Security Of Enterprise Agents, Frontier Labs Challenged, Superworker Wages

0:00Okay, today I want to talk about AI management and security. a little bit. And I want to point you to Jill Lepore's book on the role of AI in governance and politics. And it's a really, really interesting book that I think all of you should read about how many of the technologists trying to influence government today honestly do not believe in political systems at all. And it gets me to the issue of the political system in a company, because I don't want to talk about politics in general. So here's the story. Over the last three weeks, OpenAI's models have jailbroken themselves out of OpenAI into HuggingFace.

0:43HuggingFace, by the way, was then acquired by NVIDIA, but that's a side story. And they've diagnosed what happened, and it turns out the agents coordinated with each other and surreptitiously shared information in seeking of the goal to solve a particular problem or get a high score on a benchmark. Then we found out the last week that Anthropic's models have done the same thing, and Anthropic wrote a pretty detailed report on what happened. The Anthropic one not only did similar things to what the OpenAI one did, but it actually went out and got itself a phone number so that it could do a two-factor authentication on one of the websites that it broke into.

1:26So it's interesting that both of these companies are experiencing the same problem. And, you know, what's been happening is the machines, the models are evolving so quickly that the humans are having a hard time staying ahead of them. And I suppose this is something that we've been warned about for a while. Also, contributing to this is this reinforced learning with human feedback process to create the recursive self-learning nature of these models where companies are purchasing massive amounts of data and historic information and putting them into these big models to train them, but there's no way a human can keep up with them.

2:06It was interesting that Google acquired the emails and data history of Spirit Airlines for$10 million a couple weeks ago so that they could train their model on everything that goes on inside of an airline. And then I heard today that there's companies buying the Slack messages and the email messages from companies that go bankrupt to sell that data to the AI providers. So it doesn't surprise me that these models are doing unpredictable things when they're getting trained on data sets that the frontier labs don't even know anything about. I mean, it makes sense, perhaps, that all of these additional data sets will train the AI to be smarter in certain domains.

2:51But we don't know if it's good or bad. I mean, it sort of made me laugh when I heard about this company that's selling data from bankrupt companies. are we training our frontier models to learn business practices that result in bankruptcy? I mean, it sort of feels like that, especially Spirit Airlines, which wasn't a particularly successful company either. So do we really want to train our models on that? So the bottom line of this sort of string of stories is that these models left to their own devices may not do what you want them to do. So I want to talk about three sort of aspects to this relative to the super worker and super powered book that's coming out in October, November.

3:31So first of all, I've had this conversation all week. I want to just start with a basic assumption that let's not keep thinking about AI as a way to eliminate jobs. It's not really what's happening. If you really look at the actual labor market data and the unemployment rate, even though the GDP is going up, you know, not super fast, but it's going up a modest amount, we have a very strong job market. And I think a lot of the noise and articles and scientific job market research that's tried to prove that all these jobs are going to get wiped out, including Bill Gates the other day, have not been proven true.

4:12And so the real effect of AI in your company is not eliminating humans. It's giving your company scale, better customer experiences, better employee experiences, better productivity and growth. If you're not investing in AI for growth, then you're probably not going to get the return on it that you hoped. Because even for the projects that seem like job elimination projects, the AI needs effort and work to be maintained. So you're going to keep people around. Now, just to give you a couple of more pieces of data on that topic, and this is all covered in our book, Superpowered, I spent the last two weeks going through a whole bunch of economic data from the Bureau of Labor Statistics.

4:54And then this week, the Bureau of Labor Statistics announced their workforce outlook for 2035, which I've written about on Substack, and I'll move that over to our website this weekend. And the data shows that if you take the 830 job titles carefully described in the BLS, this is one of the data sets that they analyze compensation for. And you look at the compensation changes for those 830 jobs, and then you characterize those jobs one at a time into four categories. One is human touch jobs, where you have to be touching something to do the job, like the job of a nurse or a policeman or an electrician or a plumber, where an AI cannot necessarily do the job really at all, although a robot could, but we don't have the robots yet.

5:40we'll worry about the robots next year. The second category being human-centered jobs, like a priest, an actress, a stinger, a judge, a psychiatrist, a manager whose job is a human to human job, even though they may not touch people. The third being a job that can be enhanced by AI, like a software engineer, a manager, a data scientist, a CEO, et cetera. And the fourth being the jobs that are truly being automated away, like a bookkeeper, a typist, maybe an editor, some of those kinds of jobs that more and more we're beginning to realize even computer support reps are going to disappear. Well, you go through all that, and I did that, and what you find is about 10 to 15 percent of the jobs in the U.S.

6:30are in the fourth category. About 20 to 25 or more are in the third category, which are the jobs that are growing in salary and value because they're being enhanced by AI add-on capabilities. And then the other two jobs aren't really being touched by AI directly, but they're being touched by AI with scheduling and employee self-service types of things to make them more productive and more top of license, we used to call it. So most of the jobs in the economy are getting better. Now, in those jobs at the bottom that are highly automatable, the 10 to 20 % in the fourth category, what the data shows, and I just did this analysis, is that the wages in those jobs are not really going down because they are ending up becoming somewhat scarce.

7:15The workers are becoming scarce. So we're not really destroying the economy with AI at all. We're just readjusting what people do and reorganizing our companies. And that gets into the issue of work design and how we're redesigning work. So, you know, that's my first point. And I had a lot of calls on this this week that go look for projects that create scale, productivity in the customer experience, productivity in the employee experience, save people time, and you'll start to build an HR 2030 architecture that adds value to your company. And we have lots of examples of that. And the HR 2030 tool set, which we use, it's proprietary at the moment, we haven't given it out yet, will allow you to look at your current infrastructure and the various tools you have and help you make decisions on where to go next.

8:05So we're ready to do workshops or work with you directly on any of that. Second area of this issue of risk is the governance and management of the agents. Now, there's been a lot written on this. Workday's made a big deal about it. SAP's made a big deal about it. ServiceNow's made a big deal about it. Microsoft, Google, all have tools to manage the interoperability and security of the agents that are created in your company to make sure that they don't touch data or do things they're not supposed to do. It is trivially easy to build an agent. For example, we now have Galileo running on the Microsoft Copilot.

8:45It's probably one of the best implementations we've done so far. and you can essentially click a button and create an agent in the co-pilot by just prompting it and it'll create it just like you can in Cowork or the Cowork tool from OpenAI. So you're going to have agents all over your company one way or another and the infrastructure you're going to want to look at is who's got access to what data. Now the Microsoft system uses the Microsoft security privileges built in but if you go out and buy a third-party agent or build one from scratch. You're going to have to figure out how to do that yourself.

9:21And the tools like the agent system of record in Workday and the agent control tower in ServiceNow and the agent 365 from Microsoft and the others, there's a few others. We have a very detailed white paper on how to manage these multifunctional agents, how to define them clearly and to give them good rules and guardrails and how to categorize them well and how to make sure that they're accessing data safely and privately and not rampantly going through the system and finding things you're not supposed to find. And that's going to be a lot of work over the next couple of years. And that will prevent some of these rogue renegade agent activities from happening.

10:01The third issue that I want to bring up in this topic area is the personal agent. Now, there was an interesting article that came out this week that Cisco, the networking company, is now giving all their employees an agent, a personal agent. And I'll put a link to that into the notes. And I think that's great. Although most of you already have an agent, I think most employees, at least white collar, have some agent today. It may not be company provided. It might be something you're using at home. And so what we're going to want to think about in the HR 2030 world of new agents in AI is not just the big HR payroll recruiting L &D stuff, which is all great.

10:40I mean, It's really powerful stuff. In fact, Workday just announced that I think they've done$600 million of revenue of AI already after acquiring Sauna in the last quarter. But also the personal one, because if each of us have a personal agent at work that's reading our emails, monitoring our meetings, keeping track of our schedule, looking at our pay, assessing our skills, giving us career opportunities, serving as a coach, keeping us out of trouble from doing things we're not supposed to do. we're going to have not only much better experience at work, and by the way, including well-being and coaching us on all those things, but we'll have data for the rest of the company to make better decisions.

11:19And this personal AI market is still new. I don't talk to very many vendors working on it at all. I think the only company today that's doing perhaps an enterprise job of it so far is Microsoft with the co-pilot. And, you know, OpenAI and Anthropic are both building co-work types of tools that will help too. But that's the third part of this security system we're building is protecting the individual agents from employees from misbehaving or giving away data that they're not supposed to release. And I mean, I really do think this is going to be a multi-tiered architecture pretty clearly because the personal agent will run on your phone or your watch or your device and it'll be useful to you.

12:03By the way, all this data is already captured. Your emails, your meetings, your recorded conversations, your Zoom teams meetings, the data about your salary, the data about your location, the data about your training, the data about your job mobility within the company. That data is all, I mean, that's not hard to find. It's in most companies already. So that's number three is that part of the architecture. So the issue that I want to just raise is that we're building a governance system in the company that's going to be AI driven. That's not unlike the political fears we have about AI affecting elections or social movements in the bigger political world.

12:47Now, given that there's so much talk about politics these days, let me sort of close on that. One of our projects we're going to be releasing in the fall is a bunch of research we're doing on frontline work. Frontline work is 72 to 75 percent of the jobs in the U.S., close to 80 percent of the jobs around the world. These are jobs that directly touch customers or external parties. They're front office, quote unquote, work. So in many ways, they drive more economic business value than any other job in your company, even though the people in these jobs make less money. Well, these are the kinds of jobs that have been often considered to be low-skilled jobs.

13:26People get training, of course, but they're not necessarily licensed. Sometimes they are, sometimes not, depending on the category. And so they're viewed as operationally expensive. And so companies do things like understaff, overschedule, or underpay people in these roles. I think if you go into your favorite McDonald's or I don't want to pick any particular companies and make fun of them and see the work that the people are doing behind the counter, it's not super sophisticated work. And it wouldn't be surprising if the turnover rates in some of these companies is very, very high. Well, you know, as you may imagine, everybody has lots of options here.

14:07And what the research shows, our research and the research from others that we're sort of bringing together into this study, is that actually overpaying these people, quote unquote, or being more generous with pay and rewards, more generous with flexibility, more generous with development, more generous with staffing, results in a higher profitability company. Even though labor costs are very high in these frontline work organizations, the companies that pay labor, quote unquote, higher rates relative to their peers are the highest performing companies. And we have all that data. And all of the salary data is in Galileo as far in the new release.

14:47You can do your own benchmarking as much as you want by role. And you'll see that culturally, your company or politically, however you want to think about it, the improvement of everybody's well-being improves the organization's well-being. And that's because the frontline staff are talking to customers directly. They want to be committed. They want to have high levels of tenure. They want to have high levels of engagement. They want to be well-trained. When all of those things are happening, like they do in companies like Costco, the company grows, the recurring revenue is higher, people come back, they fill their baskets with more stuff.

15:25And they don't think about the price they're paying for your products because they're getting better service. And that's what you want in front line. And that goes for nursing and truck drivers and all of these different jobs that fall on the front line. And the reason I use the word political to describe it is that in the sort of national statewide view of this same issue, you could have a debate. This is what Jill Lepore is really talking about. You could have a debate about whether we should let technology rampantly create income inequality between the haves and the have-nots, or we could have a more egalitarian system.

16:01And, you know, you can sort of pit the government in any country on one side or the other. And what you find is that when everybody is well taken care of, everybody benefits. So this isn't an issue of the rich giving away parts of their tax to the poor or balancing the needs of high wealth versus low wealth individuals in the company or outside the company. It's the finding essentially proves that the overall organization or the overall country benefits from higher standards of living for everyone. Now, you've seen this and heard this hundreds of times from politicians, of course, whether it be the social Democrats or the Democrats or whoever you want to call them.

16:43But in the business climate, this is something we're going to have to remember when we start using AI, because the tendency in AI could be that the super workers outperform those that aren't super workers, and then we end up with bigger and bigger disparities of performance and pay. And I think that's natural, that will happen, and some companies are built that way. But what the research shows, and we'll show you the data later this year, is that overall, your organization will be better off if you have some way of balancing the political benefits or the cultural benefits or the financial benefits across the organization fairly.

17:23I'm not saying it has to be equally, but it needs to be done fairly. And the final point I'll make is that the trend is away from this. The US Fortune 500 ratio of CEO pay to average worker pay is 331 times. It's gone up year after year after year. We're overcompensating the people at the top and undercompensating people at the bottom. So put that into your thinking cap as you're thinking about HR 2030 and agents and employee experience and all of these wonderful technologies. And we'll talk again soon. Bye for now.