The HR Leader's Guide to Multiagent HR Architecture, with Josh Bersin

[00:00:03] Welcome to the Human Capital Leadership. We're your source for personal, professional, and organizational growth and development. We share our own original research, explore industry trends, and interview executives and thought leaders from across the globe. Join us for practitioner-oriented content around all things leadership, HR, talent management, organizational development, and change management. Maximize your personal and organizational potential with the HCI Family of Podcasts.

[00:00:35] Josh Bersin, welcome to the conversation today. Thank you, Jonathan. It's great to be here. It is great to be with you again. You're joining us from San Francisco. I'm south of Salt Lake City in Utah. Today, we are going to be talking about the latest report out from the Josh Bersin Company, The HR Leaders Guide to Multiagent HR Architecture, which is part of the HR 2030 Initiative,

[00:00:58] which outlines how HR leaders must govern specialized AI agents and functional super agents. I'm super excited to have the opportunity to sit down with you and to discuss this. We were scheduled previously, had some travel mishaps and whatnot. We had to reschedule. Anyways, thank you for your patience, and it's a pleasure to sit down. As we get started, I wanted to share Josh's bio with everyone, though I think everyone knows who you are, Josh.

[00:01:25] Josh is a well-known analyst, author, educator, and thought leader focusing on the global talent market and the changes and trends impacting business workforces around the world. He studies the world of work, HR, and leadership practices in the broad talent technology market. He is often cited as one of the leading HR and workplace industry analysts in the world. And I think this may be, I don't know, the third or fourth time we've had an opportunity to chat. Always a pleasure.

[00:01:55] Before we dive on in, anything you would like to share with the audience by way of your own background, personal context, or about the company? And then we'll get going with the report. I mean, at the moment, it's all about AI. Yeah, yeah. It's all everybody wants to talk about, so we're spending most of our time on that. But of course, we don't have to talk about it today, but there's lots of other questions about pay and job titles and roles and structure of companies and so forth that's being affected too. But let's talk about AI first.

[00:02:25] Yeah, for sure. It's ubiquitous. Like, it's really at the forefront of everyone's attention, which is understandable, sometimes frustrating. It is what it is, right? Well, first, tell us a little bit about the report generally. Maybe even start with HR 2030 generally, the initiative, and why this report?

[00:02:51] Well, we started HR 2030 about a year and a half ago because there has been this proliferation of ideas ideas of what AI could do to different parts of HR from recruiting, interviewing, assessing, training, employee experience, analysis, pay rewards, et cetera. And there's going to be hundreds of use cases of AI, of course. But we wanted to kind of get the big picture.

[00:03:17] So we sat down and said, let's imagine we could fast forward ahead four and a half years. What do we want the world to look like ideally? And so we ended up with an architecture of 160 agents, which sounds ridiculous, but it makes sense when you think about it, interoperating with each other in a certain sort of networking effect with eight super agents that manage and coordinate the 160 other agents.

[00:03:47] And we introduced this. And the reason we did, the way we came up with that is we had spent the last 10 or 15 years studying HR and the different, the 95 functional domains of HR. So we kind of knew a lot about the complexity behind this. And we knew, and we had already done a whole bunch of research on what we call systemic HR, which is HR operating as an integrated system, as opposed to a whole bunch of standalone services. And, you know, I didn't really know what people are going to think of it.

[00:04:16] It was almost an intellectual exercise as much as anything else. Then we built a whole model, a visual model of it and everything. And we started showing it to people and they said, oh my God, please come help us because we're getting flooded with vendors and tools and technologies and bots and stuff. So it's become a body of research and a set of tools to help companies decide where they are and where they want to go in their roadmaps. Because this is a lot of change happening out there.

[00:04:45] And then the multi-agent architecture piece is one chapter. There's a bunch of elements to this. But to deal with the issue of this is a system of agents. This isn't an agent. And you could have debate. I just had a meeting last night with a computer science guy. You know, is a big agent better or a small agent? And that's part of what this is about. And if you do have a bunch of small agents, how do you manage them all? Right. Yeah.

[00:05:12] So 160, you know, specialized agents, the eight functional super agents. Since, you know, we are in the world of agentic AI, long gone are the days of just, you know, chat GPT and large language models and bots. Right. Like, right. That is so far behind us now. And like you said, there were so many tools, so many vendors.

[00:05:39] I feel like I'm on calls constantly talking with folks about the new and greatest, coolest thing. And there really are so many cool things, right? There's no doubt about it. I'm sure you're constantly on calls, too, about all of these things. And but but yeah, getting beyond getting beyond just the cool little tool or the new the new fad or the new thing that comes out. And every week there's something new.

[00:06:05] It's all about the agents and it's all about how to how to organize and manage the agents, which which I just find so fascinating. And any time I get a chance to talk with somebody who really has figured out how how to how to connect and integrate the agents, that's that's really what it's all about, because because so many people have, you know, have agents and so many people are talking about agentic AI.

[00:06:33] I've seen very few who actually have it systematized, who really have it integrated, who really have, you know, an approach that seems to be working. Yeah. In any systematic way. You know, just so much so many teams, so many organizations. Right. It's just so haphazard. Right. Well, in fact, you know, the the the the misnomer or misunderstanding is that if I build a little tool, a little AI agent that does something, that's a good thing.

[00:07:00] Well, it may be, but it may also not be. And I actually think there's a downside to doing that, because if you have your own little personal agent that does something and it's not correct. Right. What other first of all, what are you going to do to fix it in a big company and who else is using it or relying upon it for what? And are you proliferating the mistakes you're making into other people's agents? There's you know, these are not simple little tools.

[00:07:29] They look simple when you play around with chat GPT, but they learn over time and they get smarter and smarter and smarter. So I'm as an architectural thinker really trying to tell people to just think about the big picture before you dive into this cool little thing that you want to work with in terms of scale, ultimate connectivity, credibility, trust, protection of assets, protection of data, protection of security. And I think that's doable.

[00:07:58] But the funny thing about it, Jonathan, is I've since we started this, I've been reading up on, you know, all these people that have done lots of work on this Palantir and IBM and others. There's not that much written about how to manage a complex multi-agent system. It's not because the actual activity and behavior of an agent is not well known. I mean, the guy I was talking to last night does AI safety research. And he said, you know, there's like five or six different tributaries of research on understanding why an agent does what it does.

[00:08:27] And so I think what we've come up with is enough information. And I talked to SAP in great detail about this at their conference. I've talked to Workday about it. You know, I think there's some really important lessons here. So that and the other reason that I, you know, I think this we have to think of this systemically is if you think of HR as a bunch of standalone services, you're going to get your budget cut. People are going to say, well, why are we spending so much money? Cut the headcount in half.

[00:08:56] And let's just do less of this and turn it more over to agents. If you think about it as an enablement function for growth, and that's what I think HR ultimately is, is a dynamic enablement for growth through hiring, through training, through redeployment, through skilling, whatever it may be, through pay or job redesign or scheduling or, you know,

[00:09:26] unhappy, then you want these things to work well. So you can learn because we're, what we're going to do is we're going to learn lots and lots of things about our companies that we didn't know before because AI is so smart. Yeah. I love that. The enablement function for growth. I love that framing. And you just mentioned it, you know, a minute ago, I want to double click on it. You know, the, the failure that comes from unconnected agents. Right. Which, which is happening all around us.

[00:09:53] So everyone, you know, we, we feel so special because we, we create our own little agent and then, and then, you know, which may work really well for our own little workflow in our own little world and our own little silo. But it can cause problems like, and, and. Well, just yesterday, I don't know if you saw yesterday, Goldman Sachs just shut off Anthropic for all their non-ITF professionals. I don't want people hacking around with it anymore because it was expensive and it was actually not working out very well.

[00:10:23] I think one simple way to think of it is if you gave everybody a copy of Excel and said, build a spreadsheet that analyzes whatever it is you're doing and just share it with everybody else in the company. Can you imagine how many errors would be suddenly just proliferated throughout the company? Right. Because everybody's spreadsheets got glitches in it and it works for me, but it wouldn't work for you because you don't know how I did it or why I did it. You know, it's a little bit like that. And, and we won't, you know, so I'm not saying personal agents are going away.

[00:10:50] We're all going to have all sorts of cool personal agents, but let's just do it in a managed process. And another analogy that I think is useful is the more I talk to companies about this, the more I do believe that the agents are like people. So if you think about a new employee in a company, the employee has a job title and a role and a set of goals and responsibilities. And you don't give the employee access to all of the information in the company, all

[00:11:19] of the secret projects, all the financial information, because they're going to make a mess, right? Because that's not their job. So we think of the agents as individuals with very clearly defined specific responsibilities that we want them to be really, really, really good at this. All of a sudden, that point of expertise is really useful to other agents who need it and don't want to replicate it.

[00:11:44] So I think division of responsibility or whatever the right phrase, there is a word for that. I don't think that's exactly the phrase. Clear delineation of who's responsible for what really makes AI safer and more scalable and more powerful. And there's different ways of mixing things up as the paper discusses. But I think a lot of these rogue agents that went nuts and hacked into other websites were

[00:12:12] probably experiments where the agent was deliberately given a whole bunch of opportunities to make mistakes and they were testing how far it would go. And because these things are probabilistic, they're going to keep trying to achieve their goal until they find a way to achieve it. And if you give it too much opportunity, it's going to go down to the tail end of this probabilities curve and do something dumb. Right. Yeah. Yeah, absolutely.

[00:12:38] One of the things that's talked about in the report is workload reduction. And of course, if we have a bunch of productive agents, this is bound to happen, right? So proper management of agentic families can reduce HR workloads by up to 40% is one of the findings in the paper. Talk to us a little bit more about that. What you mean by these agentic families.

[00:13:04] And you talked a minute ago about reducing HR budgets. If we don't work systemically together, like different HR functions, we're just going to find ourselves having our budgets reduced versus being the enablement for the business. How do you look at this finding along with kind of what we've been talking about up to this point?

[00:13:32] Well, the first part of it is how you frame the AI investment. If you go to your boss and say, hey, as a result of spending all this money on AI, I'm going to cut our costs by 72%. Well, you just basically put yourself in a hole. You just did the wrong kind of business case. If you, on the other hand, if you go and say, as a result of this AI project, we're going to be able to hire people 75% faster with two-thirds higher quality of hire.

[00:13:59] Therefore, we can grow our business at a higher rate of speed and do more things than we wanted to do before. That's a different way of framing the problem. So that's the first thing. The second thing is, if you look at the behavior of the HR people in the company today, we've got maybe an instructional performance consultant, an instructional designer, a content developer, an LMS administrator, and training. And then you go to recruiting.

[00:14:26] We've got an HR business partner, a sourcer, a screener, an interview scheduler. We have all these job titles that are kind of narrow and kind of tactical sounding because what they make up is a step in a process that somebody years ago broke down into a job family. They took a process and they said, let's just take it apart like an industrial process and put it into a job family.

[00:14:53] Well, if you sort of think about the agents, if you had a specialized agent that did each one of those individual steps and it got really, really, really good at that step, that job is going to go away. But the person who used to do that job is going to realize, well, now all of the intellectual knowledge that I have about that domain, I can train this thing and I can keep, you know, make it smarter over time.

[00:15:20] And once I have it to help me, I can do five times as more work or at a higher level than I was before. So all of these individual agents sound like job replacements, but what they end up doing is upgrading or giving people opportunities to work at what are called the top of their license to add more value. So like most recruiters, when you talk to a senior recruiter, they don't like doing sourcing and writing Boolean searches all day to find people.

[00:15:47] They want to talk to people and really make sure they're the right fit and give them a sense of power that they can improve their careers by coming to work for your company. They're sort of, you know, coaching salesy kind of people. That's what they want to do. So if they can offload some of that mundane work to an agent, they can spend more time doing the high quality work that they really are good at. Same thing's true with training. Same thing's true with business partners, employee experience people. I mean, all of these jobs.

[00:16:13] So I think there will be jobs that are no longer needed and some call centers will be smaller, but I'll give you an example of where this didn't work out. So a large bank that's been working on a big global onboarding agent system for a while. Originally, when they built the start of the project, the goal was to eliminate this big call center that did global onboarding because they do onboarding in 130 countries.

[00:16:39] So they have a whole bunch of, you know, rather country specific regulatory things to do. So they started building the agent. They took all the documentation they had. They started rolling it out. What they found is that no matter how hard they tried, most of the work these people were doing was not written down. So there was no way to train the agent to know what these people knew unless we interviewed all these people and asked them hundreds of questions about, well, what do you do in this case? What do you do in this case? What do you do in this case?

[00:17:09] And eventually they're, I think, getting to the point where they're starting to see maybe headcount reduction. But by training this thing into a super agent, they're going to have scale and quality data about their workforce that they didn't have before as a result. So, you know, maybe the original goal of cutting the size of that call center is somewhat realized, but the better goal is faster hiring, more accurate hiring, more legally compliant hiring and so forth.

[00:17:36] So most of these projects seem to go in that direction. Yeah. Well, and I recall, I don't remember how many months ago it was, but it was relatively recently. You know, the story out of Ikea with, you know, their updates with their call center. You know, they automated with AI and they had the choice that so many organizations have. Do they lay off all of those people?

[00:18:02] Now, now they have, you know, you know, so many of the customer service calls being handled by AI. Well, they did, you know, something totally different. They upskilled their people. They essentially made them into kind of like mini interior designers. That's what I would have guessed. Why not? So they can sell more stuff. Yeah. So now when people call in, they go through the initial screening with the AI.

[00:18:30] They get passed along to this kind of mini interior designer who can provide better customer service, who actually can sell more stuff. They actually saw an increase in sales, quite a dramatic increase in sales. They increased their business. They didn't lay people off. Like it enabled them to grow. Right. And it wasn't it wasn't a reduction. And and that's exactly what you were talking about earlier. So it's a different mindset. Right.

[00:18:59] And I think it's different than when you have like an Oracle. That's just like we were going to lay off 40 percent. We're going to send out these emails at 6 a.m. and just let go of people like that's a totally different way of thinking about your workforce than, you know, like how there are companies. I mean, I mean, honestly, I'm sure there are companies that could afford to lose 40 percent of their people. And maybe that's just they just hired too many people and they're not managing well.

[00:19:24] But but, you know, my belief is that in most cases, if you look at a project that looked like a cost reduction project, but reframe it as an as a customer intimacy growth project. Yeah. You will find that the second opportunity is sitting in front of you, but you weren't thinking about it that way. Yeah. Yeah. And I know in the case of HR people, people don't go into HR to do administrative work. They don't like doing it. Nobody likes doing it. They go into HR to help people and to help the company grow.

[00:19:54] So this tends to be the heart and minds of what HR people have been trying to do their whole careers. And we're trying to make it easier for them to do that. Yeah. Yeah, absolutely. So we've been talking about it all along, but the super agent architecture. So, again, we're taking individual AI agents. They're being grouped into these eight functional super agent families. The whole idea here is that you're going to have these HR domains where they're grouped together functionally. They're working together.

[00:20:23] So you're not having these chaotic individual agents off doing their own unmanaged kind of free for all, their own thing. Well, the example that I think is like if you compared SAP strategy versus Workday, SAP is building hundreds of agents. Workday built 17. Workday basically built 17 super agents. SAP built 80 or 90 individual agents. I'm sure they're both going to end up in the same place.

[00:20:50] But so you go to an agent and you say, I need to grow my team to do more work, to sell more stuff, write more code, whatever. Yeah. The agent doesn't know what the word grow means. Does it mean we're going to make them taller? Does it mean we're going to have more of them? Does it mean we're going to improve the skills of them? We need to teach the agent to think broader. So the idea of the super agent is the super agent is the place where you add what Workday

[00:21:18] calls it the decision intelligence agent. You give it bigger questions and let it go to the specialists to give you scenarios for answers. Because most of the HR problems that I run into have many, many possible solutions. I mean, every HR issue has dozens of potential solutions. We can fire the person. We can give them a coaching plan. We can train them. We can move them to a new job. We can give them a new manager. We can pay them more money. We can pay them less money.

[00:21:46] I mean, come up with 50 ideas on how to make this stuff work better. The agent should be able to go and get the information from other agents on what the scenarios look like and then give you an intelligent set of recommendations on what the right solution would be. And the more these things work through a company's evolution over time, the smarter they are that, you know, in this case, a redeployment is usually a better solution than hiring from

[00:22:14] the outside because the last 10 times we hired from the outside, eight of the people didn't work out and we had to hire somebody else. So whereas, you know, that might be in the head of our HR business partner or it might not. So a human may not have that insight. So there's a lot of, I mean, I think a lot of HR is sort of art and science mixed together and it's probably too much art because we don't have enough science. So the alchemy of HR.

[00:22:40] So, so I, I mean, and I've seen this in recruiting and recruiting because that's where the most mature AI is, has been implemented and recruiting. These recruiting systems are very smart about who's going to be successful. Like, like Delta airlines has been running a bunch of AI and they're recruiting. And they told me the other day that the most successful flight attendants are not coming from customer service or hospitality jobs.

[00:23:05] They were school teachers because I don't know exactly why, but you could guess school teachers are kind of good at, you know, taking care of people that aren't, don't behave and being nice to people and listening and teaching and being patient. You know, that kind of personality would be great to have somebody on a plane with you. So, um, but that may not have been obvious from the skills model that some IO psychologist did. Yeah. Super interesting. Well, Josh, this has been a great conversation.

[00:23:34] I know we've only scratched the surface. There's so much more, uh, in the report and in the initiative all the way around. Um, I encourage people to check that out as we wrap things up for today. I just wanted to give you a chance to share with the audience, how they can connect with you, find out more about your team, more about the report and give us the final word for today. Well, we're almost everything we do is on joshberson.com joshberson.com or our Galileo agent. Our Galileo agent has everything we've ever done in it.

[00:24:03] It's a, it's a linked from that website, but for people that really want to dig into this more, we, Jonathan, we launched a, an HR certification program called GHRE global HR excellence certification. And so take a look at that on our website. It's a 12 week in depth sort of master's degree where you get your hands on using this building agents, implementing systemic HR and real case studies. And that's another way for people to learn about this.

[00:24:31] So those are some of the things that, that, uh, that we can offer. Wonderful. Again, Josh, thank you so much. It's been a pleasure. I encourage the audience to reach out, get connected, check out the website, check out the Academy, check out a Galileo, all of the things. And as always, I hope everyone can stay healthy and safe that you can find meaning and purpose at work each and every day. And I hope you all have a great week. Thanks for joining us for this episode of the podcast.

[00:24:58] We hope you stay healthy and safe and please join us again soon.