 00:06:43,040
employee relations, it is still early days. Nearly half of organizations that we surveyed in this

68
00:06:43,040 --> 00:06:48,560
say that their use of AI is nonexistent, and only about a third are experimenting or piloting.

69
00:06:48,560 --> 00:06:52,560
And it sort of makes sense, and that's why we're having this conversation today. ER

70
00:06:52,560 --> 00:06:58,240
deals with incredibly sensitive and complex issues, so legal, ethical, and compliance concerns are real,

71
00:06:58,240 --> 00:07:01,760
and I think it's causing organizations to move more cautiously. We're going to talk about that

72
00:07:01,760 --> 00:07:06,720
today because just because we have to be cautious doesn't mean we can't use AI. It doesn't mean there

73
00:07:06,720 --> 00:07:11,760
aren't applications. We're excited about it at HR Acuity. There's real opportunity, but we have to be

74
00:07:11,760 --> 00:07:16,240
intentional and responsible. So, you know, I think this shows you don't feel bad if you're sort of

75
00:07:16,240 --> 00:07:22,720
behind the eight ball. You're going to get there. And we're going to

76
00:07:22,720 --> 00:07:27,440
we're going to help you do that in the way that actually supports what we're doing.

77
00:07:27,440 --> 00:07:31,760
It doesn't eliminate what we're doing, but it really supports and makes us better as

78
00:07:31,760 --> 00:07:36,480
employee relations professionals and investigators. Interestingly, we also asked—if you can just move

79
00:07:36,480 --> 00:07:42,000
to the next slide—we asked last year about interest in AI. So we started talking about it.

80
00:07:42,000 --> 00:07:47,360
You can see here the top interests were things like uncovering trends, analyzing case data, improving

81
00:07:47,360 --> 00:07:53,440
workflows. Over 75% of respondents said they were interested in these areas. So fast

82
00:07:53,440 --> 00:07:59,920
forward to this year, not a lot has changed, right? Only 28% are using AI for writing assistance, with much

83
00:07:59,920 --> 00:08:05,600
smaller numbers for things like policy recommendations, case transcription, and analysis.

84
00:08:05,600 --> 00:08:11,280
So, I think that's part of the conversation today. People want to use it, but we need to make sure if

85
00:08:11,280 --> 00:08:17,440
we unlock the benefits, we're doing so responsibly. There are guardrails. They enhance, not replace,

86
00:08:17,440 --> 00:08:24,240
the human work of ER. And this is just so relevant to today's conversation. So this and

87
00:08:24,240 --> 00:08:30,640
more in the benchmark study. So download that after the session. But for now, I do want to introduce our

88
00:08:30,640 --> 00:08:35,040
presenters and really thank them for being here. So we have two great guest speakers that are going to

89
00:08:35,040 --> 00:08:41,360
share with us their wisdom on this. We have Eli Makus. He is the managing partner of Van Dermyden Makus

90
00:08:41,360 --> 00:08:46,480
Law Corporation. It's a preeminent investigations firm with attorneys across California and Arizona.

91
00:08:46,480 --> 00:08:51,600
They are devoted to promoting fair workplaces and safe campuses through industry-leading neutral

92
00:08:51,600 --> 00:08:57,680
fact-finding services. Drawing upon his extensive employment law background, Eli conducts

93
00:08:57,680 --> 00:09:02,480
complex and sensitive investigations involving a variety of workplace complaints for public

94
00:09:02,480 --> 00:09:07,360
and private employers throughout California. And if you can handle California, you can handle anything

95
00:09:07,360 --> 00:09:12,880
else. It's a country in and of itself. Eli regularly trains on employment law topics, including training

96
00:09:12,880 --> 00:09:17,920
internal and external investigators on how to conduct impartial workplace investigations, and he

97
00:09:17,920 --> 00:09:23,200
is the immediate past president of the Association of Workplace Investigators Board of Directors and

98
00:09:23,200 --> 00:09:30,160
regularly serves on the faculty of AWI's ANAB-accredited training institute. So thank you, Eli,

99
00:09:30,160 --> 00:09:36,640
for joining us. And Oliver McKinstry is the vice president, associate general counsel of labor,

100
00:09:36,640 --> 00:09:42,560
employment, and investigations at DaVita, a Fortune 200 kidney care company. He leads DaVita's labor

101
00:09:42,560 --> 00:09:47,680
and employment law and HR compliance teams. And in his role, Oliver oversees all of DaVita's labor

102
00:09:47,680 --> 00:09:52,400
and employment advice and litigation, including wage and hour and union-related matters. He also

103
00:09:52,400 --> 00:09:56,960
leads a team of lawyers and HR professionals focused on workplace investigations, employee

104
00:09:56,960 --> 00:10:02,080
relations, HR policies, personnel files, and license and certification requirements.

105
00:10:02,080 --> 00:10:08,000
Since joining DaVita in 2016, Oliver has helped streamline the investigation process, implement

106
00:10:08,000 --> 00:10:13,120
a process for addressing threats of workplace violence, standardize investigation reports, and

107
00:10:13,120 --> 00:10:17,680
implement artificial intelligence to help with the intake process and draft corrective actions.

108
00:10:17,680 --> 00:10:21,840
Oliver is currently the vice president of the Association of Workplace Investigators

109
00:10:21,840 --> 00:10:27,520
and will become president in November 2025. Congratulations. So welcome, both of you, and

110
00:10:27,520 --> 00:10:32,800
looking forward to an awesome conversation. We encourage people to ask questions

111
00:10:32,800 --> 00:10:37,120
in the chat—we're going to monitor that—or if you prefer to do it anonymously, there is a Q&A

112
00:10:37,120 --> 00:10:42,160
option as well. And we're going to try to fit in your comments and questions in between where it

113
00:10:42,160 --> 00:10:46,800
makes sense. But I know we have a lot of content to cover. So thanks for joining us, Eli and Oliver.

114
00:10:47,600 --> 00:10:52,000
Great. Thank you so much, Deb. We're so excited to be here. This is a topic that Oliver and I have both

115
00:10:52,000 --> 00:10:57,840
talked a lot about and presented on before, and one we think is just particularly pertinent

116
00:10:57,840 --> 00:11:02,640
and relevant to the experience of being a human resources professional right now. I wanted to

117
00:11:02,640 --> 00:11:08,240
mention one more thing about our firm. So our firm, which is based in California and Arizona, we

118
00:11:08,240 --> 00:11:14,000
focus on conducting workplace investigations when employers hire us to do that, but we also recognize

119
00:11:14,000 --> 00:11:20,480
that most investigations should be conducted by well-trained internal HR professionals. And

120
00:11:20,480 --> 00:11:26,080
so we have a separate training company called VM Mastered, which is dedicated to training, creating

121
00:11:26,080 --> 00:11:31,040
customized training programs for internal HR professionals who are conducting workplace

122
00:11:31,040 --> 00:11:36,800
investigations, and in those conversations, AI continues to be a big topic. And so we're

123
00:11:36,800 --> 00:11:41,840
excited to address it today. With that, I'll turn it over to Oliver, and Oliver is going to kick us off.

124
00:11:42,800 --> 00:11:49,760
Yeah, great. Thanks, Deb, for the intros and thanks, Eli, for passing it over. I'm very happy to be here.

125
00:11:49,760 --> 00:11:55,600
Uh, I am, I call myself a recovering lawyer, although I am a practicing lawyer and helped

126
00:11:55,600 --> 00:12:01,040
DaVita with our legal practice. I think my heart is in HR, and so I identify first as

127
00:12:01,040 --> 00:12:08,080
an HR professional, an investigator, and an employee relations professional. So hopefully some of

128
00:12:08,080 --> 00:12:13,520
what I'm going to talk through, you'll feel a little bit of kinship with me and what

129
00:12:13,520 --> 00:12:19,600
we're going through at this time with AI and what to do in the HR profession. So let's

130
00:12:19,600 --> 00:12:28,560
move to the next slide. Eli and I today wanted to really dive into AI from the perspective of

131
00:12:28,560 --> 00:12:35,840
general HR and look at it with an eye toward where are our greatest opportunities and maybe

132
00:12:35,840 --> 00:12:40,720
where our greatest risks. And so what we're going to talk about is the top five risks and top five

133
00:12:40,720 --> 00:12:48,160
opportunities for AI in HR. I think Deb mentioned a little bit of research on AI in the employee

134
00:12:48,160 --> 00:12:54,640
relations space. Let's take a little bit of a broader look at some research just in the HR

135
00:12:54,640 --> 00:13:02,400
space in general. Um, Eli and I looked into some of this, and it turns out a large majority of people

136
00:13:02,400 --> 00:13:11,040
in the workplace are already using AI. Over half of employees, and probably over half of you

137
00:13:11,040 --> 00:13:18,240
in attendance, are already using AI somehow in your daily work. And I think it's really interesting

138
00:13:18,240 --> 00:13:25,760
to think about the ways different generations use AI in the workplace. The research indicates that

139
00:13:25,760 --> 00:13:33,120
older employees are really using AI and specifically ChatGPT more as a search engine,

140
00:13:33,120 --> 00:13:40,640
something like Google, where they can type in some kind of prompt and get an answer, whereas

141
00:13:40,640 --> 00:13:48,640
twenty-somethings and those kind of just older than twenty-somethings tend to use it more as a life advisor. And

142
00:13:48,640 --> 00:13:55,440
I identify with that age group, and I will admit that I often use AI to help me plan my

143
00:13:55,440 --> 00:14:01,680
travel, where to go, what to see, what to eat. So I think I fall into that category. And

144
00:14:01,680 --> 00:14:07,040
it turns out college students are actually using it basically as an operating system. It's doing

145
00:14:07,040 --> 00:14:13,440
many different functions for them. They really have adopted it very thoroughly. So interesting to

146
00:14:13,440 --> 00:14:20,480
start from the place of many people are already using AI in the workplace. I think it's important

147
00:14:20,480 --> 00:14:26,400
to think about why they're doing that. And the first main reason that comes from the

148
00:14:26,400 --> 00:14:33,440
research, most people are saying, is efficiency, trying to do things in a more streamlined way

149
00:14:33,440 --> 00:14:38,480
and save time by doing that. Certainly one of the reasons I have it make my travel plan,

150
00:14:38,480 --> 00:14:45,600
so that makes sense to me. Uh, another top reason is just simply access to information, trying

151
00:14:45,600 --> 00:14:51,120
to find information that we think is out there, but maybe we don't know where to get it, having

152
00:14:51,120 --> 00:14:57,680
AI go and do that for us. The next one is just innovating, coming up with some kind of

153
00:14:57,680 --> 00:15:04,080
new process or approach to things. That is another function that AI can do for us. And

154
00:15:04,080 --> 00:15:09,760
then the last one is improving our work quality. And people are already doing that,

155
00:15:09,760 --> 00:15:15,920
using AI to try and improve their work quality. So probably not a lot of surprises

156
00:15:15,920 --> 00:15:22,080
here. And I think many of the opportunities we're going to discuss will certainly hit on

157
00:15:22,080 --> 00:15:29,680
these use cases for AI in HR, and then we'll talk about some of the risks with that, too.

158
00:15:31,600 --> 00:15:38,800
So a couple of the maybe challenging components of using AI in the workplace, of course, data

159
00:15:38,800 --> 00:15:46,720
sensitivity is one of them. And the research that we reviewed shows that a lot of people are using

160
00:15:46,720 --> 00:15:53,760
AI in a way that they actually are uploading sensitive company or customer information into

161
00:15:53,760 --> 00:16:01,600
public generative AI tools. And that is a big challenge with these tools, is once you upload your

162
00:16:01,600 --> 00:16:07,040
information, you may lose control over it, and you don't know how it's being used. And unfortunately,

163
00:16:07,040 --> 00:16:11,440
it looks like a lot of the workforce is already out there doing that. I hope that's

164
00:16:11,440 --> 00:16:18,480
not anyone on the call today, but that's going on. So that's something we should be aware of. Also,

165
00:16:18,480 --> 00:16:26,000
um, a large majority of people have relied on AI output without actually critically evaluating the

166
00:16:26,000 --> 00:16:32,960
results. And I think Deb mentioned this previously. It is essential to maintain the human judgment

167
00:16:32,960 --> 00:16:38,640
component even when we're using AI. But as you can see in this data, that is already a challenge that

168
00:16:38,640 --> 00:16:44,560
people are not using their critical thinking when using these tools. And then last here on this slide,

169
00:16:44,560 --> 00:16:51,040
um, people are making mistakes when they rely on AI tools. And I think in the legal profession,

170
00:16:51,040 --> 00:16:57,920
we've seen this in some filings in court where lawyers are using AI to come up with some case

171
00:16:57,920 --> 00:17:04,640
law citations, and the AI is hallucinating cases that actually don't exist. Uh, those have ended

172
00:17:04,640 --> 00:17:10,080
up in briefs that have been filed with the court, and it turns out judges actually read the briefs

173
00:17:10,080 --> 00:17:16,240
that lawyers file, and they have figured out that those citations don't exist. And lawyers

174
00:17:16,240 --> 00:17:21,440
have been sanctioned. And I think it's just a good example of a profession where we think

175
00:17:21,440 --> 00:17:27,360
people are very careful with the things that they file, but when relying on AI, you really have to

176
00:17:27,360 --> 00:17:35,280
double-check the results that it's giving you before using them. Thanks, Oliver. And before

177
00:17:35,280 --> 00:17:40,800
we really jump deep into the opportunities and risks that we see around artificial intelligence,

178
00:17:40,800 --> 00:17:46,000
we think it's worth just level setting on what it is that makes this such a unique and

179
00:17:46,000 --> 00:17:53,280
tantalizing and also challenging technology to adopt, to understand, and why the nature of

180
00:17:53,280 --> 00:17:58,960
the technology itself makes it particularly concerning in an HR context, given the type

181
00:17:58,960 --> 00:18:04,000
of work that we do. So I want you to think about this distinction between deterministic versus

182
00:18:04,000 --> 00:18:09,840
probabilistic technology. And I'll break it down with this example. If you think about using a

183
00:18:09,840 --> 00:18:18,080
calculator, if you enter 6 * 8 into any calculator anywhere at any time, you're always going to get

184
00:18:18,080 --> 00:18:24,640
48 as the answer. You're never going to get an answer that says, "Well, 51% of the time it's 48,

185
00:18:24,640 --> 00:18:32,560
but 32% of the time it may be 23, and you know, one or 2% of the time it could be five." It's always

186
00:18:32,560 --> 00:18:38,000
going to give you the same output. And that's what we're used to when we're talking about existing

187
00:18:38,000 --> 00:18:45,760
technology. When we use a computer, we expect that if we type "deterministic" into Word, we are always

188
00:18:45,760 --> 00:18:52,160
going to get deterministic output from our keyboard. So the technology that we're used to

189
00:18:52,160 --> 00:18:57,200
is based on predetermined conclusions. We know we're always going to get the same answer and

190
00:18:57,200 --> 00:19:02,320
the same output based on the context or the information we provide to it. But artificial

191
00:19:02,320 --> 00:19:07,440
intelligence is actually very different. It's probabilistic technology. The interesting thing

192
00:19:07,440 --> 00:19:13,280
about this is it's actually... this is actually how people think. People are probabilistic technology.

193
00:19:13,280 --> 00:19:19,120
If you give us a bunch of information and you ask us for an answer, we're going to weigh that

194
00:19:19,120 --> 00:19:23,840
information, and based on our experience and our knowledge and everything we've learned over the

195
00:19:23,840 --> 00:19:29,200
course of our lifetime, we're going to give you our best estimate or best guess on what the right

196
00:19:29,200 --> 00:19:35,440
answer is to your question based on our experience and knowledge. But it's just a probabilistic

197
00:19:35,440 --> 00:19:42,080
outcome. It's our best guess based on what we know. And that's how AI tech, artificial, generative

198
00:19:42,080 --> 00:19:47,440
artificial intelligence works as well. The decisions or the outputs that you receive from

199
00:19:47,440 --> 00:19:55,600
an AI tool are based on the tool's best estimate of the answer as it goes out and looks at all of

200
00:19:55,600 --> 00:20:02,480
the data that it's been trained on. So this means that output can be pretty inconsistent under the

201
00:20:02,480 --> 00:20:08,640
same conditions. And my guess is many of you have experienced this where you put a question into an

202
00:20:08,640 --> 00:20:14,400
AI tool at 3:00 and you get one answer, and then at 3:05, you put the same question in and you get

203
00:20:14,400 --> 00:20:19,440
a slightly different answer. And maybe they're both good answers, but they're different. And

204
00:20:19,440 --> 00:20:24,000
I think this is one of the things that's most confounding and both creates opportunities, but

205
00:20:24,000 --> 00:20:30,480
also obviously flags significant risks around using artificial intelligence. So I want you to

206
00:20:30,480 --> 00:20:34,960
think about that distinction between deterministic and probabilistic technology as we talk through

207
00:20:34,960 --> 00:20:39,840
some of these opportunities and risks. I also want to highlight another challenge with AI is

208
00:20:39,840 --> 00:20:46,000
just the variations in how it's used. Oliver already touched on how generational, different

209
00:20:46,000 --> 00:20:51,440
generational differences impact different ways of using or thinking about artificial

210
00:20:51,440 --> 00:20:56,800
intelligence, but there are other factors, too. There are socioeconomic differences. This, you

211
00:20:56,800 --> 00:21:03,680
know, much of the writing on AI talks about the concern around whether different socioeconomic

212
00:21:03,680 --> 00:21:09,440
uh, communities will have the same access to to AI and whether they'll have the same education and

213
00:21:09,440 --> 00:21:14,800
training and understanding of the risks of it. So that's a significant factor we have to consider as

214
00:21:14,800 --> 00:21:20,320
well. Then there's just organizational differences. My guess is that there are some folks on this

215
00:21:20,320 --> 00:21:25,760
uh, webinar who work in an organization where the organization has said, "Absolutely not. Under

216
00:21:25,760 --> 00:21:31,120
no circumstances may you use AI in the execution of your work duties." I know that

217
00:21:31,120 --> 00:21:35,840
there are large law firms who have taken that approach, even though AI has a lot of potential

218
00:21:35,840 --> 00:21:40,080
in the legal field. So there's just going to be different experiences for organizations

219
00:21:40,080 --> 00:21:44,960
and how to navigate AI. And then there's just individual comfort. Some people are going to

220
00:21:44,960 --> 00:21:50,400
be more comfortable using the technology, others less so. And bringing a whole organization up to

221
00:21:50,400 --> 00:21:56,880
a common level of experience and utilizing AI in a function like human resources is going to

222
00:21:56,880 --> 00:22:02,800
be pretty challenging. The reality is, too, there's just not a... there's a lack of clear direction on

223
00:22:02,800 --> 00:22:07,600
best practices for using AI in human resources. And I think that's what's really amazing about

224
00:22:07,600 --> 00:22:13,200
the Empower community and that discussion that's coming up later, I think in in June,

225
00:22:13,200 --> 00:22:18,720
I think it was, is just the opportunity to sit and discuss with your colleagues, "How are you actually

226
00:22:18,720 --> 00:22:23,360
using this in a safe place where you can talk about the challenges and the pitfalls in doing

227
00:22:23,360 --> 00:22:27,520
so?" So continue that dialogue because that's ultimately going to be how we come up with

228
00:22:27,520 --> 00:22:37,600
what the best practices are. So with that, why don't we jump into our top five opportunities?

229
00:22:37,600 --> 00:22:45,840
Yeah, great. Hopefully that is uh thought-provoking already. Uh, starting out with one of the most

230
00:22:45,840 --> 00:22:54,000
significant opportunities for AI in my mind is really using AI as kind of an HR generalist type

231
00:22:54,000 --> 00:23:00,160
function. If you're anything like me, you probably spend a lot of your day or your

232
00:23:00,160 --> 00:23:06,160
time really pointing people in the right direction. That may be just telling someone where to find a

233
00:23:06,160 --> 00:23:12,560
specific policy, maybe it's a travel policy, maybe it's connecting someone to your internal leave

234
00:23:12,560 --> 00:23:18,560
of absence team or the external provider that you have for that. Um, maybe it's just contact

235
00:23:18,560 --> 00:23:24,240
information that you're providing. But I think a lot of HR generalists, a lot of HR people spend

236
00:23:24,240 --> 00:23:31,920
time really just connecting employees with the right resource. But all of that information

237
00:23:31,920 --> 00:23:41,200
usually exists in policy documents or practice one-pagers that AI tools can really easily digest

238
00:23:41,200 --> 00:23:48,720
and learn. And then a valuable way to use AI is to have your employees go to this type of

239
00:23:48,720 --> 00:23:56,080
AI agent and ask those questions to the AI agent, whether that's a bot or some other kind of tool,

240
00:23:56,080 --> 00:24:02,880
and it can just simply guide them to those resources based on the policy documents that

241
00:24:02,880 --> 00:24:09,440
it has access to or the practices that you've trained it on. And in doing so, think of how much

242
00:24:09,440 --> 00:24:16,160
time that might take off of your plate or your HR generalist's plate. So I think of this as a really

243
00:24:16,160 --> 00:24:21,280
great opportunity to create those efficiencies that we talked about that people are looking for

244
00:24:21,280 --> 00:24:28,880
with AI, by really helping that HR generalist function. Often it's the more junior folks in HR

245
00:24:28,880 --> 00:24:34,720
who are doing those things, but can really create efficiencies for them and do some

246
00:24:34,720 --> 00:24:41,200
of those simpler functions that they spend time on. Eli, anything you'd add there? Well, I think, I mean,

247
00:24:41,200 --> 00:24:45,040
as we talk through these opportunities, there's always another side to that, too, which is there

248
00:24:45,040 --> 00:24:50,720
probably are some challenges or risks around delegating this responsibility to an AI assist...

249
00:24:50,720 --> 00:24:57,120
an HR AI assistant as well, right? So I can imagine concerns around accuracy might be something that

250
00:24:57,120 --> 00:25:04,000
one would be concerned about. Is that fair? Oh, for sure. You know, if you're anything like

251
00:25:04,000 --> 00:25:10,000
companies I've worked for, sometimes you have multiple policy documents or you even have

252
00:25:10,000 --> 00:25:16,720
um, some conflicting practices depending on an organization, or maybe it's just you've

253
00:25:16,720 --> 00:25:22,560
had several iterations of the same policy. So it's possible that the AI might not get

254
00:25:22,560 --> 00:25:28,160
the accurate answer. Maybe it goes to the wrong document or can't compile all of the documents

255
00:25:28,160 --> 00:25:33,040
to come up with something. So I think that's a realistic challenge with this approach. You have

256
00:25:33,040 --> 00:25:40,160
to be careful about the inputs, so you get the right output. And I want to address Paula's

257
00:25:40,160 --> 00:25:44,800
question in the chat here, and I want to just sort of acknowledge one of the more

258
00:25:44,800 --> 00:25:49,120
challenging aspects of a webinar like this, and I'm afraid that we're going to do the same thing that

259
00:25:49,120 --> 00:25:53,520
you've probably heard in other webinars. We're going to talk a lot about what AI can do, what

260
00:25:53,520 --> 00:25:58,240
it, what some of the risks are, but we're not going to give you any guidance on what tools you should

261
00:25:58,240 --> 00:26:04,400
actually explore or adopt or identify even what specific tools are worth evaluating.

262
00:26:04,400 --> 00:26:09,120
And that's largely because there are so many different tools and they're at such an early stage

263
00:26:09,120 --> 00:26:12,960
that it's virtually impossible to make that kind of a recommendation. Well, I think what our goal

264
00:26:12,960 --> 00:26:18,720
is that you take away from this discussion what the potential is so that as you are going out in

265
00:26:18,720 --> 00:26:23,920
a more systematic process in your organization and beginning to evaluate tools, you have some

266
00:26:23,920 --> 00:26:31,360
framework for questions and how to evaluate what might be an appropriate tool for your organization.

267
00:26:31,360 --> 00:26:37,760
And so Paula, to your question, you know, how does this work from a—and I'm no technologist—but from

268
00:26:37,760 --> 00:26:45,440
a technical perspective, basically current AI tools are already trained, their large language models

269
00:26:45,440 --> 00:26:50,560
are already trained on large volumes of data. And so a lot of what the new AI tools are doing is

270
00:26:50,560 --> 00:26:56,320
they have an LLM as a foundation for the tool, and then they are developing specific agents which

271
00:26:56,320 --> 00:27:02,800
have a specific purpose. An agent is a technical term that describes a basically a plug-in to the

272
00:27:02,800 --> 00:27:08,320
AI tool that goes out and, for example, in this scenario, it would go out and it would look at what

273
00:27:08,320 --> 00:27:14,800
your policies are, which would be uploaded into the tool, and it would evaluate based on its estim... you

274
00:27:14,800 --> 00:27:18,960
know, based on its best guess review considering all the data it's been trained on, it would

275
00:27:18,960 --> 00:27:25,440
evaluate looking at your policies what the best answer is to the prompt that was given. So, this

276
00:27:25,440 --> 00:27:31,760
is a really good question and one that as you're evaluating products that are promising that

277
00:27:31,760 --> 00:27:36,640
they can do all sorts of amazing things in the HR space, you really want to make sure that the vendor

278
00:27:36,640 --> 00:27:41,200
understands the answer to questions like this and can explain to you how it's going to work and how

279
00:27:41,200 --> 00:27:48,080
you're going to avoid errors, for example, or inaccuracies or hallucinations. So with that, why

280
00:27:48,080 --> 00:27:53,600
don't we move on to another opportunity? And with all of these, there's a double-edged

281
00:27:53,600 --> 00:27:59,520
sword, right? But I think one area that AI certainly has potential for is summarizing and triaging

282
00:27:59,520 --> 00:28:05,360
conversations that occur in the HR space. In our experience, in the tools that we've experimented

283
00:28:05,360 --> 00:28:11,280
with—and I should just acknowledge one thing—in our law firm, we've decided to develop our own

284
00:28:11,280 --> 00:28:16,560
tool. So we're working on a customized tool that is unique to our firm, to workplace investigations, the

285
00:28:16,560 --> 00:28:21,680
way we do them. And so a lot of our experimentation has come through use of that tool. And we found, for

286
00:28:21,680 --> 00:28:29,200
example, that AI-supported transcription is really quite accurate. And so one tool or one function

287
00:28:29,200 --> 00:28:34,560
that AI does pretty well is transcribing or creating a transcript of a conversation. It could

288
00:28:34,560 --> 00:28:41,360
be transcripts of a hotline complaint. It could be transcripts of an intake call that you have

289
00:28:41,360 --> 00:28:47,520
with somebody who's raising an employee relations concern. And once it does that first step of

290
00:28:47,520 --> 00:28:55,120
summariz... of transcribing the call or the interaction, an AI tool can begin summarizing

291
00:28:55,120 --> 00:29:00,880
or triaging the key elements or proposing next steps based on what the issues were that

292
00:29:00,880 --> 00:29:08,240
were raised in that conversation. So this can reduce the note-taking burden.

293
00:29:08,240 --> 00:29:13,520
I know that there are some HR departments where they will have two people in every meeting, which

294
00:29:13,520 --> 00:29:18,160
may still be a good practice, but one, the purpose of one of those individuals is to take notes

295
00:29:18,160 --> 00:29:23,280
or to try to take verbatim notes while the other person is carrying on the conversation.

296
00:29:23,280 --> 00:29:31,760
And this tool would help lift that burden. Any thoughts, Oliver, from your perspective on

297
00:29:31,760 --> 00:29:40,800
this opportunity? Well, I love the sound of it. Um, I have struggled to take notes accurately in

298
00:29:40,800 --> 00:29:48,080
many an interview and, you know, had to go back and interpret them. And so I really do see this as

299
00:29:48,080 --> 00:29:55,360
an opportunity, but one of my big concerns is just the accuracy component of this tool. And I

300
00:29:55,360 --> 00:30:00,640
think that's really important to analyze because I think for those of us who have done investigations

301
00:30:00,640 --> 00:30:08,400
or in employee relations, sometimes it's just not so cut and dry. Someone may say something to you

302
00:30:08,400 --> 00:30:14,640
um, that is coded or may not even be clear in the words that they're using, but you pick up on the

303
00:30:14,640 --> 00:30:20,480
fact that there is a bigger issue and you need to dive into it. And I think some of that doesn't

304
00:30:20,480 --> 00:30:26,480
always come across in a transcript that you get. And so it is important that you are reviewing

305
00:30:26,480 --> 00:30:33,200
transcripts even when you use a tool like this to make sure that the context is included and that

306
00:30:33,200 --> 00:30:38,560
what is coming across in the transcript makes sense and is accurate. And unfortunately, I think

307
00:30:38,560 --> 00:30:43,520
that means it's maybe not always as efficient as we want because if you have to go back and review

308
00:30:43,520 --> 00:30:49,360
it all, you know, that that takes time, too. So I think some concerns with using a tool like

309
00:30:49,360 --> 00:30:54,960
this, but I am hopeful that there is improvement and this just continues to get better and better

310
00:30:54,960 --> 00:31:01,120
because I do think it could be an efficiency save. Yeah, Oliver, I'm just going to add to that as well.

311
00:31:01,120 --> 00:31:06,880
Um, it can work, but I mean, you need to review it like right after you have the conversation, too,

312
00:31:06,880 --> 00:31:12,080
right? You know, I know in interviews, they're sort of in my head, I can hear that person talking and

313
00:31:12,080 --> 00:31:16,480
emphasizing things, but sometimes transcriptions, you'll go back a week later and you'll be

314
00:31:16,480 --> 00:31:21,040
like, "What were they talking about with the the owl and the pink purple suitcase?" You know what

315
00:31:21,040 --> 00:31:26,160
I mean? Like, or someone asked here about sort of accents and whether they picked up something. And

316
00:31:26,160 --> 00:31:30,880
if you were to go back right after and look at it and review it and sort of make changes and

317
00:31:30,880 --> 00:31:35,360
say, "No, no, that's not what they said, or here's the context," it could be helpful. But, you know, we

318
00:31:35,360 --> 00:31:40,960
know ER people are really busy. They're moving from one thing to another and they're humans. And so

319
00:31:40,960 --> 00:31:46,720
will they do that? And that is a concern. Yeah, and Deb, there's a concept called verification drift

320
00:31:46,720 --> 00:31:52,800
that some researchers around AI have talked about, which is a recognition that AI can be effective so

321
00:31:52,800 --> 00:31:56,960
long as you're doing what you just described, which is reviewing it really carefully. But there's this

322
00:31:56,960 --> 00:32:03,200
recognition, sort of a psychological recognition that people tend to begin to rely more and more

323
00:32:03,200 --> 00:32:09,360
on the output of the tool because a lot of it is seems accurate, seems good. Referencing

324
00:32:09,360 --> 00:32:14,960
back to the attorneys—I am an attorney—and and referencing back to the attorneys Oliver

325
00:32:14,960 --> 00:32:21,600
mentioned who have submitted case law that was not actually real, they're not bad lawyers. They relied

326
00:32:21,600 --> 00:32:26,720
too heavily on the tool before submitting their briefs. And that's something as HR professionals

327
00:32:26,720 --> 00:32:32,960
we have to be really careful about. I want to just address quickly June's question here about

328
00:32:32,960 --> 00:32:40,080
discovery of transcription. And that's an issue that I can certainly see coming up, which is if you

329
00:32:40,080 --> 00:32:45,520
use a transcript in your investigation and your subsequently, that matter is subsequently

330
00:32:45,520 --> 00:32:52,400
litigated, and you have to produce that transcript and explain how it was created, we have not seen

331
00:32:52,400 --> 00:32:57,760
that yet at all at such an early stage in the use of this technology, but it is something that we're

332
00:32:57,760 --> 00:33:02,560
talking with our investigations team about is how do we make sure that we have an explanation of how

333
00:33:02,560 --> 00:33:08,480
we use AI technology in our investigations and what are our answers going to be when a

334
00:33:08,480 --> 00:33:15,920
plaintiff's attorney challenges us or disputes what we said or what conclusions we found.

335
00:33:15,920 --> 00:33:21,120
And I think that's something we'll have to address with the transcription piece of it, too. And to

336
00:33:21,120 --> 00:33:27,120
Katrina's comment about AI in other languages, I do believe that AI transcription is pretty effective

337
00:33:27,120 --> 00:33:34,000
in other languages so long as the tool is fine-tuned for that. But background noise or

338
00:33:34,000 --> 00:33:40,320
folks who have an accent that can impact the quality. So it goes back to what Oliver said, which

339
00:33:40,320 --> 00:33:49,440
is you have to be vigilant about double-checking this. Back to you, Oliver. Yeah, thanks. I I'll stay

340
00:33:49,440 --> 00:33:56,720
kind of, um, off the top here on this one. Please, I recommend working with your employment counsel,

341
00:33:56,720 --> 00:34:02,720
your lawyer, your in-house counsel on some of these topics, you know, especially if you're talking

342
00:34:02,720 --> 00:34:09,120
about discovery. They'll likely have opinions about that, so make sure to check with them, too. And

343
00:34:09,120 --> 00:34:15,200
I think benefits is another topic that comes up as an opportunity, but you wouldn't want to explore

344
00:34:15,200 --> 00:34:21,840
without consulting your your internal counsel, too. I have benefits on this list because

345
00:34:21,840 --> 00:34:28,800
with benefits comes a lot of data. This is a very data-rich area. For those of us who provide benefits

346
00:34:28,800 --> 00:34:34,160
to our employees, which I think is probably most people on the call, it comes along with a lot of

347
00:34:34,160 --> 00:34:40,400
data that's both data of who's using the benefits, how often they're using the benefits, you know,

348
00:34:40,400 --> 00:34:47,520
once they register, do they continue using them? But also the vendors that we use collect

349
00:34:47,520 --> 00:34:52,320
a lot of information from our employees in providing those benefits. Think about health

350
00:34:52,320 --> 00:34:58,960
benefits, you know, the conditions that employees have and how often they're seeing

351
00:34:58,960 --> 00:35:04,640
providers because of those. It's a very data-rich environment around benefits. And that means it's

352
00:35:04,640 --> 00:35:10,880
an opportunity where AI can help us with some of the things we need to do. And that could be

353
00:35:10,880 --> 00:35:18,320
uh, maybe personalizing some of the offerings that we have to our employees, um, you know, helping them

354
00:35:18,320 --> 00:35:24,800
navigate what can be very tricky benefits by making suggestions based on demographics. People

355
00:35:24,800 --> 00:35:29,680
who have similar demographics used these benefits in the past, so maybe they would be

356
00:35:29,680 --> 00:35:36,080
interesting or useful to this type of employee. Another way that this could be helpful is

357
00:35:36,080 --> 00:35:42,080
just tracking and determining what benefits are most used. You know, we often do that just

358
00:35:42,080 --> 00:35:48,880
based on registration, but we don't always know after registration how the benefits are being

359
00:35:48,880 --> 00:35:56,800
used, how thoroughly. So analyzing that data, having some AI help can help understand how useful

360
00:35:56,800 --> 00:36:01,360
our benefits are and whether you want to continue them. Then the last would just be costing, you know,

361
00:36:01,360 --> 00:36:06,720
predicting the costs based on the data that you have and helping you better understand your

362
00:36:06,720 --> 00:36:17,840
costs. So some benefits there potentially from AI in the benefits space. Our next opportunity is

363
00:36:17,840 --> 00:36:23,920
certainly close to my experience, which is around employee relations and investigations. And

364
00:36:23,920 --> 00:36:28,000
I think this is a... this is a challenging opportunity because it's so wide-ranging,

365
00:36:28,000 --> 00:36:32,480
like there are so many different things or ways that an AI tool could be incorporated

366
00:36:32,480 --> 00:36:38,080
or relied upon in the employee relations space, and specifically in the context of investigations.

367
00:36:38,640 --> 00:36:44,560
What I hear a lot of people talking about when they think about using AI in investigations, and

368
00:36:44,560 --> 00:36:50,000
this is something that we've talked with Deb about, too, is everybody says, "Oh, well, I want to use it

369
00:36:50,000 --> 00:36:54,880
to summarize my interview notes," or, "I want to use it to summarize evidence." And as we've already

370
00:36:54,880 --> 00:36:59,280
alluded to, there are some real risks around that. And so I want to focus our attention rather on

371
00:36:59,280 --> 00:37:06,320
some other skills that AI can do pretty effectively in the context of investigations.

372
00:37:06,320 --> 00:37:14,000
One of the ways that we found AI, our AI tool to be particularly effective is just vetting or

373
00:37:14,000 --> 00:37:20,800
or querying the evidence we've gathered. So we might want to say, "How many witnesses

374
00:37:20,800 --> 00:37:28,400
uh, talked about this meeting on June 3rd?" And you can, you know, upload, in our case, we upload

375
00:37:28,400 --> 00:37:34,400
transcripts from interviews that we've conducted and emails that we've gathered, and the tool is

376
00:37:34,400 --> 00:37:40,000
pretty effective at going out and searching which employees referenced or talked about that meeting

377
00:37:40,000 --> 00:37:46,960
on June 2nd. Another way that we've experimented with it is saying, "Okay, well, how many witnesses

378
00:37:46,960 --> 00:37:55,760
corroborated that, you know, John made a comment about age in that meeting on June 2nd?"

379
00:37:55,760 --> 00:38:01,360
And the tool is pretty effective at going out and identifying where in the evidence and

380
00:38:01,360 --> 00:38:07,040
which witnesses made these particular comments or can corroborate that they heard John make

381
00:38:07,040 --> 00:38:12,720
that comment, for example. So querying evidence and querying information you've gathered during

382
00:38:12,720 --> 00:38:20,320
an investigation is a skill that an AI tool does pretty effectively, and it doesn't have some of the

383
00:38:20,320 --> 00:38:27,440
risks that delegating the full responsibility for summarizing a witness interview has. The

384
00:38:27,440 --> 00:38:31,600
reason that we're concerned generally about the idea of summarizing a whole witness interview

385
00:38:31,600 --> 00:38:38,560
with AI is that if you go and tell the tool, "Go summarize, you know, summarize Sally's complaint,"

386
00:38:38,560 --> 00:38:43,920
if that's all the information you've given it, the tool is going to come back with radically

387
00:38:43,920 --> 00:38:49,440
different responses on three different occasions that you give it that that direction.

388
00:38:49,440 --> 00:38:54,240
So the point is, if you think back to what I said earlier about probabilistic technology, it's making

389
00:38:54,240 --> 00:39:00,720
its best estimate based on the information you've given it as the user or the prompter, and based on

390
00:39:00,720 --> 00:39:06,080
the volumes of data that it's been trained on. So if you want a better summarization or you

391
00:39:06,080 --> 00:39:12,240
want a better output on a summary, for example, of Sarah's complaint, you have to as the user provide

392
00:39:12,240 --> 00:39:20,000
a really detailed, narrow context or prompt about what specific information you want the tool to

393
00:39:20,000 --> 00:39:25,840
summarize. So that's one of the challenges that you have to think about. Another skill

394
00:39:25,840 --> 00:39:30,800
that we found it to be pretty effective about in the context of investigations is drafting,

395
00:39:30,800 --> 00:39:34,800
identifying witnesses who you might want to interview, or identifying documents that you

396
00:39:34,800 --> 00:39:39,920
might want to gather, identifying the potential scope of the investigation. You can upload the

397
00:39:39,920 --> 00:39:44,720
complaint, for example, and it does a pretty good job of identifying the scope if you include your

398
00:39:44,720 --> 00:39:50,880
policy, it can identify what policies it might be... the conduct might violate. Uh, it can also

399
00:39:50,880 --> 00:39:58,160
propose questions that you should be considering during the investigation. So one way to think about

400
00:39:58,160 --> 00:40:04,160
an AI tool in this context is sort of like an eager associate investigator, super smart, really

401
00:40:04,160 --> 00:40:08,800
eager to help and support you, but actually has no understanding of the investigation process

402
00:40:08,800 --> 00:40:13,040
without you providing it lots of information and training. So those are a couple thoughts

403
00:40:13,040 --> 00:40:17,760
where I think there's an opportunity there, but as with everything, there's some risk around it.

404
00:40:17,760 --> 00:40:24,720
Deb, what are your thoughts on that? And you're on mute. Thanks. Yeah, I have lots of thoughts on this,

405
00:40:24,720 --> 00:40:29,680
but I agree with everything you said. I think the problem with a summary—a summary by definition is

406
00:40:29,680 --> 00:40:34,560
sort of taking something and condensing it. And as you said, it's deciding what's important and what's

407
00:40:34,560 --> 00:40:41,440
not. You can assume that AI bot has never worked in a plant, has maybe never done a sales job. So those

408
00:40:41,440 --> 00:40:46,640
little things that actually, you know, being in your organization or just, you know, being an outside

409
00:40:46,640 --> 00:40:52,400
investigator knowing enough to ask those questions may get glossed over or may get eliminated. I'll

410
00:40:52,400 --> 00:40:59,200
never forget, I had, this was not AI, it was years ago, but I had someone take notes for me and they

411
00:40:59,200 --> 00:41:03,280
weren't a very good notetaker, and there was one thing that the person said that afterwards I was

412
00:41:03,280 --> 00:41:06,080
like, "Oh my god, could you believe..." And I went to their notes, they didn't even mark

413
00:41:06,080 --> 00:41:12,800
it down because to them that was not relevant. To me, it was very short of a smoking gun. And so

414
00:41:12,800 --> 00:41:16,720
um, because I was thinking about the connection, everything like that. So I think

415
00:41:16,720 --> 00:41:22,080
I love, and that's some of the things, you know, not to give away things that we're doing in

416
00:41:22,080 --> 00:41:26,880
HR Acuity, because we know there are ways to assist the investigator, make sure you scope it well, you do

417
00:41:26,880 --> 00:41:31,040
a comprehensive investigation, which we know is one of the things that as an investigator

418
00:41:31,040 --> 00:41:37,920
you need to do to make sure it's compliant. Um, and that's exciting. Oliver, I want to ask you a

419
00:41:37,920 --> 00:41:41,680
question in this in the context of employee relations and investigations with AI because

420
00:41:41,680 --> 00:41:47,600
I think one of the one of the sort of the devil's... you know, the devil's in the details

421
00:41:47,600 --> 00:41:53,600
issues here is the risk of delegating authority for an investigation or for an employee relations

422
00:41:53,600 --> 00:42:01,840
decision. Oliver, how would you evaluate that risk? Oh, very... I would say it's high. I would not want

423
00:42:01,840 --> 00:42:08,400
um, to delegate authority. And for some of the reasons that you all mentioned, I think it's really

424
00:42:08,400 --> 00:42:18,080
important to have that human decision-making point all through the investigation. So I think AI can

425
00:42:18,080 --> 00:42:26,080
help the human make decisions. And I think, Eli, you raised a great point about how a human can use AI

426
00:42:26,080 --> 00:42:32,240
to be more prepared to make a better decision by querying the evidence that exists in the tool. But

427
00:42:32,240 --> 00:42:38,560
I really think that it should be the investigator who uses their critical thinking to make judgments

428
00:42:38,560 --> 00:42:43,920
throughout the investigation and certainly to make the ultimate finding. You know, there are so

429
00:42:43,920 --> 00:42:50,000
many things that go into that. I don't think that AI is sophisticated and advanced enough yet to

430
00:42:50,000 --> 00:42:56,800
be delegating that to. And, you know, that's a good thing for our profession. We're still

431
00:42:56,800 --> 00:43:03,520
necessary as humans, HR professionals, investigators to do that. I think one of my other concerns is

432
00:43:03,520 --> 00:43:11,200
some of the things you're saying AI can do, that's how I train people on investigations and

433
00:43:11,200 --> 00:43:16,480
employee relations. You know, Deb mentions having someone take notes and then kind of issue spot.

434
00:43:16,480 --> 00:43:22,560
Well, that's how I learned, and I assume that's how a lot of people learn how to really understand

435
00:43:22,560 --> 00:43:26,960
what's important when you're talking to people and how it plays into the bigger picture in an

436
00:43:26,960 --> 00:43:33,600
investigation. So if you're having AI do that, you know, do we miss a generation of investigators?

437
00:43:33,600 --> 00:43:39,360
You know, I think that's incumbent upon us as leaders to think through if we use AI in this

438
00:43:39,360 --> 00:43:43,760
way, how are we going to train the next generation? I don't have an answer to that, but it's definitely

439
00:43:43,760 --> 00:43:50,160
on my mind. Yeah, that's a great observation. Go ahead, Deb. The other opportunity or the issue is

440
00:43:50,160 --> 00:43:56,320
just bias, right? We know AI is inherently biased. And what have we been trying to do for years

441
00:43:56,320 --> 00:44:00,800
and decades is take bias out of the investigation process. So now you're layering something on that's

442
00:44:00,800 --> 00:44:08,960
known to have that problem. So also a big concern, obviously. For sure, great point. Well, why don't we

443
00:44:08,960 --> 00:44:13,440
talk about our last opportunity, which I, Oliver, I hope there are no risks with this one. I hope

444
00:44:13,440 --> 00:44:19,040
we can just start using it immediately to do this. Yeah, I'm sorry to be the bearer of bad

445
00:44:19,040 --> 00:44:26,880
news. This one is complicated, but, you know, a lot of employers have data when it comes to

446
00:44:26,880 --> 00:44:32,960
performance. And especially in retail, healthcare space, things that are important to us include time

447
00:44:32,960 --> 00:44:40,080
and attendance and productivity. And those are both very data-rich environments. And AI can help

448
00:44:40,080 --> 00:44:44,800
us there, compile, you know, how people are living up to their time and attendance requirements

449
00:44:44,800 --> 00:44:50,480
and how productive they're being, and then put that into a nice little package for a leader on

450
00:44:50,480 --> 00:44:56,960
their each individual employee. And I, you know, I'm sure you've all heard this. It's hard

451
00:44:56,960 --> 00:45:04,800
to write PDRs. It's it's time-consuming. And so this is an area where AI may be able to help. However,

452
00:45:04,800 --> 00:45:12,960
it is a very risky area, and performance is something where context matters, and the human

453
00:45:12,960 --> 00:45:19,440
decision-making discretion is very important. So I've listed this here with the caveat that

454
00:45:19,440 --> 00:45:24,720
you got to be very careful. Yeah, it sounds like a tremendous opportunity, and I know that there are

455
00:45:24,720 --> 00:45:29,760
going to be tools that are that are fine-tuned to to help with this, but it also sounds like, as with

456
00:45:29,760 --> 00:45:34,640
many of the opportunities we're talking about, the importance of human judgment and making

457
00:45:34,640 --> 00:45:41,360
sure that the human is still the decision maker, is still the one driving the analysis,

458
00:45:41,360 --> 00:45:48,560
and is only using the tool to the extent that it can be helpful in gathering or consolidating,

459
00:45:48,560 --> 00:45:53,680
you know, summarizing or collecting information, but ultimately the user has to be the one who's

460
00:45:53,680 --> 00:46:01,680
driving this, right? Yep, totally agree. Well, that probably felt like five opportunities with a

461
00:46:01,680 --> 00:46:06,960
lot of risks. Now we're going to talk about five risks that also have some opportunities. I think

462
00:46:06,960 --> 00:46:10,960
that's the challenge with this topic, but it's also the thing that makes it kind of exciting.

463
00:46:10,960 --> 00:46:18,640
Uh, so why don't we move on to the first risk, which is using AI to make employment decisions?

464
00:46:18,640 --> 00:46:24,640
Now, this obviously is a significant area of risk because as we've talked about, and I see some

465
00:46:24,640 --> 00:46:29,680
of the comments in the chat even expressing some concern around delegating too much responsibility

466
00:46:29,680 --> 00:46:36,480
to AI, and certainly relying on the tool to make an employment decision, which that might look

467
00:46:36,480 --> 00:46:43,680
like saying, "Okay, here's the investigation report. What should the action be?" Well, that would be an

468
00:46:43,680 --> 00:46:48,960
inappropriate delegation of authority. The user, the HR professional, simply has to be the one

469
00:46:48,960 --> 00:46:54,800
who's evaluating past practice, has to be the one who's evaluating the impact and relevance of

470
00:46:54,800 --> 00:46:59,680
the evidence that was gathered, the findings that were made in the investigation or the analysis. So

471
00:46:59,680 --> 00:47:06,320
I certainly think there's tremendous risk in using AI to make employment decisions. On the other hand,

472
00:47:06,320 --> 00:47:13,120
Oliver, there probably are some opportunities here around using AI to inform or help an HR

473
00:47:13,120 --> 00:47:20,000
professional in this way. Is that true? Yeah, I think so. I do think there are organizations

474
00:47:20,000 --> 00:47:26,480
that don't have other systems that that can do some of what you're saying, you know, track

475
00:47:26,480 --> 00:47:32,560
what they've done in the past. And so there may be opportunity for AI to give you data on that

476
00:47:32,560 --> 00:47:40,800
rather than make a decision, really give you that data. Also could help give context around

477
00:47:40,800 --> 00:47:49,520
what past practice has been or what other company documents, policies, practices would

478
00:47:49,520 --> 00:47:55,920
indicate. So I do think there's some opportunity to leverage AI here. Actually making employment

479
00:47:55,920 --> 00:48:06,000
decisions, I think that has to really rest with the professionals and rely on the human judgment.

480
00:48:06,000 --> 00:48:13,760
Uh, related to that is the hiring decision. And I include this here because a lot of the tools

481
00:48:13,760 --> 00:48:19,200
that I'm coming across right now are focused on helping in the talent acquisition space,

482
00:48:19,200 --> 00:48:26,240
making that process more efficient. And you know, some of that's about reviewing résumés,

483
00:48:26,240 --> 00:48:33,920
um, using AI to to review résumés and then spit out the best ones based on the job description. But

484
00:48:33,920 --> 00:48:41,040
I've also seen AI used to conduct interviews. You know, it can display questions that the person

485
00:48:41,040 --> 00:48:48,320
answers on video, and then it can ask follow-up questions to whatever the person responds. And

486
00:48:48,320 --> 00:48:56,480
so that's another function that I'm seeing. And I think all of this is very risky because that

487
00:48:56,480 --> 00:49:02,800
hiring decision and, you know, even whether you advance someone's résumé to the next stage of

488
00:49:02,800 --> 00:49:09,120
the hiring process, there's risk in all of that. That's all part of an employment decision, and

489
00:49:09,120 --> 00:49:16,800
so it needs to be very, very careful. And bias is a huge factor in this process. And, you know, for

490
00:49:16,800 --> 00:49:23,280
those who have looked into kind of résumés, for example, um, it's very common knowledge, I think,

491
00:49:23,280 --> 00:49:29,280
that names on résumés can have a lot of bias. You know, we have seen just changing the name

492
00:49:29,280 --> 00:49:35,600
at the top of the résumé can change the way it's interpreted, and AI will learn based on some

493
00:49:35,600 --> 00:49:42,880
of those practices. So it will not be perfect, and I do have a lot of concern about using these tools

494
00:49:42,880 --> 00:49:49,920
um, in the hiring process. Any opportunities you see there, though, Eli? Well, I think the bias issue

495
00:49:49,920 --> 00:49:54,160
is fascinating, and we're we're going to dive into that a little bit more as risk number three. But

496
00:49:54,160 --> 00:50:02,400
in this context, you know, we also have to remember that there is bias in human nature and that humans,

497
00:50:02,400 --> 00:50:07,680
especially if we have folks who are maybe not well trained or don't have the experience in

498
00:50:07,680 --> 00:50:14,160
making hiring decisions, that they actually could insert more bias into the process and actually

499
00:50:14,160 --> 00:50:19,760
a technology tool may who that's well-formulated and well-designed and is specifically

500
00:50:19,760 --> 00:50:26,640
has agents to help address bias, it actually could result in better consistency and accuracy

501
00:50:26,640 --> 00:50:32,960
and could take out some of that human bias that we so much want to avoid. So I think it's

502
00:50:32,960 --> 00:50:39,840
an interesting dichotomy, is that while I think there's a lot of risk here in helping employers

503
00:50:39,840 --> 00:50:45,920
make responsible decisions around recruiting and hiring, I suspect that there will be AI tools

504
00:50:45,920 --> 00:50:52,800
that are fine-tuned to help eliminate some of that risk that humans actually insert into the system.

505
00:50:53,360 --> 00:51:00,080
Yeah, I think it's an area where we should be looking for studies and data where the

506
00:51:00,080 --> 00:51:06,400
vendors who are offering to do these things have really looked into what happens when

507
00:51:06,400 --> 00:51:12,400
they do things like change names on a résumé or change the appearance of someone that the

508
00:51:12,400 --> 00:51:19,600
AI is interviewing, and they've proven in their their data that there there isn't bias or that

509
00:51:19,600 --> 00:51:25,920
they've really improved that that component. Well, and so with that, why don't we talk a little bit

510
00:51:25,920 --> 00:51:31,680
more about bias? Because it's sort of the, you know, the the 800-pound gorilla in the room is the fact

511
00:51:31,680 --> 00:51:38,560
that every AI system seems to have a significant amount of bias integrated into it or built

512
00:51:38,560 --> 00:51:44,320
into it. And so let's understand why that is first. And the reason that is, the reason that

513
00:51:44,320 --> 00:51:52,800
we are so concerned about bias in AI tools is that as a large language model is trained on all

514
00:51:52,800 --> 00:52:00,160
of the volumes and volumes of data, it is trained on data that is human-created, that has intrinsic

515
00:52:00,160 --> 00:52:10,160
biases and that already has intrinsic bias or unconscious bias built into it. And so

516
00:52:10,160 --> 00:52:15,440
one of the challenges is there's... and if you've, you know, read up on how these tools have been

517
00:52:15,440 --> 00:52:20,960
trained, there's not a lot of focus on how to make sure that these tools are unbiased when they are

518
00:52:20,960 --> 00:52:26,720
being trained on all of these volumes and volumes and volumes and volumes and volumes of data. And so

519
00:52:26,720 --> 00:52:32,800
a lot of the concern is that, and if you think about it like any human being who has their own

520
00:52:32,800 --> 00:52:38,160
lived experience and their own background and their own perspectives, we all come to our jobs

521
00:52:38,160 --> 00:52:45,600
and to this day with our own experiences and our own biases. And that is magnified in an artificial

522
00:52:45,600 --> 00:52:52,080
intelligence tool. And that's why that's so concerning about how the bias may

523
00:52:52,080 --> 00:52:58,640
manifest in the outputs that these tools create. For example, this is several months ago now and,

524
00:52:58,640 --> 00:53:08,080
um, but there was an experiment done where an AI tool was asked to describe individuals,

525
00:53:08,080 --> 00:53:15,280
including gender, race, age of individuals in a boardroom who were board members of a large

526
00:53:15,280 --> 00:53:22,080
corporation. And pretty consistently, the tools came back and described the board members as

527
00:53:22,080 --> 00:53:29,280
middle-aged white men. And that is obviously reflects significant bias. And so that's

528
00:53:29,280 --> 00:53:34,080
the type of bias that we're concerned about in large language models. And so if we're delegating

529
00:53:34,080 --> 00:53:40,960
responsibility on decision-making or hiring or things like that and we haven't ensured that the tool is

530
00:53:40,960 --> 00:53:48,880
uh, designed to eliminate or mitigate that bias, we could be inserting exactly the bias that we've

531
00:53:48,880 --> 00:53:57,360
been working for so many years to try to get out of our decisions as attorneys and HR professionals.

532
00:53:57,360 --> 00:54:03,760
Fully agree. And I hope that everyone has been working to remove those biases. And I think the

533
00:54:03,760 --> 00:54:10,160
only, you know, counterpoint to this is, uh, for those who haven't worked on that, the

534
00:54:10,160 --> 00:54:16,720
the these vendors are working on that. So, you know, um, I think it goes back to what you

535
00:54:16,720 --> 00:54:23,440
said earlier, Eli, uh, it is possible that the AI tools are better than humans.

536
00:54:23,440 --> 00:54:28,080
You know, they don't have conscious bias. They have these built-in biases, but,

537
00:54:28,080 --> 00:54:32,160
um, you know, I think that's something we need to think deeply about. And hopefully everyone

538
00:54:32,160 --> 00:54:42,640
is working on those biases. Um, but I think this is a very important thing to be aware of with AI.

539
00:54:42,640 --> 00:54:51,920
One of my um, fears in reliance on AI is that we really lose our our listening skills and ability

540
00:54:51,920 --> 00:54:58,560
to hear what's really going on, maybe not the words people are using, but um, what's really going on

541
00:54:58,560 --> 00:55:06,160
under the surface. And I think that is one of my biggest assets as an HR professional, is um, those

542
00:55:06,160 --> 00:55:13,520
listening skills. But if we are teaching people to rely on artificial intelligence to listen to

543
00:55:13,520 --> 00:55:22,080
conversations, summarize them, even identify the most important issues, I worry that over time,

544
00:55:22,080 --> 00:55:26,720
um, we'll start tuning out. People will start tuning out and won't really hone those

545
00:55:26,720 --> 00:55:32,160
listening skills as much. Uh, and I think that's one of the most valuable parts of our profession.

546
00:55:32,160 --> 00:55:38,960
And I hope that that doesn't happen. But uh, I think there is some risk there. I think that really plays

547
00:55:38,960 --> 00:55:44,720
into our last risk here, too. Uh, and I think with these two risks, I don't know that there's

548
00:55:44,720 --> 00:55:48,400
a lot of opportunity. I think they're really just cautionary tales that we have to be thinking about

549
00:55:48,400 --> 00:55:54,640
as we're evaluating how to integrate AI into the HR profession. I think one of the areas, and we've

550
00:55:54,640 --> 00:56:00,720
alluded to this a little bit, is just the concern about reducing the reliance in an organization

551
00:56:00,720 --> 00:56:06,320
on the true value that HR professionals bring to the table. Uh, HR professionals are well

552
00:56:06,320 --> 00:56:12,400
trained, they are experienced, and probably the experience that many of you had is more valuable

553
00:56:12,400 --> 00:56:19,520
than any training you can give in, you know, a week or, you know, five days. Um, it's really that in the

554
00:56:19,520 --> 00:56:24,240
trenches experience dealing with individuals in unique, uncertain situations and using

555
00:56:24,240 --> 00:56:30,320
your judgment and critical thinking skills and emotional intelligence to come up with a solid

556
00:56:30,320 --> 00:56:36,560
solution to a given employee relations issue. And I think that's one of the concerns is both

557
00:56:36,560 --> 00:56:42,080
reducing or suggesting to an organization, to leaders in an organization, that the true value of

558
00:56:42,080 --> 00:56:49,520
an HR professional can be replaced in somehow by an AI tool. Another concern is something that

559
00:56:49,520 --> 00:56:55,200
Oliver mentioned earlier, which is how do we ensure that the next generation of HR professionals is

560
00:56:55,200 --> 00:57:00,960
well trained if we're delegating some of these ministerial tasks to an AI tool and not using

561
00:57:00,960 --> 00:57:10,160
those as an entryway to the HR profession? Um, thoughts on those uh concerns, Oliver? Yeah, I have

562
00:57:10,160 --> 00:57:17,360
all of those concerns. And I think that that makes it important for us to really start grappling with

563
00:57:17,360 --> 00:57:24,400
AI and understanding it and leveraging it in ways that make sense and are lower on the risk profile,

564
00:57:24,400 --> 00:57:30,480
um, but that make us more efficient, because I think leaders and organizations will demand

565
00:57:30,480 --> 00:57:35,680
that we understand AI and that we're able to leverage it. And I hope that that really helps

566
00:57:35,680 --> 00:57:42,400
us focus kind of at the top of our license where those HR skills, the listening skills, the judgment

567
00:57:42,400 --> 00:57:48,880
skills, where that's really useful and that people still rely on us for that. I think they will if we

568
00:57:48,880 --> 00:57:57,360
use AI to be more efficient on the other tasks. So wrapping it up, uh, let's talk about what

569
00:57:57,360 --> 00:58:03,680
does this really mean for HR professionals? Yeah. Well, first of all, don't be left behind. There's

570
00:58:03,680 --> 00:58:09,280
a lot going on in this space right now. You're you're here today, so that's a great start. But

571
00:58:09,280 --> 00:58:15,520
um, you know, keep learning as much as you can about this and play with tools on your personal time. You

572
00:58:15,520 --> 00:58:21,760
know, it's fun. And there are ways you can build those skill sets. So don't be left behind and get

573
00:58:21,760 --> 00:58:29,600
in the game here. Make sure you're staying on top of this area that's rapidly developing.

574
00:58:32,000 --> 00:58:36,320
Make sure that you understand the limitations and that when people ask

575
00:58:36,320 --> 00:58:42,160
you to implement a tool, you can talk to them about why that might be a challenging idea or

576
00:58:42,160 --> 00:58:48,240
even a bad idea that can include bias or the reliability of the information or even just

577
00:58:48,240 --> 00:58:55,360
the adequacy of the output. Um, but make sure you're able to talk through those issues, too.

578
00:58:57,120 --> 00:59:02,080
Make sure you're using it as a support tool, but not to replace your independent judgment. I think

579
00:59:02,080 --> 00:59:08,640
we've talked about that a few times today. And then make sure you're analyzing the various

580
00:59:08,640 --> 00:59:15,360
risks. I think I said earlier, work with your legal teams, work with your privacy teams, your IT teams.

581
00:59:15,360 --> 00:59:19,760
A lot of them are focused on this right now. So make sure you're bringing in partners to analyze

582
00:59:19,760 --> 00:59:24,880
all that risk. And then last but certainly not least, hopefully you heard in much of

583
00:59:24,880 --> 00:59:31,040
what we said today, your independent judgment remains very important even if you're using

584
00:59:31,040 --> 00:59:38,800
these tools. So keep that at the forefront, even when implementing AI. Great. Thank

585
00:59:38,800 --> 00:59:42,960
you, Oliver, and thank you for the discussion, all of you. It's great to have comments and

586
00:59:42,960 --> 00:59:47,440
uh, questions from you. And pleased to be here. It sounds like you have some opportunities coming

587
00:59:47,440 --> 00:59:53,760
up to discuss this more with the HR Acuity and Empower team. So I'll turn it back over to Deb.

588
00:59:53,760 --> 00:59:58,640
Yeah, thank you so much. This was great. Obviously timely. I think everybody was sort of leaned into

589
00:59:58,640 --> 1:00:03,520
everything you were saying. Of course, for me, every time you mentioned Oliver, I thought Oliver, our ER

590
1:00:03,520 --> 1:00:08,240
companion who's in our platform, would show up. But Oliver, you're on the same wavelength, so

591
1:00:08,240 --> 1:00:14,160
you know, you're... all Olivers think alike, I guess. So it's great. But thank you.

592
1:00:14,160 --> 1:00:17,920
You know, Eli and Oliver's contact information is here if you want to reach out to them for more

593
1:00:17,920 --> 1:00:21,600
information. You should. But we love having them as part of our community. And thank you

594
1:00:21,600 --> 1:00:26,080
again to everybody for joining us. And stay tuned for more conversations. This is going to... we're going to

595
1:00:26,080 --> 1:00:32,320
keep talking about AI for a long time. Thanks, everyone. Have a great afternoon. Thank you so much.