Episode Transcript
[00:00:03] Okay, guys, today I want to talk about the next big wave of AI stuff, which is called agentic AI. It's a weird word, but the really big idea is for these AI large language models to now be able to take actions on our behalf. And so they move from, in a sense, large language models to large action models. What that basically means is that the AI systems and chatbots that we're starting to use in all sorts of applications in our consumer lives, and our companies are going to get hooked up to other systems, and they're going to be able to send transactions, send emails, and access and activate and use other systems in the company, including sending messages, composing emails, creating documents, etcetera. Now, there's many, many, many applications of this, but the way to think about it is the agent becomes like an employee, and you can ask it to do things, and the agent will have potentially a list or a workflow of activities that it will actually do on your behalf. Let me just mention two of them that I've talked to vendors about, and they're going to be hundreds more.
[00:01:27] If you're a learning and development professional, for example, and you have to build a course on a new topic or maybe do a video or some other kind of instructional activity, there's a bunch of steps that have to be involved. First you have to get the content and perhaps interview the subject matter experts that contribute to this content. Then you have to create the course or object or video that your employees are going to consume. Then you have to launch it. Well, first you have to review it with the people that are responsible to make sure that it's accurate. Then you have to publish it into your learning management system or whatever system you use for training. You have to launch it. Send communication emails and messages to people so they know that it exists.
[00:02:13] Then you have to continuously track it and manage it and analyze it to see if they're actually taking it and if it has an impact.
[00:02:21] I wrote a whole book on this, actually called the training measurement book. A lot of that can be done by an agent. The agent can send emails to the subject matter experts, click them into a link so that they can author or simply record an audio or video recording of the information that they believe should be in the course. You can upload documents to the agent. The agent can consolidate the audios or videos from the subject matter experts, which is called needs analysis. By the way, to build the instructional material, the agent can have some instructions on how long to make the course and how rigorous to make the course, and the audience and the reading level and other aspects of the course, what languages to translate it into. The agent can then build the course, by the way, can do that automatically. You can tell the agent if you want the course to be a mobile activity in audio, video, or a traditionally learning course, and then it can publish it into the LMS and start tracking it and send emails to everybody in the system that's supposed to get it and on and on and on. And all of those steps would have had to be done by hand, if you even remember to do them all. And so that is a relatively repeatable, mechanical set of steps. While each one of them are important, that can very easily be done by an AI agent. And we can call that agent our training manager, our training assistant, our training teammate, or you can give it a human name, whatever you want. And there are going to be products just like that. And as you'll read about in the article I just published today, companies are selling these agents. Now, in some respect, what's happening is the whole idea of software as a service is now becoming services driven by software. Because when you buy one of these agents, you're not really buying an end product, you're really buying a service, and you're training the agent and enabling it and tuning it and essentially coaching it to do what you want. So, interestingly enough, Sarah Franklin, who's the new CEO of Lattice, introduced a feature of Lattice, which is an HR platform to manage these agents as if they were employees on the chart. I and a whole bunch of people like Sherm and other people kind of insulted this idea, and they thought it was really disrespectful. But actually, I don't think it's disrespectful at all. I think this is real, and I think this is really going to happen. And we're going to have these things in our organizations doing really valuable work. Now, just like any other employee or real teammate, the agent needs to be onboarded. The agent needs to know your culture and your rules and your language and your products and your services. They need to know your organization and who reports to who. Because if they're going to send emails, they need to know who to send them to and who not to send them to. You're going to have to, in a sense, coach the agent. So if it's building a course, you're probably going to want to have it review the course so you can edit it or tweak it in the process. And so in some sense, we're going to be embarking on a new job, which is to train and manage our agent talent.
[00:05:45] And that sounds so silly when I say it, but it's actually kind of a good analogy, and these agents are going to do a lot of cool things. A second example that I want to point out is one that you're all very aware of is in recruiting. In recruiting, there are just a lot of really complex steps that are involved in finding and assessing and hiring somebody. First, someone has to write a job description or at least a concept of what the job is. You have to give it a level, you have to give it a title. You have to write a description so people who are applying for the job know what it is. You have to list the skills that are required for the job. The agent could actually get that information from the line managers or a group of people in the company, just like I talked about earlier on the l and D example, through videos or audios, the agent, the AI agent, could then build the job description and the set of skills and assign a level to this job.
[00:06:47] The agent could then ask the manager or hiring manager to approve or edit this job. Once the job description is finalized, the agent could create a job rec. The agent could post it on LinkedIn or other indeed, or other services, advertising services. The agent could watch the incoming candidates. The agent could look in the ats of existing candidates that have been maybe silver medaled in the past and they're available for this. The agent could go through the list of people in the company that have related jobs and look at their skills. Then the agent could consolidate this into a list of candidates. The agent could look through their locations, their credentials, their job histories, their salary histories, and give you a short list of the candidates that you're interested in. You could go back to the agent and say, okay, I want to go after these three people. The agent could send those people video interviews that they could do asynchronously with a set of questions that are even generated by the generative job description. The answers could come back. The agent could look at the answers and the agent could actually figure out which one's the most highly qualified candidate. And there's lots of AI systems that already do that. They're called interview intelligence systems. The agent could then take that short list of candidates, run a background check on them, do a sourcing check, look at their social media posts and some of the other information about their credentials, check their education, send an email to their universities to get their educational credentials, put that portfolio together and send it to the hiring manager. Now there's obviously manual human steps involved in that, but you can see it wouldn't be that hard. And I just did this in the last couple of hours to design that workflow and implement it in your company if you had a good intelligent agent, and that's going on all the time in companies. I have watched recruiters doing these steps manually and a lot of times they'll outsource them because it's actually just a lot of work that isn't very value add. Recruiters want to spend more time on the human part of this and basically what the recruiter could end up doing is after all this other work is done, they could get on the phone with the top two or three people and really validate that they're kind of the right culture, fit for the company, for the organization, for the job, for the team.
[00:09:06] I could go on and on and on. There's going to be lots and lots of these. Now if you read the article that I just published today, what you'll see is first of all, there are startups building all sorts of agents like this, sales development rep agents, customer service agents, you know, sales agents, iT security agents and so forth. And some of them you don't even know they exist because you interacted with them as a consumer. Those companies are building these on these generative systems, so they're training them in the use cases that they've decided to go after. And they're building the inequations so that those particular domain agents have AI technical interfaces into Salesforce, into Zendesk, into workday, into whatever it may be. So what these agents are going to do, these new digital employees you could call them, is they're going to sit in front of our IT systems and they're going to change the architecture of the enterprise landscape. Now as you'll hear about in my speeches I'm giving over the next couple of weeks on the road in the hr tech world, the average company has about 90 to 95 employee application systems. And I'm not kidding, these are large companies, small companies have eight to ten. And each one of these systems, corporate systems applications, has data. Data about your employees, data about their training, data about their work, data about their jobs. Salesforce has data about salespeople. Zendesk has data about customer service people. GitHub has data about your engineers that if they're software engineers, etcetera, etcetera, etcetera. So what these agents can do, and by the way, the way we usually deal with these heterogeneous data systems is we buy a complicated piece of software that sits in front of all that.
[00:11:00] We go to servicenow or we go to workday or somebody else, or we use visir and we build either an analytics tool or a portal, and then we spend all sorts of money on sharepoint or other consultants to make it easy to find things in these systems and bring the data together into different applications. Well, once you have an intelligent agent that is hooked up in a sense and has the metadata about these backend systems, that becomes the interface. And as Satya and Adela mentioned really almost two years ago, the bot becomes the application. Now, I know all of you love to log into workday and poke around and look for the right screen or HubSpot or Salesforce or whatever it is that you use. Nobody really gets that much out of doing that experience. If I can ask the agent to get that for me. That in and of itself changes the technology landscape. And so not only is this going to change the jobs and the work and the activities we do in our own functional domains of HR and other aspects of our jobs, but this is going to change the stack of tools we have in our technology landscape. And for those of you that are involved in HR technology architectures, and I'm sure a lot of you are, because this is a massive, massive part of HR, this is a very, very different way of thinking about employee experience.
[00:12:24] Now, I'm not the first one to think of this. This is a well discussed topic out there in the tech world. ServiceNow has built a whole framework for the now assist to be able to participate in some of these transactions. But not everybody can do everything. The l and D agent is very unique to L and D. The recruiting agent is very unique to recruiting. The agent that goes out and does employee surveys and asks people questions about their employee agent is another type of agent. Maybe you could call it the IO psychology agent who likes to talk to people about their work experience. By the way, Galileo is going to become one of these and it will be sort of your Hrtaine business partner agent. And these things are going to get very sophisticated in very different domains and then they're going to talk to each other. Because one of the architectural aspects to this agentic AI is not just agents talking to systems, but agents talking to each other. Interestingly enough, the Microsoft copilot, which is really attempting to be the agent of agents in a way that's certainly the vision that Microsoft has. If you read through the Microsoft documentation, we talked to them about this and we've been doing some testing of hook and Galileo up to the copilot. And the interesting thing that happens is once you get that plugin to work, the copilot and the Galileo agent say things to each other that we don't know what they're talking about. I mean, they're kind of asking each other questions and getting answers, but that's not necessarily visible to us unless we instrument the system to see it. So, you know, a world of multiple agents acting on our behalf, acting on the behalf of employees, is going to need some management. It's going to need some governance. And of course, each one of these agents is going to need training to make sure that the things it's doing are getting smarter and smarter and better and better, not worse and worse, and hallucinating and making mistakes and irritating people with things that are incorrect or even worse. Giving the wrong people access to confidential information that they shouldn't have. The reason I'm really, really high on this is actually for many reasons. One is, I'm just an enthusiast, but we've talked to a lot of companies about it. I've talked to lots of vendors, and I have seen these things working, and they are actually here now. They're not all in the market yet, but they're coming very, very quickly. And we had a fascinating conversation three or four weeks ago with a large healthcare company in the northwest that has built one of these agents for their nurses and their employees. And they built it on an old technology stack from IBM. So it's a little bit before the days of LLMs, and they're moving it into more of an LLM architecture. But what they said happened was once the employees became comfortable that this HR employee, agent, or at Chatbot, was the place to go for information, all of the things that the HR department did, whether it was a news alert, I, a compliance program, an organization change, a training program, some new structure, or some new initiative the company is doing all goes through the agent because it's a way to reach people in a very centralized way. You know, how complicated it is to build a mobile app or a. Or any other kind of communication interface that really works all the time. And by the way, nobody wants to get emails all day. These things are very, very powerful ways of delivering information when people need it in the form they want it, and giving them the ability to access information, get work done, and do things much faster and much more efficient than ever before. Now, interestingly enough, based on where we are in the stock market and the slowing economy, there's this other weird economic thing going on. And that is a lot of financial analysts are getting worried that the cost of the electricity, the Nvidia chips, which are very, very expensive, the software platforms from OpenAI or Microsoft or Google, is that this stuff is too expensive. And I talked to a vendor two or three weeks ago that's building a really cool new system I'll tell you about at the conference. And they love the system. It's very, very powerful. But they said the problem is it costs $50 to run it every time we run it for one employee, so it's too expensive. So there's this ongoing debate from Goldman Sachs and many other investors that the ROI, quote unquote, of AI has yet to come. And at some point, people are going to stop spending all this money on infrastructure with the anticipation that some great outcome will come sometime in the future. Not everybody's like Elon Musk and just spend a billion dollars for kicks and see what happens. Well, I think Agentic AI is going to unlock that, because once these agents are configured to do real work and provide real solutions, not just answer questions to interesting queries that we have, we will be able to measure the productivity improvement of these systems. And to give you an example of how far along this is going, there was an announcement today by, I believe, Zendesk, who basically said, look, because of the cost of delivering the support agent in the Zendesk platform, we are not going to charge you for it unless it adds value. So rather than charge you a license fee for something that you may or may not use and may or may not have value, you will get charged based on the quality of the answers you get. So in some sense, that's like buying a consulting service. If you buy a consulting service and the consultants sit around all day in the conference room and drink coffee, you're not going to pay them. You're only going to pay them if they get the work done. Well, that's where we're going to go with agents, too. We're going to be able to measure the work, we're going to be able to instrument the work, and we're going to be able to capture the ROI of all of this massive infrastructure investment that's going on in AI. So there's a little bit of education for you guys this weekend. I will tell you a lot about the vendors and the space at the HR technology conference in Vegas in a couple of weeks. I'll be at unleash, I'll be at workday rising, and we'll be announcing some really exciting things along these lines for Galileo late in September as well. If you would like to learn more, read the article I just put on the website, or give us a call or sign up for the AI and HR course in the Josh Gerson Academy. Thanks everybody. Have a great weekend.
