#352 AI Agents at Work: What Actually Breaks (and How to Fix It) with Danielle Crop, WNS

It's totally worth it to upskill yourself. People do need to just experiment within some guardrails and have some fun with these new tools. Like, I've been what I would call a pragmatic tech optimist.

AI agents suddenly got suspiciously good, and that means you probably urgently need to reinvent half your business. If you work at a large enterprise, that's a daunting change management task. How do you make sure that your AI strategy feeds into your business strategy?

Once a resource becomes more, becomes cheaper and more productive, the need for it goes up. Are those agents actually providing value or are they not?

Fortunately, we have Danielle Krop to guide us. Danielle is Executive VP of Digital Strategy and Alliances at WMS, which means she spends all day helping customers run transformation programs to make better use of data and AI. Previously, Danielle was the first Chief Data Officer at American Express, where she built out a data capability from scratch. I'm curious as to how she'd approach that now.

Who wants OpenAI to have access to all of their Slack data? It's a probabilistic model. It's never going to be perfect. It's going to do unexpected things, which is that, you know, it will hallucinate. You still have to use your brain.

Let's find out how to reinvent everything. Hi, Danielle. Welcome to the show.

Hi, Richie. How are you?

I'm feeling great, actually. Almost the weekend. Looking forward to this. Although not for everyone listening because it's coming out on a Monday. Yeah, let us kick off. So if you are in an industry that is overrun by agents, what should you do?

Oh, overrun by agents. So I think that first of all, you have to ask a few questions, right? One is, are those agents actually providing value? Right? Or are they not? Two is, are those agents in some sort of secured environment, right? So that they're not creating undue security issues for your organization. But the other aspect of it, I'll just have fun with it. Like, I think this is a moment in which people do need to just experiment within some guardrails and have some fun with these new tools and see what they can build with them. Because I think we were just at that that tipping point between infrastructure and software in this world, right? Um, and you know, we'll see what comes out of it. It's going to be a lot of fun to see what people— like, I mean, I think, oh, you know, you had to be living under a rock if you didn't see what OpenClaw did, right? Like, um, and it's just really fun to see what creative and interesting things people are doing, um, with this technology.

I do love the idea of just have fun with it. Um, so if you're interested in having some fun, like, what's a place to get started?

I would say wherever you want. Like, so the, the best, the easiest place to like get familiar with it, um, if your organization is relatively uncomfortable with it, is to just get your own consumer version, right, of the tools. Um, and just play. Like, it's like, I know that there's some trepidation about like, okay, well it's gonna be like some monthly cost. It's totally worth it to upskill yourself. And learn the art of the possible about what these things can do. So whatever version, like, you know, Anthropic, OpenAI, whatever version you wanna use, Gemini, just get familiar with the tools, understand what's possible with them, figure out how to do maybe even more creative things in just your day-to-day life. Like, one of the things that I built was just this agent to just do competitive analysis. So ring-fenced it about around, you know, um, our competitors' earnings statements, earnings calls, transcripts, etc., and just did a compare and contrast between, you know, what our company is reporting and doing and saying and what they're doing and saying, so that as a, as a strategy executive, I can think about like, okay, where are the gaps? Where are the opportunities? Where are we? You know, um, and it just is a lovely way to, you know, to use those tools. On a side note though, you have to be, you still have to use your brain. So, which is that, you know, it will hallucinate. In my case, this agent actually hallucinated and said one of our major competitors bought a company we bought, even though it was ring-fenced and completely ragged and only primary source information. It still hallucinated in a very major way. So. While it's a great tool, you have to like be, you know, you have to like, it doesn't take the place of, you know, knowing and being a critical thinker and going, yeah, that doesn't sound right, I gotta dig into this further.

That's very cool. I do think competitive analysis is one of those things that no one really quite has time for in general to do properly, so that's a good thing to outsource. But yeah, it seems like you've got to have two mindsets then. So you've got to have this sort of naive, childlike innocence to be like, okay, I'm going to try new things. But also you need to be very critical in terms of, you need cynicism about the output of the AI because sometimes it does make things up. Do you have a sense of like how you go about balancing these things and like how you approach your mindset?

So I think that, you know, for me, and personally, I've always been the type of person who likes new technology, right? So I never, I never, and I think that the genie's out of the bottle, guys. Like, you gotta embrace it, right? Like, it's, and it's like, but I've always been that kind of, okay, what can I do with this? I was a very early adopter of like online banking and online, right? Like I was just one of those, like, whatever can make my life easier, faster, more convenient, um, I'm all for. And so I've always been that way my whole life. Um, you know, and, uh, so I just keep embracing that and I would say that it's it stood me in good stead, right, throughout my career and throughout my life that like I've been what I would call a pragmatic tech optimist, right? Like, so it's like, it's not that I think that technology can solve all the problems of the world. It cannot. Technology is a tool. And when human beings get better, greater tools, we can do better, greater things, but only if we choose to do so.

Yeah, I mean, there's certainly a case for you've got to choose to solve the right problems. Otherwise, you know, you can cause a lot of problems with AI. But maybe we'll focus on doing better things for today, or at least for now. So there are lots of— we were talking about agents, and there are lots of different types of agents. So some of them is like, okay, you've got an existing workflow, you're going to throw a bit of AI magic in there. And there are some companies that are claiming they've got AI that are going to replace entire employees. Do you want to talk me through like how you decide what sort of AI you should be using?

So, I think that you should, I mean, as far as like what AI you should be using, it very much depends on your industry, your regulated, you know, how heavily regulated you are, how much risk you're able to take. So, your risk appetite as an organization is number one. And then you can decide a little bit more about what, you know, which direction do you, like if you need, the more and more I see things, particularly with like the Open Claw scenario recently, the more I like feel that, the unlocking of a lot of this opportunity, the early opportunity, right? Of like just automating major parts of jobs. And I don't think it, I don't think it automates the full job, right? Because there's always that element of like, okay, now I have to evaluate, did it actually do what I wanted it to do, right? Is it doing the right thing, right? Like that's, LLMs are math on, but they're qualitative outputs. Right? And so you have to— a human being has to evaluate a qualitative output. It's like— and even in a mathematical output, you still have to like check and make sure that the computer does the mathematical. So this is not new. We have to keep checking things. Um, so I'd say that what tools you use, if you want to be more— if you want to keep all your data basically in-house, I think that the frameworks will evolve to be able to allow that. So basically you'll be able to have the tools, entirely within your ecosystem and manage all your— because like, who wants OpenAI to have access to all of their Slack data, right? I don't think anybody does. But the reality is that there are some organizations that are currently like, you know, feeding, you know, an agent all of their emails for as an org. Like the smaller organizations are figuring this out much faster than the larger organizations, right? Feeding all their emails. All their Slack, all their— so basically they have kind of like this agent that can tell them everything that's going on in their organization and do a summary of it at any given point in time, which is really incredible, right? And then they can do like— think about like the, the, what that enables in terms of strategic decision making and timely strategic decision making. It's pretty incredible. So, I think that that's where it's like, okay, you lean into the tools, but you don't, you can't, just like we, you know, let's face it, we don't all fully trust our human colleagues, okay? So, we can't fully trust our agent colleagues. And that's the same mindset that we have to bring to it.

Absolutely. I love the idea of just checking absolutely everything, 'cause you know, things can, you get the wrong answer whether it comes from AI or from a human. Now, you've mentioned OpenClaw a couple times. So, just for anyone who has not been following this story, just talk them through like, what is OpenClaw and Why might you want it and, and how can it go wrong?

Well, I think that the, you know, like where I discovered OpenClaw mostly was in like the All In podcast, right? Like they were talk, they talked about it. Um, and it was, they explained kind of what I just explained about like what they're creating with it. And the market reacted very strongly, right? So it basically said like, okay, SaaS is dead. Services are dead, like all, like all, like because of what, you know, this one guy created with OpenClaw. Okay, so he used the existing, you know, capabilities and just created this amazing new, you know, framework to bring things together in, in an agentic fashion. You have to be really trust it. You have to be willing to give it to, you know, access to all of your systems and access to all your, right? And it's like, you know, so do you wanna trust it that much as an organization? You have to decide. But it showed that there's a lot of possibilities. So in this particular case, like the, the, there was a AI, they created an AI producer for a podcast actually. So you might wanna look into it, Richie, if you haven't seen it already. But basically goes out and it like prepares, you know, lists of like who should be on the show next and why and what they're putting on LinkedIn and all those kinds of things. Sends emails to the person, sends, right? Like automatically does all this stuff to like, you know, create the, create the pipeline for the next podcast. And so like, does it automate most of what a producer does? Yeah. Does in the end, do, are you actually gonna get more quality people on your show that if you don't actually vet what it gives you back? Probably not. Right. But it is gonna automate a lot of the mundane tasks that nobody wants to do. So that's— there's nothing wrong with that. And think about for sales. Sales is going to be like all those things, all those places. It's going to make— it's going to make a huge difference in what people are able to— capable of doing.

Yeah, I mean, lots of, uh, interesting sort of, um, things to think about there. So I mean, I guess, uh, the idea is you just feed like everything you're doing into an agent, and it's kind of designed— it basically works like a personal assistant for you.

You give it your login for LinkedIn, you give it your logged in— like, you give it all of the identity, like you give it a separate identity. So that's a huge issue that's coming up is like, how do you do that? How do you control it? How do you— right? Um, but you do, you have to, you know, you do that and then you let it go.

There've been a few disaster stories with OpenClaw already. So there was an example of, um, someone using like a GitHub version, uh, like an OpenClaw version of their GitHub persona. So it was, uh, being mean to other people who are making pull requests. There was an example of it deleting someone's inbox. This was like a Meta security researcher as well. So is especially ironic. So there are some dangers here. Do you have a sense of how you do this sort of thing safely?

I think you have to be very clear on what you ask it to do for you and how you ask it to do. So this goes back to like, okay, well, you have to give it very specific commands, right? And very specific instructions. And some people are not like, so even about like tone, like, you know, you can constrict it and say like, I only want you to use this type of tone with people, right? Like, so it doesn't do those kinds of things, right? That doesn't mean it won't still do that, but you can contain the risk by giving it very specific instructions on what you want it to do, right? So that it's as much like you as it could possibly be, but it's still not, it's still not gonna, it's never gonna be perfect. It's a probabilistic model. It's never gonna be perfect. It's gonna do unexpected things. There's an error rate. So as long as you're comfortable with what that error rate is, or you can build small language models that like make that error rate smaller in the context that you want. So there's different ways of handling that problem, I guess, is what I'm— is like, what, what, you know, what model do you use? What, you know, what tools do you use? What, you know, what servers do you— there's a whole stack of how you set, like, how you handle those problems. Yes.

Only think about like all the infrastructure, the infrastructure and technology. Those are big decisions, but I like the idea of just thinking about like, if you have something that's interacting with other people and you think about the tone of the output. So actually recently, DataCamp's head of data science created a bot to review other people's dashboards, just to make sure they work okay. It was like, wow, this bot is really mean. It's supposed to be talking like me and it's horrible. It's being incredibly critical and not in a nice way about people's work. So that required a bit of tuning. You do need to think about this stuff.

There's a lot of, you know, tuning that's already been done. Like, it's like, it's actually quite amazing. You just, you know, that's another one of like, play with it, ask it to do something and then ask it to give it to you in different tones. Like it will, it's actually amazingly accurate. You give it very like simple instructions, like make this cheekier. It actually understands what cheekier is and gives you something that's cheeky. It's kind of amazing. Like, how does it? So there's been a lot of great work that's been done. By the AI researchers in tuning this, and that in particular in the area of tone. Why they focused a lot of time and attention on that, I'm not exactly sure, but it's kind of fun to play with. Like, you can tell it to be really mean, you can tell it to be super, super nice, you can tell it to be verbose, you can tell it to be concise, you can tell it like, you can give it all. It's very, very interesting.

Absolutely. And certainly if you're using something, if you're creating AI for making social posts or for marketing materials, then the tone, getting that writing is going to be absolutely essential.

Like, well, we still do. Everybody still has brand guidelines, right? So I just anticipate there's going to be like a brand guideline for your AI tones.

Absolutely. Yeah. I mean, like everywhere I've worked, there's been like an extensive brand guidelines document that no one can quite bother to read. So it's good. At least the AI can read it. It's not going to complain about that. And hopefully it'll stick to the rules. The other aspect of this was around like, what do you feed into your agents? So, presumably, I mean, you mentioned like sending Slack messages into some AI, it's a little bit risky. How do you go about deciding what AI, what, uh, what data the AI gets?

So I would, I mean, it's like, I'm going to be a bit of a redundant here, right? It's really about what your risk tolerance is again, right? Like, so if your risk tolerance is high, feed a lot of it in and then dial it back. Or feed a lot of it in, in a proof of concept, in a ring-fenced environment, and then tweak it from there and decide, you know, but you want, until you do it, you won't know, you won't really know what you want to feed it or what you don't want to feed it. You have to like, you have to play with it and learn. And then, yeah, it's 'cause I think to some extent, the risk tolerance has been, like, there's been two ends of the spectrum in my experience. One is like the people who have like, don't even consider the risk at all. And then the other end of the spectrum are the people who don't even try it because they think that it's— and it's like there's some happy balance in the middle. But I would like, going back to that, what you said earlier, which is like being very positive and happy about, but then skeptical, be skeptical as well. Like, like, so, but you've got to try it out because we won't know all of the risks until we, until we try some of them in a controlled environment. right? Similar to that example of like, oh, look, it said our competitors acquired this company when we did. Like, you won't know that that's the kind of things that it will do until it does it. And then you're like, okay, well, now how do I— now what do I do about that? Right? So, that's where I would just once again tell people, you gotta, you gotta, you gotta learn the tool. It's not— it's, it's an— it's a— you have to, you must.

Absolutely. It does seem essential to get started, but also you maybe don't want to be gung-ho and try stuff without thinking about what can go wrong. So, you need some kind of culture of iteration, experimentation, testing. I mean, you've managed many departments. How do you build that culture into your team or department?

So, I built the culture by rewarding curiosity, rewarding creativity, rewarding, right? Like, so, and demonstrating it myself. Right. So that's, I think those are really key things to, you know, like, whether it was, you know, back in digital transformation or any of the data when we moved from one data, it's like, I mean, I've been through like 4 major data transformations in my career, right? So every single time you had to demonstrate to your team that like, yeah, you're gonna, you know, you're gonna move from SAS to SQL and you're gonna be okay. And so here's how you learn. And so, but as a leader demonstrating the skills and using it yourself, And reminding your, and like using it yourself so that you can be the art of the possible for your team. So in this case, it's like, you know, I have to remind my team, a strategy team that creates content for a lot of different organizations, like you have a copywriter in your back pocket, like it help you, right? Like it's like, it's, it's let it help you, right? It's, and, and see what it does. Um, so just reminding them where they can use it and how they can use it, but then, and then like unleashing them in, you know, controlled ways. I think is the most important thing, and definitely doing it yourself, leading by example.

I do love that idea of leading by example because it's one thing to tell your team to do stuff, but if you're actively doing this and showing them that, then they're going to be more motivated to do that. Do you ever set any targets around AI usage?

No, not at this point. We will see. Like, I think that we're all going to— we're all very fast coming to a point in which it's extremely expensive. Right? So I think that that's where people need to like, that token cost is non-trivial, right? Hopefully over time that will go down because of, you know, like as we get, you know, more energy, hopefully cheaper energy, et cetera, you know, then we'll, you have token cost and more data centers and more like the, the cost will come down over time like it does for every general purpose technology. But for right now you have to be, you know, so that's one of the other aspects I would tell organizations to be careful, the ones that are just letting people go play. You have to, you have to watch your token cost, right? Like, and I think there's going to be more of that, you know, just restricting how many tokens any individual, individual employee can use. It's going to, you know, I'm sure there's probably places that are already doing it today. My particular organization is not, but it's fast approaching. So I don't know if that entirely answers your question, but that's kind of what I'm thinking about right now is like, okay, yeah, I get them to use it. You have to get them access. But I think once they have access and they know like, oh, how, how, how helpful it can be to everything that they're doing, you end up with like, okay, well now we have to like constrain the costs because we're new. Like, you just can't have it run wild, right? Like it's the same thing as like, um, in the beginning of PCs, right? Like not everybody had a PC when they first rolled out, right? Like they rolled out over time as, as the technology came down in cost. So I do think there's an aspect of this of like, Okay, how do you, how do you make those decisions? How do you strategically make those decisions of who has access to this to begin with? How do you constrain the costs? It's a little different than a PC because you don't have to buy a piece of hardware to stick on everyone's desk, right? It is, it is software. So yeah, anyway, go ahead.

No, absolutely. Yeah. Um, so cost is definitely something to think about. Okay, we'll find a, uh, like a pro subscription if somebody's, uh, chat tools is like, what, $20? But then if you're doing like, uh, coding stuff, you can burn through thousands of dollars of tokens in a month. So it certainly can get expensive. Uh, you mentioned strategic decisions. Like, uh, talk me through how do these get made? Like, who needs to be involved in deciding AI strategy?

So I think the area that needs to be involved in AI strategy is— this is a little bit of a different shift from like a traditional— so for example, like, I'm in a strategy role, but I came from a chief data officer background, right? There's a reason for that, right? Like, at this moment in time, my leadership said that like what we need is somebody who really understands this ecosystem to help us make business strategy decisions. And so I do think that that is more and more important is like, okay, having that view of like, like as stacks are converging, right? That's what you're seeing in the market, right? Is that like all this market angst about like SaaS and we're seeing the stacks converging across cloud, across SaaS, across, you know, the, the native AI companies. Across the Databricks, the Snowflakes, like we're seeing all of these companies basically converging and trying to build the same feature sets. And that's an interesting point in time, but you have to be able to see that. And I think that unless you have both the business savvy as well as like the data and tech savvy and are able to connect the dots across, you're not going to be able to, to make the strategic decisions of who should you partner with and why. Right. For example, am I like— am I like, who are we like, are we going to partner with a major SaaS provider to do agents? Are we just simply going to go directly to OpenAI or Anthropic? Like, those are very important strategic decisions for the future of the company. And what are our clients? What do our clients want? Right. Like, what is the stack that they're going towards? And I think everybody's converging into this more simplified most likely stack, which is like, I mean, companies only want to have to play with companies they have to play with, right? Because it creates a very complex ecosystem for them to manage. As a chief data officer, I know this well, right? Like, you just don't want the ecosystem of partnerships to be more complex than it absolutely needs to be. And I think what we're seeing is like, do you need AgentForce and OpenAI Frontier? I think this is a still, it's an open question, but we'll, we'll see what happens. But it's like, I don't, I don't think the, I don't, I think the answer is no, you will not need both of them in the longer term. You will just, but you do need the LLMs regardless. So there's one player you have to, like, you have to play with the clouds. You have to play it right. There's things you have to play with and there's things you don't have to play with. And that's, I think, what we're seeing in the marketplace at the moment.

Ah, that's interesting. So, um, I suppose that's been a big stock market story in the last few months is that a lot of these SaaS providers, so you mentioned Agentforce from Salesforce, and I guess there's a few others like ServiceNow where they have these huge SaaS companies, but they've also—

I do think it's an overreaction. Okay, so I'm like not saying that, like, I think there is an overreaction in all of that, but I do think there is a real threat that needs to be You know, 'cause the, the moat in this world is the data. And those SaaS companies actually don't own the data. Services company can, service companies like the one that I work for can help companies navigate and go, okay, which direction should we go? And what, you know, how should we integrate with like these different, right? But like the SaaS, the SaaS providers themselves, they're, they're, it's an interesting moment for them. We'll see how they navigate through it.

Absolutely. I mean, some great products, but yeah, they're being sort of a bit— they're struggling in the market at the moment because there is maybe a worry about like, do you need these SaaS products long term if you've just got an LLM in the nation?

And if you must have the LLMs and you must have the data, and do you want the data to remain your own? Right. And it is in the SaaS environments today, but like you have this weird situation with, you know, how do you, how do you navigate all of that? How do you make sure that your protecting your data, your proprietary data as your moat, because it is. And you don't want to make that just free and available for, you know, OpenAI or Anthropic, right? Like, because then, like, that's, that's a problem, right? You don't want your strategic decks to be available and queryable by anyone who's using OpenAI, right? Who's using ChatGPT. So like the more, like, that's going to need to be ringfenced, kind of like how it is with SaaS, but. You know, it's an interesting moment. I do not have the answer for this one, but I think it'll be fun to see what happens.

Yeah, I mean, part of the fun is the chaos at the moment, I think. So, we were talking about strategy, and so you mentioned like about simplifying tech stacks. I guess the other big question is how do you make sure that your AI strategy feeds into your business strategy?

Business doesn't change, right? Like, so it's not, you still have, so this is the advice I give to, to clients. Your business still has all the same core business capabilities that it needs to do to make money today as it did 5 years ago, right? In many cases, unless they're creating new products and new lines of business, right? And so think about what those bus— business capabilities that you need are, the core business capabilities that you, that you need are, um, and how you currently drive, you know, top line or bottom line through those capabilities. And then determine which AI use cases you're going to go after based on the size of that opportunity, right? It's like, and it's going to be different in different verticals. It's going to be like, that's, that's okay, but it doesn't change the fact that you have to start with the P&L Right? Like it's, it's still business. You're still doing business, right? Like it's, and so I think that's where people get a little lost of like, okay, well, we've had a lot of bot, like I, I, I do love the bottoms up stuff, right? Sometimes it's really interesting, but like you have to, we're like in that moment of like, okay, you need the bottoms up because you need people to learn what the case these tools are capable of. But you also need the top down of like, we need to focus on the things that actually really drive our business. And not the things that just make your job easier, right? Like, so it's a little bit of that, you know, and strategically I think companies have to make some decisions about like what's going back to who has access, how are we making, how are we holding them accountable for what they're doing with it, and how are we, you know, making sure the costs stay low, right? But that's a, you know, that's a strategy, senior strategy decision to go, Who should have this? Why should they have it? What business capabilities are they gonna be working on? What is the business value? It doesn't— business case is still essential. Like, it doesn't— so I feel like people have lost a little bit of the, the plot of like, you still have to have a business case. That doesn't mean that you should not do this even where you don't have a business case because it's like the PC revolution, um, general purpose technology. You can't just say like, oh, because I don't have a clear business case, I'm not going to do this at all. That's not a good strategy either at this moment. So, it's like balance, right? You should have a business case. Try to find— you probably do if you dig in and you get some really good people in your org on it. You'll find those business cases and then deploy it. But if you can't find that business case, like for whatever reason quickly, you still need to deploy this. Like it's not, you don't, And I was like, I, like, I, I recently made the analogy of between like the, the companies that were like, let's just take FedEx, right? FedEx was born out of the PC revolution, right? Like the whole business model was built around the new technology and it killed, right? Like it did amazingly, like in the '90s, right? It's like, it was like, I don't, I'm old enough to remember this, right? Like I'm old enough to remember like FedEx really hot stock. And like, but this is because— and Walmart, for example, Walmart, like, you know, took the PC and took the, you know, distributed computing and made, made a whole moat around their business about it. They— and they— but they still are a retailer. They just knew that— they, they knew how to use the tools better than their competitors. Sears saw it as a line item. They're dead. Walmart saw it as a moat. They're trained. And like, that's like, people, like, companies really need to think that way and make the technology tool available because if we, you know, like, you don't want to be Sears.

Okay, so you can think of AI as being a cost, like you start worrying too much about the token costs and think, we've got to try and minimize spend here and you're probably not going to benefit. But if you think about how can we harness the power of AI, then that's going to, give you new business opportunities and hopefully greater growth. I guess that's the gist of the argument.

It's striking the balance, right? You can't like have a completely like risk-averse, I'm not going to even try this, I'm not going to engage in the new world. But you also can't look at it as just like a line item. So because that's not, you know, if you look at, we had the same arguments during the PC revolution, right? Like, okay, it's going to, we had literally the same arguments if you go back in time and look at it. We had, you know, the Harvard Business Review saying, you can see, you know, the PC and all the numbers, but not in productivity. We know how much the PC drove productivity in the '90s, but when you were actually trying to do finance accounting in a particular company, you couldn't find it. That doesn't mean that we shouldn't have rolled out PCs, right? So, it's that balance of like, but should a company have bought so much PCs that it drove them out of business? No, like it's a balance.

Yeah. I mean, you gotta buy a lot of PCs to drive you outta business, I think. But yeah.

But might your cost to be the point where you like, you can't make a profit anymore, right? So it's that, you know, it's that balance that, you know, a company always has to strike regardless of what era. And it's just like, I think people get a very, they get very nervous when a new general purpose technology comes out and what does this mean and how it's, And it's normal angst. It's historically very normal angst. Whether it was the, but you wanna go back and look at the railroad. You wanna go back and look at the telephone. You wanna go back, any of these types of things, they cause these and you'll see it in the, like the news articles. You'll see it in the, like the, the expert journals. You'll see it's, it's the same. And so I find that incredibly reassuring. This follows a pattern. Human nature stays the same. We still all need to make a profit in our businesses. So there are some things to just anchor yourselves around so that you're making good decisions.

I love that phrase, historically normal angst. I think if you're anxious, you think this is new to me. This is like an incredibly personal thing. This never happened before. And that's why you're anxious. But actually, if you put it in the context of history and this has happened before, it's okay. Take a breath. You can relax and move on and do nice things with your life.

And try to have some fun. Right? Like, we're all right? 'Cause it's like, if you just get, I mean, like, I mean, I recently went on the record, I completely disagree with the idea that this is gonna take jobs.

Oh, well, tell me about this then. If it's not gonna take jobs, how is the interaction between AI and careers gonna be?

So, I mean, like any general purpose technology, right? Like, it's that drives productivity. Like, that I, it's going to create, like, so in the short term, it will have disruption. I'm not saying that it won't. So, it's like, you think about, like, I like to use the analogy again of the PC, right? Like, we had a lot of typists that lost their jobs, and there was even conversations— like, I found this article that was quite hilarious that was talking about, um, the discussion in the boardroom and the C-suite about who would do the typing in the PC world. Like, it's kind of fun, it's kind of reassuring and hilarious, right? Like, okay, they did have these. So that whole Like, yes, jobs will be, jobs will shift. And this has happened, by the way, very quietly in the background during the, you know, the initial, what I would say, AI era that most people didn't see, right? It didn't hit the news because it was happening quietly in the background of companies. So companies that were building machine learning models to make better fraud decisions, et cetera, were displacing a ton of people who were answering the phones for fraud calls. Like, whole organizations, like, were let go that were just answering the phones because we were able to do model-based risk decisioning and routing in a new way. So, overall point, this has happened before. It happens all the time. This might, I mean, I don't, I don't know if the, everybody seems to think this is going to happen faster. I'm— if you see what's happening right now, not so sure it's going to happen faster, right? Like, in the smaller brand new startup organizations, yeah, it's going to happen lightning fast. But in the existing enterprise organizations, based on the risk, based on the change management issues, based on like, is it going to happen like overnight? I don't think so, right? So, I think that there's a lot of angst about, okay, it's going It's because it's taking, it's being more productive, we'll need less people. But what's happened the last time that the, the times that that happens where people say, oh, we're gonna have more productivity, so we need less people. Well, those people's, those people move from here to here, right? They'd start doing different things. They don't go away, they just start doing different things, new things, new opportunities that arise. So going back to like FedEx, right? Like there was a new opportunity in shipping and logistics because of the PC, right? Like it didn't, like, you know, in the '80s and before, it was all USPS. No, like that's how you sent stuff, right? FedEx created an entirely new market, competitor to it. We'll have stuff like that. We just don't know what it is yet.

Okay, I love the optimism. And on your point about what happened to all the typists, I think if you go back and watch like TV shows from the '60s or '70s, like every middle manager had their own secretary. And now it's like, you've got to be, you got to be C-suite before you get your own executive assistant. So I think all these admin operations type jobs, there's been a long trend of the last 50 years of technology kind of making this build more efficient and reducing need for that sort of labor.

Absolutely. And think about how much easier it is to book travel today than it was, you know, 20, 30 years ago.

Absolutely. You have to go to a travel agent and then they take their cut of whatever and And yeah, now it's just like, oh, on my phone, easy peasy. One of the big stories around careers over the last year or two has been around the junior hiring crisis. Now, it seems like for people who are just coming out of university, it is more difficult to get a job in the last year or two. So, do you see an effect where it's different for people who are just starting their careers versus mid-career or late career?

I think because of the moment that we're in right now, it's harder. I don't anticipate that lasting. So once again, I anticipate that like, as we realize the Jevons paradox, which is that, you know, once a resource becomes more, becomes cheaper and more productive, the, the need for it goes up, right? The demand for it goes up. I think we will have, I think entry-level folks that have the right skillset and the right mindset, more importantly, will do just fine. Just curious, creative, self-starters, like all these are all things we looked for before. We look for them now. That also hasn't changed, right? Like it's, it's yes. Do I, do I think that we made that there's one major shift that I see between the way that we educated in the past and the way we should educate, I think, in the future because of the nature of these tools? We spend a lot of time and money in the West in particular, right? I think to some extent catching up with the East on STEM education. Do I think that people, high schoolers in the future, should just be focusing on STEM? And no, no, do not. Because think about some of the things we've talked about today. I am a statistician, yes, but I talked about history, I talked economics, I talked about business, I talked about like, you have to be the type of person in the future who's just curious and interested in all sorts of things, then you will connect the dots and, and move the business forward. Um, and that was true before, it's true now, it'll be true in the future. But do we, do I think that, you know, will we, you know, will we see possibly really long-term 40-year careers as a software developer? Maybe, maybe not, right? So I would say don't be a specialist. Like, that's not like, learn the tools, but you don't need to be a deep subject matter expert in the future. The models will have that covered. But what you do need to be is the type of person who can imagine the future with those tools.

Okay, so I love the idea of, I mean, you said that if you're curious, if you're a self-starter, if you've got a good imagination, these are all things that are going to help you get hired. Do you have any other ideas about skills that are going to help you? That will help you in your career?

So I would say that, like, so, like, I'll just take, like, you know, kind of like the, the thought process I have for my own 15-year-old daughter, which is I'm trying to educate her broadly, right? So that she can have relatively moderate level understanding of a lot of different things, right? And then understand how to use the tools when she needs to understand something more deeply, right? But by having that, like, that level, you have to have, like, this level of understanding across a lot of different things so that when the LLM, when the, when the, when it returns something to you that just, like, makes no sense, right, you can, you can have that critical thinking and go, that's not right. Like, you have to have enough knowledge to be able to call BS on the tool, right? Or to just know that actually this isn't, this tool, an LLM actually isn't the right tool. I need a small language model for this because it's got way higher risk, right? So being able to understand the risk tolerance of the world, like, do I think that a, do I think that a, that a, any sort of bot in the future, whether it be a robot with LLMs on it and world models and all, like, do I ever, do I think that they will ever be able to understand, like, say what, when Darwin discovered, you know, the theory of evolution, right? Are any of them gonna be able to go like, okay, now I'm gonna go and challenge the entire world on their existing ideas? No. Only humans will be able to do that.

Um, yeah, I guess we hope. If there's some sort of AI revolution challenging humans on our ideas, that's gonna be tricky.

But the math doesn't support them being able to create novel new ideas. They'll make stuff up that we can then decide, like, is that right or is that wrong? And should we, like, should we take that and move that forward, right? But do they actually create novel new ideas and then make decisions about that impact the world and go and argue with policymakers in Washington or whatever. Like, no, they don't do that today. I don't anticipate they will ever do that. I could be wrong. We'll see. But like, I just, it's like, it's just, it's a tool. Like, I don't see it any different than any other technology tool than we've, that we've ever created as a species. It's, it's the same thing. So I refuse to, I refuse to think that I wrote, I wrote something about this too. I'm like, Terminator's not coming. This is AGI is not a thing. Okay, maybe I'll, maybe somebody in 5 years will be like, Danielle, you were totally wrong.

We have this conversation from a bunker. Yeah, no, there do seem to be a lot of companies who are trying to build bits of Terminator though, which I do find quite terrifying.

Oh yeah, yeah, yeah, yeah. There's, that's not, that's not yet what it, like the idea that Skynet's gonna become sentient and like try to kill us all, like that's like, is that gonna happen? Do— am I more concerned about concentration of power? Right? Absolutely. So if you have concentration of human power, whether that be like— and in this moment in time, concentration of power would be concentration of compute. Okay? So if you have a few people who have access to the compute, that creates a very unstable situation for all of us. Right? So that's kind of like, we need democratization of the tools, we need democratization of compute, we need like, that's what will keep us from any sort of Terminator-esque scenario in my mind, right? Like, it's just about, it goes back to like the basic understanding of freedom. Like, if you have a concentration of critical resource amongst just one group, Like, think about if just Saudi Arabia had access to oil and none of the rest of the— wouldn't Saudi Arabia dictate the entire global policy for everything? Yes, they would. So you, you— if any sort, any critical resource, in this case, can be like, you know, the more data centers we have, the more distributed they are around the world, the more freedom we will have, the less control. Like, it just— that's what— that's, I think, what I— that's in the Terminator scenario. That's what I think about. Like, I go, what am I worried about? I'm not worried about the models. I'm worried about concentration of compute.

Yeah. And that does seem to be a big thing because there are a very small number of companies that are now spending tens of billions of dollars on, I guess, hoarding compute or in some cases building more compute. But yeah, there is that concentration there. So, as was at least one of the big topics of the last few months was around sovereign AI and the idea that you need to kind of own some of your tech stack in order to remain independent. Do you have any position on this? Like how, how should individuals or organizations worry about like controlling, uh, their AI stack?

I think it depends on what you mean by sovereign AI, right? So like, do I think that individual countries, so like, and I've had conversations, like it's like people have different, just like when we first, and even to this day, we have like, we have lots of different opinions on what agentic is. We have a little, we, we have the same thing with sovereign AI. Okay, so does sovereign AI mean you own the compute in your own data center and everything's all, or does it mean that you have your country in defense of itself has its own AI models and its own infrastructure? That second thing is inevitable, right? Like obviously countries are going to invest in AI to defend themselves. That's not even a question. Right? Which means they're gonna have their own infrastructure, their own compute for those particular defense applications. Does that mean that they're gonna then create some sort of like sovereign AI that takes over social, political, other aspects of society? I don't, like, I don't know. Like I would, we see this with China, right? China has done this. It is not a very pretty picture for those of us who like freedom, so, or privacy, right? Like either one. If you don't, if you're— it's a very— it's not a pretty picture, right? If the Chinese Communist Party doesn't like you, you don't get to transact on certain things. You just— it's, it's very dystopian from my perspective. So, um, like we've seen that. Is that what people want? Like, I think that's the decision that each individual, you know, democratic country is going to have to make. Like, is like, are you gonna push your policymakers to keep these things free and independent? Are you gonna let them create, you know, centers of power that are really not good for society? And I think that's just a, that's just something people have to contend with at this moment in time. Okay.

Yeah. I mean, certainly some, uh, incredibly like, uh, important decisions on the societal level about the interaction between government and technology providers and I guess, uh, all the citizens, uh, as well. All right, that was a fascinating digression. We somehow got sidetracked from, we were talking about skills and careers. So I would like to talk a little bit.

It's all connected, Richie.

Everything's connected. Yeah, I would like to talk a little bit about what happens to data teams, 'cause you've been a chief data officer. So with the advent of like AI and particularly agents that can do a lot of data analysis, Does it change how you approach building a data team?

Yes. So, I think that there's still this core, you have core functions that you need to build as a data officer, right? Like the most, almost all, I mean, there's a little bit different depending on where the data officer reports to the organization. But the core things are, you know, data governance, data management, data products, data science. Data monetization. Those are usually the common 5 pillars to build for a chief data officer when they start a new job. Like, okay, well, do I have these? Do I not have these? So, I think that how you build all of this, you have to really think in terms of now structured and unstructured data, right? As, so that's one, one pillar, that's one, one functional thing that's different in the last 3 years that you didn't really have to think that much about for, right? From a data governance and a data management perspective, how are you going to manage, like you've got every, most chief data officers, like we all understand, oh, structured data, how do you manage structured? Like now it's more like, okay, now you're like, how do you manage unstructured data? How do you create that provenance for it? In the same way that we used to do data lineage, but it's different, it's just different. So how do you do that? And what tools, providers, capabilities that you can use to do those things are different than they were 3, 5 years ago, right? So how you organize your team can be different. Like, how do you, you know, what do you want to do in data governance? Do you really want to do it? Like, data governance is going to be a very important but very fast-evolving space in my mind, right? And I think that there's going to be a lot of tools that are going to be built out of this agentic era that we're in. That are going to make it a lot easier to do data governance, a lot faster, a lot, because data governance was something a company just never wanted to invest in. Right. And now I think you don't have to invest nearly as much in it, which I think is good, but you still have to have those people who know how to do it. It doesn't change the fact that you have to have people who are skilled in what does data governance look like, what does— like, you still have to people have people that are skilled with that. They just have better tools to do their job.

Yeah, certainly data governance tooling has evolved a lot over the last few years. It is dramatically easier than it used to be. But it's interesting that, you know, the big worries around unstructured data. So I suppose this mostly means like the documentation of how your business runs. I guess there's also going to be— do you seem to care about images and video as well?

Yeah, I think absolutely. Like, how do you know? I mean, like, we're all dealing with this at a societal level, right? Like, how do you know the— like, I use the word provenance. Right? Because I think it's the most apt one. I know it's not like the one that most people necessarily use, but I think it's the most apt for the moment. Is like, does this have history to it? Does this have reliability to it? Does— do we know where it came from, right? Just like you would do like for an antique, right? Like it has a provenance, right? So data has a provenance, and like historians already know how to do this stuff very well, right? Like how do you critically evaluate source material and know that it's is actually real and know that it, right? And that same kind of thought process, I think, has to come to any unstructured data. So, for, not all books written in the 18th century are equal, right? Some of them are good and some of them are not. And how are you judging that? It's the same thing with unstructured data. Where did it come from? Who authored it? What, right? What's, how, all those things matter. It just, they matter at a greater scale than ever before. And so, how do you create, you know, the right type of mindset and tools around that?

Fascinating that you mentioned historians knowing how to do this. I think like philosophy degrees, often considered like not great for careers, but actually now every company needs philosophers for their responsibility AI teams. Do you need historians on your data governance team now?

I think that. You know, perhaps like that might be a good way of like inculcating that thought process and mindset into a data governance team. I think that they would understand. I mean, I think, you know, librarians, like very similar type of like, how do you structure things? How do you organize things? How do you like— I think there's just these, these, there's these professions that we can learn a lot from that are more, that have, that have maybe not been the shiny professions in the last 20 or 30 years. But I think that we need to go back and look at them and go, what did they have? What did— what do they know that we don't know? And how do we organize things in a better and more structured way for the future so that when we're making decisions, like, they will— they— the LLMs won't hallucinate as much if you're feeding them good data. Like, if you're feeding them things that are true and then it goes to the societal aspect of it all, we all need to stop— like, I, you know, we all need to stop feeding social media with things that aren't true. Like, it's like we, we have our own personal responsibility that we can take in this world at this point in time as well, which is like, stop reposting the thing that you didn't check before you reposted it. Like, that— like, there's just— it's, it's a fascinating space.

Absolutely. I mean, certainly, uh, I like that idea as a A call to action is don't repost silly things on social media.

If you haven't checked the source that it came from and what their agenda, motivation, financing, you know, like if you haven't checked, don't repost, particularly if it's a very sensitive, highly political topic. Like just don't, just You can read it, you can do it, but don't like, let's not propagate things that we don't know are true. Like let's all try to remember that truth matters.

Yes, I love that. A little bit of fact-checking goes down very well. All right, so just before we wrap up, what are you most excited about in the world of AI at the moment? We talked about a lot of things here, but what's really getting you going?

I think that my, the most, like, and I've said this a couple times already, the most exciting thing for me is to see what people will do with this, right? Like, it's like, you know, it's like I'm having fun doing things with it, right? Like, I'm curious to see what people much smarter than me will, will do with these tools and what kind of things that they will create, what type of app store we will have, right, using these tools that we don't have today. Like, I, I'm really in— I'm like, I think that in the next A lot of wealth has been generated in the infrastructure phase. More wealth is always generated in the next phase. And I'm really curious and interested and see what happens. And that's why I want all the young people who may be or maybe not listening to this podcast to be like positive. And there's a lot of opportunity. Go play with this, have fun with it, and you might discover the next huge, you know, the next unicorn. You don't know, right, unless you go play with it. Um, and so I think that that's like the fun. And it's easier now for anybody who's a creative or a historian or library philosopher to do this than ever before in history, right? Like, before, they would have been like, oh, I have to go, I have to go get my 4-year CS degree. And like, so I'm interested to see what those people who have a very different mindset and a very different way of approaching problems in the world what they're gonna do with these tools, because that's a part of humanity we haven't unlocked in this technological era. So, I'm really excited to see what happens with all that.

Absolutely. I mean, yeah, it's not even young people. It's like, it's absolutely everyone. Like, we've got hundreds of millions of people who are now capable of creating stuff. So, yeah, certainly exciting times. I'm with you there that this is gonna be a big thing. All right. And just to finish up, I always want more people to learn from. So, whose work are you most excited about right now?

So I generally like, so the way I, I go about it's like I have to keep a very, you know, external lens on like what is going on, right? So I generally like to do the podcasts and I've already mentioned one, but the Dwarkesh podcast is a really good one. So if people aren't doing the Dwarkesh, I, I would, you know, and it gives you the high-level perspective of what all the big guys are saying, right? Like, you know, Dario. Elon, Sam, like, they're all like— it's like— and, and just other huge, you know, Yann LeCun, like, just other, you know, people who are real thought leaders in this space and pushing the envelope. And, and so keeping on top of what, what they're thinking, what they're doing, and just asking yourself, though— like, for me, I ask myself the question, like, do I agree with them? Do I not agree with them? Like, I often argue with them in my own writings and posts, um, but it's, uh, but it's a way of keeping on top of what's going on. And so, I'd say that like podcasts are one of the easiest way to like get deep into a topic, like spend 2 hours, right? But not have to like, you know, like for somebody like me, it's like, yes, I was in the modeling world, et cetera. Do I need to know the different, the exact difference in the technology between Claude 4 and Claude 4.5? No, I do not, right? Like, so, I'm not in the AI research space. So, if you're in that space, I love you, Okay, I need to, but I have, I have my own chief AI researcher. We have a, you know, one at WNSC, which I, who I just adore. But like, I spend my time more on the strategy level of like, what's going on, what's important, what people are talking about at this moment in time. And so I find the Dwarkesh podcast to be essential.

Okay. Yeah, maybe my second favorite data podcast or AI podcast after DataFrame. Yeah, I'm definitely all for listening to podcasts in order to learn things. Very good idea. All right, super. Thank you so much for your time, Danielle.

Thank you, Richie Cotton. Bye.