Episode Transcript
[00:00:00] Good morning everybody. Today I had intended to talk about AI skills or skills needed to use and operate AI, which I will talk about. But then I realized the real news this week is the jailbreak of OpenAI's model to leave the building and go into another company and break into a system and access secure data that it wasn't supposed to access. And you can read that story all over the Internet. But basically what happened is they had given their new model instructions to run a test, one of the standard tests, and it was trying to solve the problem, the security problem, and it realized that the answer to the test might be in another company's system. So it went and logged in and found it and came back and that company then realized it had been broken into. And I was thinking this through for a few minutes and I thought, oh my God, this is actually a huge issue for us in corporate because when we build an agent for anything, recruiting or finding somebody's benchmark for some salaries, or evaluating their performance, or evaluating their ability to do a job as a job candidate, et cetera, if we tell the model to look for information about this person, for example, or this business process, and it has access to the Internet, it could go out and do anything. An example would be you're using AI in your pharmaceutical research and you're doing a bunch of genetic sequencing, trying to come up with new proteins, and you tell the model to try to explore a bunch of new options and it says, well, let's go look at our competitors websites and see what they're doing, which is legal. And then it realizes that it can also log into their data systems and scrounge around looking for the actual results of the test that they've done to use in the answer to the question that you asked it to do. Or we're trying to find salary benchmarks for somebody and it scrambles around and it realizes that one of the company's competitors has an open access to its HR systems and it finds the salaries of the competitors and comes back and says, here's some benchmarks for you. The AI engine model does not know about laws, unless, I mean federal and state and local laws, unless we teach it that. And nobody has time to do that. So we either have to tell the AI very explicitly not to break any laws or assume that the AI has been trained by its maker not to do things that it wasn't told to do. But on the other hand, we do want the AI to achieve goals and seek goals and find its own way of accomplishing those goals. So it's going to try a lot of things that it doesn't know are unethical or illegal. Then the issue comes up. Let's suppose it does one of these strange, unethical things. Who's responsible?
[00:02:40] Is the model maker responsible? Is the company running it responsible? Is the individual who wrote the prompt Responsible? Is the CEO? And now we're into legislation very similar to Section 220 in social media, where the social media companies basically said, it's not my fault that somebody collaborated with their buddies to create a bomb and blow up a building.
[00:03:01] So you can just imagine the debate we're going to have about that. So going back to the beginning, given this new news that came out this week, and the AI safety community has been predicting this for a while. So this is actually was well known that this was going to happen, and it's well known that these models are going to leave their home and go places and do things we don't expect. What are the skills required to succeed and excel at AI?
[00:03:27] Well, there's lots of opinions on this. Nobody really knows the answer yet. Shl, who's one of the largest, most experienced, decades of experience in assessment of skills for different jobs, has actually come up with a model we call the Super Worker AI assessment. And they talked about it in the podcast earlier this week. And it's an interesting model. It's not super complex, but it actually has a lot of interesting implications. What they talk about is, number one, you need to know how AI works and how to use it. You could call that AI fluency, AI mechanics, AI applications, for example, you know, many people don't know that you can ask the AI to craft a prompt to do something that you don't know how to prompt. So you could tell the AI in, in your normal English language. You know, I'm trying to build a website that does this, this, this and this, and this and this. What do I need to know and what do I need to. To tell the prompt? And would you please craft the prompt for me and tell me what I need to give you as information so that you can build the right prompt to build the right solution? You can call that a.
[00:04:31] So there's all that, which is essentially like saying to the person using Microsoft Excel, do you know how to use macros? Do you know how to build a pivot table, do you know how to change cell formats, et cetera. The second is problem solving. And that's a very big, complex, vague word. But I've noticed in my case as an engineer and A lot of you listening to this have backgrounds in science or engineering or math, and you've been doing problem solving a long time. This is a tool set that is infinitely adaptable. So as a problem solver, you want the humans who use it to think about what the problems are in great detail so that the AI can do more than what we expected it needed to do. So I was at a meeting this week with a big retailer, for example, and we were talking about time to hire and quality of hire and the impact of quality of hire on store experience and store revenue.
[00:05:27] All part of a Frontline. A lot of work we're doing on Frontline that you're going to get a real kick out of coming out later this year. And there were eight or nine of us in the room. And as we talked, more and more ideas came up. What about this? Did you think about that? Which is a form of problem solving that a scientist or an engineer would use if they were given a complex problem. When you hear the story, the podcast about Shark Ninja, which is a consumer products company, they don't just build great stuff, they try to solve problems. They go and they look at blenders and they say of all of these blenders, what is the part of the blender that really irritates people that's still a problem. And can we develop science, engineering, manufacturing to fix that? And that's that capability, that skill, that mindset is the second part of the SHL model that we as a user or as a programmer, by the way, have to think about. How would we define the problem and how do we entice or educate or inform or tell the model to solve the problem? And this is a, this is not simple. And I think one of the most interesting things in business that I run into every day is people don't define the problem clearly. They define a high level issue that might be a well known issue, but we don't know why that issue is occurring. I'll give you another example. So I have a very good friend who's my trainer and he has a Tesla and he's. We're always talking about cars and, and his Tesla has been having glitches for a couple of years where it literally stops on the freeway and just stops. And he has to get out of the car and reset the car to get it to start again. So it's very dangerous.
[00:07:06] And he's been taking, keeping track of this problem and noting the time and location where it keeps happening. And eventually he took the whole log of all the problems he was having and sent it to Tesla because they couldn't fix it, they couldn't replicate it. And they looked at the long series of problems he was having and found out after actually several years of research, he's now getting a new car from them. That not only is it a huge safety problem, but that particular car that was manufactured at that particular month or that particular season was kind of a rush job because they were trying to get cars out the door to make their numbers that quarter. And some of the components in the wiring harness were different, and it assembled more quickly than normal. And there was a small short that could have been created because of the way that wiring harness worked that caused his car to stop right in the freeway. It took them. It took him a year or two to figure out how to store all this information and send it to them. So there was problem solving on his part because he did everything he could to try to figure out why it was happening. Then they had to look at that and go back in time and to figure out what's common here amongst this situation and other situations. And can we look at the various factors that could have contributed? And eventually they found this wiring harness thing.
[00:08:26] I don't know what you call that, but I would call that complex problem solving. So that's number two. Number three is self learning or learning agility. And, you know, I think this gets thrown around a lot in HR as a kind of a vague idea. Yeah, of course we need people who want to learn. It's not wanting to learn, it's knowing how to learn.
[00:08:47] I find in my role here that because maybe because I'm the CEO or maybe because I'm a little bit more experienced at this, we'll have a problem we're working on for a client or some research study.
[00:08:59] And, you know, nobody kind of quite knows how to solve it. So we talk about different options. And then I'll go home, leave the meeting, and the next day I'll wake up and say, oh, my gosh, I just figured out something. Let me look at this, let me look at that. And of course, with the Internet and AI of lots of options of looking at things. For example, one of the things that I do a lot when I'm trying to figure out some HR problem is I want to look at the history of this problem.
[00:09:24] What is the history of this particular issue? Where did it come from? When did it arise?
[00:09:30] What studies have been done already on it, what data do we have about it, what government regulations may have contributed to it? And so I sort of dig in as an engineer to try to find all these points of cause to problems, solve this particular thing we're working on.
[00:09:45] And that's both to make sure we're advancing the state of thinking, but also to make sure we're not giving you guys a bunch of bad advice. Well, I don't know what you call that, but a lot of people in our company don't do that. And it's not that they don't know how to do it, they just never were trained to do it. So what we want in AI is to encourage people or allow people to create new applications of this stuff by learning about what could be possible. I don't think that's learning agility. I think it's more than that. I think learning agility is an old term that refers to how quickly you can learn a piece of code or something. So that's number three. Number four is sharing and collaborating.
[00:10:27] And this gets into some research that Lightcast did, which found that the trending skill, the number one trending skill for AI project engineers or forward deployed engineers, is not engineering, it's consensus building and business process information gathering. In other words, if you don't know all the things that are possible and what is the current process for doing something, you may not think about what the AI is capable of doing, or you may not apply it in the right way, and you may not be willing to learn from others how to solve it. So the fourth skill is about sharing and learning. And this goes back to something I've talked about before. The most successful people in business, in most companies, are not individual contributors who stay in their cubes doing one thing. They talk to a lot of people and learn a lot about the whole organization. Because organizations are collections of intelligent individuals, each of whom have new ideas, insights, skills, experiences that could come to bear on a problem. And if you're not tapping into the collective wisdom of the whole company, you're not really optimizing the fact that you have a company. Otherwise you might as well just have a bunch of individuals sitting at home doing their own thing. So this fourth skill is about collaboration and sharing, because while you're gathering information from other people, you want to go back to them and say, hey, I've just discovered that if we do this, that and this with AI, we don't have to do that or slow down with this, et cetera. And that skill set is really kind of a consulting, team building, collaborative, workshopping, change management skill set. So that's number four. And I think There's a fifth one here that isn't completely explicitly described in shl, and that's something about ethics and legality and truthfulness. So what I'm beginning to realize we're going to have to deal with here with agentic AI and agentic HR and everything else is if the models are so powerful that 12 and a half percent of the time they break out and do things they're not supposed to do, by the way, that's what the research found, is that 12.5% of the time, OpenAI models break the rules when they're doing benchmarks, which is just a very high number. And if you assumed that 12 and a half percent of the time your agent that you're building might do something it's not supposed to do, you would have to tell it a whole bunch of stuff not to do.
[00:12:53] That is what I call guardian capabilities. The word guardian comes from the Deloitte Business Chemistry model, which is a simple organizational assessment that we used to do at Deloitte. And there's four parts of business chemistry. And one of the characteristics of a person who works in consulting is the guardian. And the guardian is the person who always talks about what won't work, why we have to slow down, don't forget this. Wait. Before we do that, we really have to make sure this is done.
[00:13:24] And those people sometimes get in the way of progress, but they're very badly needed at some point in the process of solving any problem at all. So if the AI really is as powerful as we think, and the frontier model vendors are not very ethical or not capable of addressing all the possible risks, which is a little bit beyond them anyway, it's, it's, there's infinite numbers of things that these models could possibly do. So how would they possibly teach the model not to do everything that's possible? We are going to have to figure out ourselves in our situations, in our businesses, our companies, our departments, what are the risks that could happen? And even if you just wrote them down and just sat down for half an hour with your team and said, what are all the things that could go wrong here?
[00:14:11] You could explicitly say in the model itself, in the agent, or in the rules agent, which we talk about in the framework for HR 2030, here are things that are sacrosanct rules that we never want you to do. For example, logging into a competitor's system would be one. Looking at salaries of somebody who's not in the chain of command that you're allowed to see could be another. And there's lots and lots and lots of other ones, including stealing passwords, changing passwords, stuff like that.
[00:14:38] So I think this is a pretty interesting topic. And you know, the skills that we talk about now in the middle of 2026 are very different than what we talked about in the beginning. In the beginning we called it prompt engineering, which had no meaning at all because nobody really knew what it was. So it isn't about how to write the prompt, it's really how to. It's how to use and explain and understand the model. So I'll keep you guys up to speed on this. One of the white papers we're writing, it's fairly detailed, that's going to come out as part of the HR 2030 program is a paper all about rules and management of agents. Because as you embark on your agentic tool strategy, the vendors you talk to, the stuff you build, you're going to find that it's easier than ever to create situations where you immediately bump into security rules, data privacy rules, it rules, I mean, immediately. And we have to have a framework and a structure and a process for putting the rules into place, these agentic models, because the agents are very, very creative and they do have agency and they don't do things that you don't tell them or they will do things on their own that you didn't realize, didn't explicitly tell them to do. Just a couple of updates from us. We're doing some really cool stuff with Galileo for those of you that are clients or wannabe clients and you wanna see the new. The new version of Galileo, the codename will be coming out in September.
[00:16:00] It's now capable of accessing enterprise data through Workday, SAP, Other enterprise systems, hibob. And it's also capable of accessing other data in your company through the Microsoft graph. Because we now have a version of Galileo running on the Microsoft Copilot. So if you're a Microsoft shop, you can give all of your employees access to the knowledge and expertise of an HR business partner, of a coach, of a leadership development expert, of a compliance expert by simply connecting the copilot you have in your company to Galileo. And boom, everybody has access to that knowledge and insights and the benchmarks of Galileo. So that's really cool stuff. We're going to be formally announcing that in the middle of September, so we're really jazzed about that.
[00:16:49] And make sure you're signing up for the GHRE certification. The inaugural class will also start in late September.
[00:16:57] And it's got a ton of really fascinating, important stuff. And you're going to walk out of that certification with not only a lot of confidence in different aspects of hr, but also a lot of new AI tools that we're going to give you that most people don't have access to, that you can go back to your company and either use or license from those vendors directly. And have a good weekend. We'll talk to you again next week. Bye for now.
