Why AI is Failing: Ex-Google Chief Cassie Kozyrkov Debunks "AI-first"

Is “AI-first” the future of business or just another tech buzzword? On this episode of Digital Disruption, we’re joined by former Google Chief Decision Scientist and CEO of Kozyr, Cassie Kozyrkov. Cassie is best known for founding the field of Decision Intelligence and serving as Google’s first Chief Decision Scientist, where she helped lead the company’s AI-first transformation. A sought-after advisor and keynote speaker, Cassie has guided organizations including Gucci, NASA, Meta, Spotify, Salesforce, and GSK on AI strategy. She combines deep technical expertise with theater-trained charisma to make complex concepts engaging and actionable for executive and general audiences alike delighting audiences in over 40 countries across all seven continents, including stages at the UN, WEF, Web Summit, and SXSW. Cassie sits down with Geoff to unpack the hidden cost of the “AI-first” hype, the dangers of AI infrastructure debt, and why real AI readiness starts with people, not technology. She reveals how leaders can architect their organizations for innovation, build human-in-the-loop systems, and create cultures that embrace experimentation instead of fearing mistakes. Cassie exposes why 95% of organizations fail to achieve measurable ROI from AI and how leaders can finally bridge the AI value gap. This conversation dives into why AI success isn’t about tools, it’s about leadership, measurement, and mindset. Most organizations chasing “AI transformation” see no measurable ROI not because the technology fails, but because leaders are still measuring value the old way. Generative AI success is hard to quantify when there isn’t a single “right answer,” yet many businesses keep trying to apply outdated metrics to a completely new paradigm. In this video: 00:00 Intro 00:44 The Generative AI Value Gap: Why 95% get no ROI 02:20 The paradox of AI productivity 05:38 Why measuring AI value is harder than we think 12:04 Leadership abdication: “Just sprinkle AI on everything” 15:10 AI infrastructure debt explained 20:17 What real AI readiness looks like (beyond tech) 23:42 Humans as part of AI infrastructure 28:00 Why “AI-first” isn’t one-size-fits-all 33:31 Building human judgment into AI systems 36:19 The risks of scaling too fast 41:34 Automation vs augmentation: where leaders go wrong 44:00 The “do the work” approach to AI success 48:35 The recipe for an AI-ready organization 53:40 Guardrails, governance, and security in AI systems 57:00 Thinking probabilistically: a new mindset for leaders 1:03:20 The human side of AI transformation 1:06:45 Leading through uncertainty Connect with Cassie: Website: https://www.kozyr.com/about LinkedIn: https://www.linkedin.com/in/kozyrkov/ X: https://x.com/decisionleader YouTube: https://www.youtube.com/c/Kozyrkov Visit our website: https://www.infotech.com/?utm_source=youtube&utm_medium=social&utm_campaign=podcast Follow us on YouTube: https://www.youtube.com/@InfoTechRG

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Geoff Nielson?

  Hey everyone, I'm super excited to be sitting down with Cassie Kozakoff, former chief decision scientist at Google and a game-changing founder, AI advisor, and keynote speaker. What I love about Cassie is not just that she's incredibly smart and thought-provoking, but how fearless she is at calling out bullshit and just how good she is at parsing what's useful from all the noise and hype. She believes that companies talking about going AI-first are getting it fundamentally wrong, and we need to completely change the conversation about what AI is capable of. I want to ask her what AI can really do for us and our jobs, and what is the real future of work? Let's find out.

• 0:43Geoff Nielson?

  I'm here with Cassie Kozakoff, former chief decision scientist at Google. Really excited to connect today, Cassie. And maybe to kick things off, you've talked recently about a question that I think is kind of on everybody's mind, which is what you've called the generative AI value gap. What does that mean? And what are you seeing in that space?

• 1:05Speaker 01

  Yeah, so I'm sure that anyone who's been watching the various surveys and numbers coming out about generative AI and generative AI deployments would have found that 95% number. You know the one, right?

• 1:21Geoff Nielson?

  Mm-hmm. Sure do. 95% not getting any value from AI.

• 1:25Speaker 01

  Exactly, exactly. Except the phrasing, I like this phrasing, is measurable ROI.

• 1:33Speaker 01

  Right? So some part of what's going on is that companies are really getting no ROI, and there are fantastically foolish ways to just try to keep up with the Joneses in AI, have no idea what you want it for, kind of send your people off to go sprinkle the magical AI on top of your business, and you hope better things happen, and then you join the no ROI bucket. But there is also some number of those 95.

• 2:04Speaker 01

  Are going to be no measurable ROI. And uh this is uh that this breaks up into two pieces. One is that generative AI is fundamentally just more difficult to measure. And I I wanna double click on that in a moment, because I know that's what you're asking me about. Uh but the other piece is that sometimes what we're actually getting is we're getting the ability to innovate next time. And I think that not enough companies appreciate that innovation demands waste. If you are doing something that you've done before, you know exactly how it's gonna go.

• 2:45Speaker 01

  Then of course, you can have these KPIs that you know you're gonna hit for sure. Because you've already done it. Now you're trying a completely new technology, a completely new use case. You have no idea if it's gonna work. You have to be willing to accept that that might be time and effort thrown, you know, um burned at the altar of innovation, so to speak. Right? That that is just the nature of innovation. And I've had companies come and consult with me who they really wanted to be innovators. But when I ask them, so what is your actual tolerance for getting no results back after you invest in innovation?

• 3:29Speaker 01

  Or how much bandwidth do you give your people to do things that are very specific work product that you expect from them? Do you give them time uh and space to to chase an idea? And quite often, the answer is no, no, we don't. We have no tolerance for innovation. We have uh absolutely no slack for our people. And we need every project to be predictable. Okay, if you're dealing with that, you're just not going to be an innovator. Or you're going to be an accidental innovator. Because you somehow accidentally hired somebody who's going to essentially work two jobs, the one you gave them. And then, you know, the other one, they'll spend nights in the office. And maybe they'll come up with something, but there won't be a lot of these folks. And uh yeah, that's not a great lottery ticket.

• 4:14Speaker 01

  So um if you don't have that tolerance for no ROI when you're trying to innovate, um.

• 4:24Speaker 01

  Just you have to be a follower, just wait for everybody else to to um show how it's done and follow them. But there's another piece with when you actually do this wasteful innovating. You learn how to innovate. And we have Solos paradox coming up again in AI. And Solos paradox um came up in um computers for productivity. And that uh you could see the the productivity everywhere except in the numbers, right? That was the paradox. So how is it that we can all feel so much more productive? How is it that we can have um individuals? Numbers like ninety percent of software engineers are um using

• 5:08Speaker 01

  generative AI to help them code um other numbers like I can't remember if it's if it's ninety or seventy or some big number of um people personally use these tools um in the surveyed population for this um ninety five, no, different study, I think. Uh anyway, uh workers personally use tools and yet a tiny uh fraction of the employers, the companies actually formally give access to these tools. So you've got this shadow AI thing going on where people are using AI, but it's not sanctioned. It's not uh held by their employers, right? You've got this big disconnect. People really like it. They seem to be productive. I'm very much more productive with it personally.

• 5:52Speaker 01

  Um and yet we don't see it in the ROI. We don't see it in the productivity numbers. Sometimes what we're doing is we are just laying tracks to be able to innovate next time, to get the next project right. Sometimes this is the first pancake. And in some sense, when you begin a batch of pancakes, the first one is an investment. And there is a return on that investment. It's just not measured the same way as return on investment of your other pancakes. So I just want to caution people as they're in the innovation games that are just getting started, as there's a lot of uncertainty. Don't expect that there's some magic here. The guarantees that the rules are now suddenly different, they're not. It's the same innovation game as before. But now, now I'm going to answer your actual question and I'll land this plane.

• 6:38Speaker 01

  Finally, let's get back to the difficulty of measuring ROI and the difficulty of value. Talking about value and why there's a value gap. And here I'll say that if you look at how we thought about metrics before with your classic machine learning ten, twenty years ago. We're thinking, um, and when we deploy it as well, we're thinking in terms of minimizing loss or minimizing error. When you have that philosophy of error, what you also have is a philosophy of correct answer, right answer, right? Because if you don't have such a thing as a right answer, you can't have such a thing

• 7:22Speaker 01

  as a mistake, so you can't have such a thing as an error to minimize, so you can't have all the optimization app that we're very used to. So it'll be things like, you know, you'll, you'll have an image classifier and it's supposed to say cat and instead it says dog and we can say that that's an error, right? We can measure that. Or you're supposed to predict the weather and it was supposed to be seventy-two degrees and we observe that it's seventy-five degrees, right? Three-degree error, all in terms of there's a single right answer that we are choosing. Now, of course, there are many wrong answers. If the weather is seventy-two degrees, there, all the other numbers are wrong, right? But there's only one right answer. Now, think about generative AR.

• 8:02Speaker 01

  We are essentially simulating from distributions here and anything out of that distribution could potentially be, if it's from the right distribution, a good-ish answer. Now, think about this, a customer service interaction, an email, a poem.

• 8:22Speaker 01

  If I ask for a, an email to have this podcast with you on a Friday afternoon, I could write that email hundreds, thousands, infinite number of different ways and it would still be a good email. Of course, there's infinite number of ways it would be a bad email. Um, I could, you know, start cursing in the middle of it. I could, uh, you know, send, send you a complete, not an email, but you know, just, just a poem and you would find that quite weird, though you'd probably be intrigued, be like, but definitely invite her on the podcast. Or, you know, I could do the classic mistake and instead of Jeff, I could name you something else, right? Um, lots of different ways to be wrong, lots of different ways to be right, though.

• 9:05Speaker 01

  What should be my tone? What should, how would I know which email is better than which other email? Right? I've got infinite, endless ways to get it right. It's not cat not cat or seventy two degrees not seventy two degrees. It's infinity of ways that I could be solving the task. And now we get the big problem. So far, these are just mathematical um tools or, you know, software based on mathematical tools, based on data, that's doing what it was made for.

• 9:42Speaker 01

  But what it can't do for a leader is tell them what good enough actually means. How do you make this cut between uh completely awful emails all the way across to, I don't know what, how you would even think of like the most perfect email you could get. But somewhere, you're gonna have to draw a line, you're gonna have to create standards of some kind, you're gonna have to talk about how you're gonna measure this. If you're going to have automated emailing, for example, if that's the system that you're gonna put in. And if you're kind of squeamish about that, you could say, well, I'll reduce it to a KPI I know about. I will maybe um

• 10:26Speaker 01

  see how much time I can save my humans if I give them an emailing co-pilot. But now I get some measurement issues as well as a as a manager. Because do I force them to use it or not? Because if I don't force them to use it, am I tracking whether they chose to use it or not? Now, how am I gonna measure a value if they're all ignoring it? And continuing to write, but maybe they are writing better for reasons unrelated to the AI. Maybe that'll look like um results, maybe it won't. All right, we've got, we've got some potential issues here. Or maybe we force them all to use it. They hate it, they haven't learned how to use it yet. Maybe what we're gonna see is decreased productivity.

• 11:10Speaker 01

  And eventually that productivity comes up, but how are we, are we sure we're measuring the right thing? And how would we think about those strange edge cases where every now and then that email is a PR disaster? Especially when we make systems where we take the human out of the loop and now all those emails are going to be sent with no human oversight. Maybe a bunch of them save a lot of time, but there's that one that gets the media interested and that tanks your company. So there are so many different ways you could think about setting up notions of what value is, how to measure it, and how to deal with this curse of endless right answers. And again, most MBA courses, most things that we think about when we think about metrics

• 11:52Speaker 01

  is about targeting a right answer and how wrong are we? This is a different paradigm, and I think it's snuck into our workplaces without us even realizing how much of a different paradigm it is.

• 12:04Geoff Nielson?

  If you work in IT, InfoTech Research Group is a name you need to know. No matter what your needs are, InfoTech has you covered. AI strategy? Covered. Disaster recovery? Covered. Vendor negotiation? Covered. InfoTech supports you with the best practice research and a team of analysts standing by, ready to help you tackle your toughest challenges. Check it out at the link below and don't forget to like and subscribe.

• 12:32Geoff Nielson?

  There's so much you said there, Cassie, that I want to follow up with and so many interesting jump off points. But, but, you know, to your last point, maybe I'll start with this one, which ties into something you said earlier, which is in the absence of really understanding what we should be measuring or what return we're actually looking for on our investment, which is, which is hard. And I think in some ways what we're seeing is like an abdication of leadership decision making to even try to do it. And to come back to what you said, just sprinkle AI on everything, right? Like just, just sprinkle AI and it'll make everything better. When you, when you hear leaders say, just, you know, we need more AI in the business,

• 13:15Geoff Nielson?

  everybody should be using AI, what, what's your reaction to that? And, you know, assuming it's a negative one based on the expressions you just made there, you know, what, what better approach do you recommend so that they can have a better, they can, you know, create more impacted value?

• 13:33Speaker 01

  Look, I am a huge fan of the word why. I'm an enormous fan of it. And I would never presume that the um surface read of a situation, the way that you've described it, is in fact the reality. Right, because sometimes um you get a request that sounds silly and that is because there are, there are things underneath it that nobody is telling you, you, you, you gotta check. Right, sometimes folks are just on the hook to have to do something, anything that calls itself AI, uh because there's a board member that must be appeased or something like this and there's just no way out of this conundrum.

• 14:17Speaker 01

  And then what you do is you sprinkle AI as far away from the business as possible. Where it won't actually touch anything, do anything important. I mean, a, a very classic version of this would be, you know, you could add some, you know, fun, harmless AI feature, uh, you know, maybe it makes music on your, on one of your webpages or something. And maybe it's awful, but you know, it's instrumental, so there's, there's no real way to mess it up, right? Like, who knows, just something, keep it away, keep it away from anything important. So we, we gotta know why, why we're chasing AI. The other thing is, if you have a sense, if the reason that you're doing this as an executive,

• 15:00Speaker 01

  is that you have a sense that the world is going AI first, and I would agree with you. But you don't yet know what you need it for.

• 15:10Speaker 01

  But you wanna make sure that if you needed it, you would be in an okay position to act quickly. Like, that, that is the panic. And the thing to do there is not to actually deploy AI. On my newsletter, I put out a piece about AI infrastructure debt. So this is a term that Cisco has just um, uh, what's it coined? In their AI readiness report that came out this week, their twenty twenty five one. And um in there, and I'm gonna forget the exact numbers, but it's something like, uh, thirteen percent of the, the uh companies that they surveyed, and this is surveys of over a thousand leaders. Um, something like thirteen percent have the preparedness

• 15:56Speaker 01

  to actually do AI and agentic AI at scale, thirteen percent.

• 16:02Speaker 01

  Eighty-three percent is how many plan to release AI agents?

• 16:10Speaker 01

  Right? These numbers, these numbers compute poorly, so to speak. Um, seems like there's a problem and they, they refer to um this AI infrastructure debt as every time that you, you're just sloppily um putting AI into production without thinking about the infrastructure, without thinking, it's, it's sort of like a cousin or an evolution of technical debt. Right? You're just hunting potential problems. Oh, we don't really have the people to do this. Oh, we can't really set up proper guardrails or human in the loop interventions. Oh, you know, if we had to scale this up, we wouldn't have the GPUs, right? All of that stuff.

• 16:51Speaker 01

  That is debt that grows quite quickly as you get a lot of pressure to be in the game of AI. And that has a really, really high uh interest. That's a really, really high interest rate credit card, that credit card. So what I would say instead is

• 17:12Speaker 01

  if you are able as a leader to somewhat blur the boundaries, and I think this is ethically fine, blur the boundaries between what AI infrastructure means to you, your board and your leadership team, and what AI means, you can start thinking about rather investing in what you would actually need when the time comes and you can watch others in your industry show you what use cases are good and not good, and you are set up to hit the ground running and scale quickly and join that thirteen percent from the, the Cisco report. So that might be the smarter thing, right? Start, start investing in the capability of doing it.

• 17:55Speaker 01

  As you're doing this, then you also have these pilots and you can be in pilot purgatory, as people call it. But it doesn't matter because what you are trying to do is set yourself up to be able to use AI properly in the future because you don't know what you want it for yet. You wanna watch, you wanna get inspired by what others are doing. You don't wanna rush in, don't wanna be ballooning that AI infrastructure debt, you wanna be setting things up. Um, that said, also going back to that ninety-five percent study. One of the findings there is that companies that partner.

• 18:29Speaker 01

  Do uh are twice as likely to deploy AI solutions as those who reinvent the wheel, build things from scratch in-house. And then I think there was a note in that report that said something like,

• 18:43Speaker 01

  most of them, or almost all of them that um the researchers talked to had considered or tried to build in-house, right? So this is like an urge. Everybody wants to reinvent that wheel. And the the guidance, and this is guidance that uh we've, I've been, I've believed in this for years, but it has made sense to me during my entire, pretty much, uh, career involved at Google. And now that you should focus on what your particular strengths are, whatever your um business edge is, and let somebody else handle for you the parts that you are not an expert in.

• 19:27Speaker 01

  And you know what you are definitely not an expert in? If you are not a company that does this, or of a tech giant, you are not an expert in AI security. So that's one you don't want to roll at home yourself. You have to go and you want to partner with a vendor that actually knows how to do that. And if it's the teensy-tiniest little vendor, and a security breach could be catastrophic to your business. Well, I mean, look, I am biased because I spent 10 years at at a behemoth at Google. But um you might want to go to a a larger company that actually has the staff, that talentful to do something like uh, secure your client-facing AI, if that's the direction you're going in.

• 20:13Speaker 01

  Um.

• 20:15Speaker 01

  Yeah, I forgot what the question was, Joe.

• 20:17Geoff Nielson?

  So that's all, that's all good. I wanted to um, yeah, we we covered a lot of ground there. And one of the things I did want to unpack a little bit more is the phrase AI infrastructure. And so when you say I in AI infrastructure, you know, is that making sure that you've got the uh, you know, basically, your data is primed and ready for AI ingestion? Is that you're working with the right, you know, stack of vendors in the space? Is that you've got enough AWS credits for all the processing you're going to do? What what is in that bucket? And what if if organizations are truly serious about AI? What do they need to do foundationally to get ready?

• 20:54Speaker 01

  Right, so, so things that you mentioned, whether it's AWS or one of the other cloud providers, those make sense, but those are, even there, we, we would wanna poke a little bit, which we will, you remind me, um, but there's a whole piece that you didn't mention. That's the, the humans, that's the humans on the inside, that's your leadership, and that's also the humans on the outside. Um, one of the, the things that leaders, in my opinion, could spend more time thinking about is how, um,

• 21:34Speaker 01

  whatever AI system that they would love to put into production, how that will actually, be taken up by the people it touches. Right, you really have to think about that. So, eh, when you, when you think about, um, the expectations of your users and what you are deploying into this, um, pot of expectations, yeah, you can immediately be defensive against some bad situations. For example, if you have, uh, all of your users very well-trained to expect a narrow set,

• 22:15Speaker 01

  but a, a unimpeachable, narrow set of correct responses from your system. And now you offer them a generative AI system, that might, though it makes mistakes, AI systems always make mistakes, um, just sometimes it takes quite a lot of scale to, to see them, right? What happens on a, even a very functional AI system, we still say you will meet the long tail, find the, the outliers, the weirdo situations that you did not see coming, even when it's highly performed. Expect, expect something's gonna happen.

• 22:47Speaker 01

  And so you anticipate that there will be mistakes. Now the question is, when a mistake touches a user who has a particular kind of expectation,

• 22:58Speaker 01

  what then happens? How much, how, how flammable is that? Your, your AI infrastructure, of course, it's all the obvious things, your actual infrastructure and your data pipelines and all the rest. But it's also things, intangible things, like

• 23:16Speaker 01

  At what stage are my user expectations? Have I managed them sufficiently where I even could be deploying to users? What about internally? If I'm making, if I'm doing some, you know, in internal corporate engineering, I'm offering some, uh, you know, and now we're looking at the digital employee experience. I'm offering some tools to my employees, some digital tools. Have I managed their expectations? Have I trained my staff? Do they know how to think about these tools? Uh, let's say I need humans in the loop. Am I sure my human will be in the loop? Or might they be asleep at the wheel? And how do I do the training? And how do I put in maybe a collection depending on the importance of the task? I might need to think about having multiple humans in the loop.

• 24:01Speaker 01

  I might need to think about consensus. I, there are all kinds of um, measurement infrastructure, things that we would need to put in place. We're doing generative AI. We've just seen endless right answers. It's a nightmare thing, a nightmare challenge for management, because we've all got to change our paradigm and we've got to think differently about measurement and metrics. Have we done that? Have we put this in place? Do we have testing pipelines? Do we have experimentation pipelines? Do we know how we're gonna roll things back if we need to? Do we know what we're gonna, uh, what versions we're gonna go to? Do we actually know what will happen in what kind of scenario? Do we know how we're gonna make our guardrails? Who sets those guardrails? How do we update them? How are we gonna react to, uh, legal changes, right?

• 24:46Speaker 01

  All this stuff. Now, okay, I know it's hegemonical. You can't say everything is AI infrastructure. But to be ready for AI, there is a lot of stuff that you would need to be ready for. And so one of the ways that you can dodge a lot of this is that you, you do outsource some piece to a vendor who was supposed to do all of it for you. And you just check that you're getting precisely what you need. You have to still articulate what it is that you need. And you have to worry about, measurement wise, that there is going to be a gap, a hole, between what the vendor sees and what you see. There's going to be some bit in the middle that nobody sees. And that could be a huge risk, not just in terms of security, but um

• 25:30Speaker 01

  in terms of your system slowly going um sideways with neither party not

• 25:38Geoff Nielson?

  So, that's, that's, it's really interesting. And, and I, I interpreted infrastructure, as you could tell, in a much narrower, more technological answer. And you're saying, no, no, it's, it's cultural infrastructure. It's the business architecture. It's, it's everything you need internally to set up yourself, success for that type of adoption.

• 25:57Speaker 01

  Look, Jeff, I, I also think, um, maybe this is just a great one to pull out. But if we take things to their limit, to their, their logical conclusion.

• 26:09Speaker 01

  You and I, I imagine, um, nerdy folks that we are, um, have a certain love for technology. Right? We still think of technology as this, um, uh, this ugly duckling thing that, that we, we have to suffer with a little bit. And after a while, we love it because we've, we've done battle with it. It's a little bit fiddly and detail oriented. You know, I have to learn some unnatural languages, that kind of stuff. Um, or it's, you know, physical hardware objects that might, and we still remember how annoying things were to set up. Right? We, we still have this feeling that the majority of

• 26:51Speaker 01

  technology is the fiddle and a long journey of fiddling and then the pale. But what we will see is that technology as the, the actual experience of building it and interacting with it becomes so much easier.

• 27:06Speaker 01

  We will all move to less fiddling and more architecture, more thinking about how everything fits together. And we will see that what everything fits together with is not just all the other complex technology, which is, there's going to be plenty of, but also all the ways in which it touches physical reality, quite often through humans. So if you're not thinking of humans as now part of this infrastructure, because you're you're still used to wires and bits being what infrastructure is, then I'm not sure, like, will that way of thinking have to snap, crack at some point within the next decade

• 27:48Speaker 01

  or so, when we realize just how human and ambiguity-filled and unpredictable technology is?

• 28:00Geoff Nielson?

  Right. And, you know, as you were saying, I was reflecting too. I mean, part of me wonders if some of like the hubris of the moment is people believing that that fiddliness isn't here this time with AI and, oh, AI is easy. You just, like, you really can sprinkle it in and it's all built. And it's like, even in this world where we get toward AI being able to remove some of the fiddly stuff, focusing more on integration, focusing more on, on, you know, sort of the magic that people bring to it, we still have to build that, right? And so coming, coming back to that initial report, and I love, by the way, so many people I talked to saw that report and were like, this, this is the end. It's a bubble, you know, AI is dead.

• 28:44Geoff Nielson?

  And, and, and I mean, I love that you come in and you're like, I'm not even surprised. Like, of course, it said that people have the wrong expectations.

• 28:52Speaker 01

  For me, I just, I just need to make sure I publish before, right? I can just tap to be like, be like, this is January. Tap, tap. Your report came out in the middle of the year. Don't make me tap at the board again.

• 29:06Geoff Nielson?

  No, it's it's it's really good, but it you know, it's making me wonder as well. With this belief that, you know, AI is so easy, you know, just, you know, plug it into anything and, you know, magic will come out. And that bubble sort of being deflated, you know, it comes back to.

• 29:25Speaker 01

  I can't wait.

• 29:25Geoff Nielson?

  Sorry?

• 29:26Speaker 01

  I can't.

• 29:26Geoff Nielson?

  Yeah, you know, I I can't either. I can't either. But I but I did want to ask you, you know, there's there's another phrase that has been floating around, which is, you know, the AI first enterprise, the AI first organization. And so I think part of what's the why people are coming back to leaders are, oh, I want to get there. I need to to be competitive. And there's this, you know, sort of race, and I deliberately choose the word race because it's like as fast as possible. Like, I'm not budgeting for any architecting. I just want results now. Is this? Is that the right approach for everybody? Like, should everybody be in AI first organization?

• 30:09Geoff Nielson?

  Or is that only right for some? And what do you need to do to answer that question and to get there properly?

• 30:15Speaker 01

  Yeah, I look, I love this question. I also want to say to the to the the question of, it's going to be relevant, I promise. I land the plane eventually. It's not blocking. It's not blocking bridge. People are like, is she blocking and bridging? No, no, it's the plane is taking a look, but we'll start here. The part of why I want that that

• 30:36Speaker 01

  AI is easy notion to die, but also to live. Because I want people to see two separate things here. And that lessons from one don't translate to the other. So the question of if I am always my own human in my immediate loop. And I am just dealing with language, right? I just have a large language model, so connected to anything, maybe the internet. But all it does for me is it it gives me language back on screen. So language can only hurt me in that setting if I let it, right? If I'm gullible, I ask a stupid question, I get a stupid answer, I don't realize it's a stupid answer.

• 31:17Speaker 01

  I go, you know, change my medical habits or something on the basis of it, right? Bad and it hurts me, but it hurts me because I took it and I did it. But mostly what I can do as an individual is I can play back and forth and ask anything that's immediately of interest to me. And if I'm bored, I leave and if I'm compelled, I stay. And as a worker, if it's making me more productive, I stick with it and right? If it's a choice, otherwise I put it aside. And in the moment, right? But the human is generally, if they're using these tools well, shaping a whole bunch of output. And that version of AI first of of saying,

• 32:02Speaker 01

  this is actually a pretty awesome tool. If you're good at seeking advice, then everyone in the company should join the new economics of advice. So what are the skills for for getting advice? Knowing what's important, not just their old judgment skills or leadership skills. They're actually kind of hard, though they sound easy, but knowing what's worth asking about, right? That's your priorities. Knowing how to ask, that is context. If you're having marriage issues and you get the best advisor in the world, don't just run up to them, go like, should I get a divorce? Right? They need a little more context than that. Not gonna work. Doesn't matter how good they are, they're gonna give you a stupid answer if you don't supply. And context, a little lot of judgment required. Um, here's not one right answer.

• 32:45Speaker 01

  What you choose to show is as important as what you choose to withhold. And that's gonna shape what we get next. So that's judgment as well. And then the third piece of asking for advice is the skill of not taking bad advice or knowing the quality. Right? So knowing what's important, knowing how to ask and knowing the quality. If you're pretty good at that and you can get better at that, you can get so much out of just the language piece of generative AI. You can learn things faster, right? Give a bunch of context, you're interested in quantum physics, I don't know, but you know nothing about it. Tell it what you do know about, tell it how to, how you like to learn, how you want the output formatted. Get help in finding sources, rabbit holes to go down, right? You will learn quantum physics much faster that way than if you just try to consume old fashioned.

• 33:31Speaker 01

  So that version of AI first, where as a leader, you have uh quite a strict mandate for everybody to, when it doesn't involve confidential information, to get a second opinion for crying out loud from a large language model, right?

• 33:55Speaker 01

  Before asking your manager, before submitting work that's non-confidential again, screen it. Maybe you were asked to create some graphics, right? Before showing them to the the next person or the client, right? Maybe screen it, ask how can this be better? And then ask how this can be better again, and then ask how this can be better? Ask about your prompts themselves, how can this be better? And keep applying your judgment, keep applying your brain and um

• 34:26Speaker 01

  take the good advice, leave the bad advice, and you see yourself supercharged. That version of AI first is not going to harm anything. We're just gonna realize how and how much of a premium judgment is. We're gonna begin to value it more, hopefully. Um and we're gonna realize those with good judgment are gonna supercharge themselves. But this is on the individual level, where you are your own intense human in the loop. So that version of AI first, right? Like, please, do that. Train, train your people, set their expectations right, tell them not to put confidential stuff where it doesn't belong. And for the rest, like, go wild with AI. Um, I had my junior admin that I I hired, um, his his manager, my executive assistant,

• 35:11Speaker 01

  um, has been very, very insistent that he does several rounds of checking things with AI before even passing it to her, let alone passing it to me. The skills that he's managed to pick up are phenomenal and how quickly. Within a week, editing some of my R code for renaming files, right? That that R code is not exactly a state secret. But this is just a young, bright kid with zero engineering background is jumping into editing R code, which is, you know, an arcane language that only only uh the masochists can love, but uh, yeah, what can you do? So.

• 35:54Speaker 01

  And and then, you know, he's gotten really great at graphics in just a few weeks, right? The what you can learn and what you can do that you have no experience in before. And how quickly you can become what I call a a chimeric worker, chimeric talent. You just pull in new skills and get AI to supercharge you when you're doing it with this individual um productivity lens is phenomenal, right? You just have no excuse. So I'm singing that version of AI first um from the rooftops. But it does not translate, does not translate. And every lesson there does not translate to automating with AI at scale. Hands off the steering wheel. Does not translate.

• 36:38Speaker 01

  In fact, it teaches you the wrong lessons and bad lessons. If you're an executive where who is intending to only boost the individual productivity of their workers, you should absolutely be, you know, banging the drum about ask AI before you ask me always. You know, ask AI everything not confidential if if you wouldn't violate your NDA by asking an external friend about something um like, you know, hey, what kind of um style is appropriate for a um uh uh performance review? And give me a a draft of how to express in a kind way the following things for a two out of five stars performer, right?

• 37:20Speaker 01

  This type of stuff without naming any names can get that advice externally. Get that advice internally as well, right? I'm a huge champion of that. But when you see how easy that is, you think that kind of easiness that is predicated on you really micromanaging your AI tool and really massaging where you point it and how you use it. You want to now take this thing and scale it up somehow and let it go. It is not the same game. And that's where all your AI infrastructure debt issues come in. And that's where you you have your meet the long tail situations where you thought, oh, AI is smart. That's fine. Let me put this chatbot in front of a customer.

• 38:05Speaker 01

  And then you have situations like um the the um.

• 38:10Speaker 01

  Car dealership chatbot offering to sell a car for a dollar. Or my personal favorite, when Virgin Money's chatbot berated a user for using the word virgin. Right? You just get this stuff and it's it's a magnet for for the press making fun of you. And it is just such a different game. So when we automate, this is a different thing. Or we take our hands off, this is a different thing. And so you have to know which version of AI first you're talking about. There is absolutely do that. You know, AI as advice, sir, don't take bad advice. It could be bad advice, right? Do that before automation.

• 38:51Speaker 01

  You're gonna have to know how to measure value. You're gonna have to say why you wanna do the thing you're doing. You're going to have to put all kinds of guardrails in. You're going to have to expect that your guardrails were wrong. You'd better have safety nets and plans in place for when things happen. You're gonna have to think about security. You're gonna have to think about, can you scale properly? You're gonna have to think very carefully about how this is gonna be received by people who were not the people who set it up.

• 39:18Speaker 01

  But there is a glorious side to this version of AI first as well, which is why it smells like catnip. And that is that just the simple equation of if AI makes anything easier. But let's imagine it makes email easier, just that. Because it does, probably. I can ramble into the chatbot and say, you know, write the, write a polite email to Jeff. We're doing this on Friday at 1pm and something polite will come up, right?

• 39:56Speaker 01

  I don't even have to prove that I I can make software engineering easier, which we can, right? AI, even just reducing email. And software engineers have more time to write more software. But if we make them productive with software co-pilots, they have more time to write more software. And then we have some of that time will be invested in software for making software writing easier. So we just have this this loop of the things that we can suddenly automate, that that space is growing. That universe is expanding much faster than it used to. And so now, from the executive's point of view, the executive has been told recently, probably, that there is something that they wanted that's impossible, technologically

• 40:40Speaker 01

  impossible, technologically infeasible, just can't be done. And if it's teleportation, that's still infeasible. There are a whole bunch of other things that are suddenly being put on the table for automation. And the version of AI first that I would love executives to have for this automation sphere is not, let's automate with AI no matter what, right? That is, that is bad news. That is leadership abdicating its role. As you said, Jeff, don't be doing that. Instead, before you go do the thing the traditional way, or before you give up thinking that it is impossible and infeasible to run your business the way that you wish you could, you think you've got a technological constraint, insurmountable,

• 41:24Speaker 01

  just take the time to revisit that frequently. Because it could just be that what you need has recently become possible. It could be. And so before assuming and committing to doing things the old way, consider the possibility that AI might have given you the impossible made possible. That version of AI first is what we want to see. But that version of AI first is also a can of worms. Because let's say that you do see that your traditional processes now have an automation option with AI that will change everything. You're going

• 42:07Speaker 01

  to have to think differently about how to measure. You're going to have to think about how to have a foundationally probabilistic system play nice with what is probably a deterministic ecosystem. So are your users prepared for that? Are your workers prepared for that? Are any other systems going to break if you start putting this in place? But things are moving so quickly, you can't, as an executive, afford not to be thinking about it. So in that way, you've got to be AI.

• 42:38Geoff Nielson?

  I love the way you, you know, kind of teased that apart into the different, you know, different definitions of AI first. And I'm trying to, this may be a dramatic oversimplification, but I'm trying to frame this up in my mind, Cassie. And, you know, if I, and I'm happy for you to disagree with the framing, but I'm almost seeing like three buckets here. There's the AI for the individual worker, which is fairly easy. And it seems like there's definitely some, you know, reward or return on it. There's the opposite end, which is trying to just go after everything and re-architect your entire organization. And then there's, you know, if I may, like a Goldilocks option in the middle, which is figure out specific, you know, cases or specific components of your business where

• 43:20Geoff Nielson?

  there's, there was too much complexity previously, but now that can, that can be unlocked. So, I mean, first of all, do you agree with that framing of, you know, the spectrum of difficulty? And then across those, where would you place the value on each of them? Like, is it just as simple as like, one's real, like, one's really easy and low value, the other one's really hard and, you know, big value and, you know, the Goldilocks is just right? Or how would, how would you categorize that?

• 43:51Speaker 01

  I would think about this differently. I would, I would, again, I wouldn't put them,

• 43:58Speaker 01

  I wouldn't even want them to be juxtaposed. I would say, the year is 2025. It is almost 2026. If you still think that you can run your business without introducing your workers to the concept of personal productivity upgrades, it is like you were trying to run your business in the year 2015, ignoring that the internet is a thing. There are a few businesses that you could probably get away with, with just pretending there's no internet. But what a wild thing to do. Right? You don't have to be a technology

• 44:40Speaker 01

  company. Just, you know, get with the program. So, that version is a, it's a really a different set of stuff. And it is, when, when that gets delegated to the IT department, I find it very funny. That is not an IT department thing. Making sure they have access to the internet, making sure the wifi works, making sure that the security system is there, that I get. But the, you know, how do we search the internet is not an IT thing, right? Not now, not in 2015.

• 45:18Speaker 01

  We have access to a store of knowledge. Well, here's human knowledge and we can, we can go searching within it. What would we want to think about finding, right? Very domain oriented. And so, once you've got people past that first lesson, just use it for something. Um, which again, uh, do every time that you notice that someone is asking you something at work where they really could have asked an LLM, just respond LLM until they get the hang of it. That shouldn't even be juxtaposed. Now, the other two things, your other two buckets. Here's how I see them. The thing you're calling the Goldilocks zone is actually the thing where leadership

• 46:03Speaker 01

  has not abdicated their responsibility. As in, leaders had to think about what is worth doing? What is worth having at the scale that they're supposed to operate? Not at individual scale, what's my priority and what should I ask advice about? But big, what do we need? What if we had this would save us a lot of time or money or allow us to open new lines of business or cha- completely change the way that we operate?

• 46:34Speaker 01

  If you do the exercise of thinking about what you actually wish you could do with your business, then, once you have that inventory, you can go through that inventory and see if any of those wishes might actually be granted.

• 46:49Speaker 01

  So that work is not a work for a, a technical PhD. That work is for somebody who's, who's committed to making the time to be strategic, giving them the space to actually think about where they want to direct their company, their organization, and articulate this, right? And this work, this is the hard work. And this is going to be the work, no matter how much we automate, it's still going to be the hard work.

• 47:18Speaker 01

  And I think of your third bucket as don't worry about any of this. Let's just get all our shiny toys in place. I see that as a very expensive way to skip doing the work.

• 47:35Speaker 01

  Because doing the work is probably faster than setting everything up for a maybe I need it. You can begin setting things up like beginning to train people that possibilities. Get people to help you with your brainstorming. Um, you can certainly begin

• 47:56Speaker 01

  delegating some opportunities where individual teams would find their own quick wins and put in some quick win automations and maybe, you know, catch the bug that way. But that's just to get your organization even able to think about or to know about that. There is a vendor ecosystem that here's how uh, here's what it actually means to to connect these pipelines, right? But mostly the work is what would we actually want?

• 48:26Speaker 01

  And again, it's a very, very expensive way to skip doing the work to try to set yourself up for every possible future.

• 48:35Geoff Nielson?

  So let's let's stick on that one for a minute because it sounds like this is obviously, you know, the the approach that organizations need to be taking if they're going to be successful, you know, in in twenty twenty five and beyond. And so in this world where you have to do the work, you have to ask the tough questions, you have to focus and prioritize as a leader. Can you, we we've talked already about AI architecture, right? It's not just a leadership exercise, then have to bring in IT. You have to actually figure out how you're going to do this. What does that recipe kind of look like for you from end to end, at least at least abstractly? And then, you know, what do you see as the biggest success factors and the areas where people tend to go off the rails here?

• 49:19Speaker 01

  Okay, one one question at a time, just so that, you know, those those planes are going to go. It's going to be a whole airshow.

• 49:26Geoff Nielson?

  Okay, let's let's start with the let's start with the recipe. We know there's kind of a right approach here. If I can just, you know, kind of cut the knot and call it that. There's an approach we need to be taking as leaders with IT, with technology, with probably an ecosystem of vendors. Can you just kind of lay out that vision abstractly from end to end?

• 49:46Speaker 01

  Um, yeah, well.

• 49:51Speaker 01

  First things first, uh, the safest way to automate is to do it at home rather than, I don't mean at home at home, I mean internally, rather than client facing. And so again, back to that ninety five percent report, which is quoted to death now. But back office automation was much more predictive of successful deployment than attempting to, you know, automate the salesperson or something like that. And you can see that first, automating the salesperson is an insult to humankind, right? Whenever we say automate the human, um. Instead, what do we want to automate?

• 50:32Speaker 01

  Repetitive drudgery things, by often things involving translation in a way that we're not used to thinking about as translation. We think translation, English to Spanish, but what about um, translation from image to text? Translation from a very boring, long legal documents to a quickly possible TLDR, right? Those types of translations. Uh, so it's repetitive, it's internal, it's back office stuff, involves a whole lot of drudgery. Things that any human could do, AI will probably do better. So if it doesn't matter who it is, if there's no holding of special context within your organization, that's

• 51:16Speaker 01

  and and you're doing a lot of that, um, that is likely a great target for AI. And so you you really want to break down, where are we wasting a lot of opportunities? Where are we spending our time where we don't want to be spending our time? And then at the same time, you you have to look at if we could do.

• 51:40Speaker 01

  Anything.

• 51:42Speaker 01

  If we now said now we've we've freed half of the the hours um of some of our staff members. Does that mean that we cut half of our staff members? Or does that mean that we get ambitious about what what we're going to do tomorrow? Or are we going to pinch yesterday's pennies because AI lets lets us do this? Or are we going to realize that if everybody is um getting into AI? Um, and actually using it to transform their businesses? Then the future is going to get interesting and weird and we're going to need to uh compete with everyone else's move. So so let's think about what that's going to look like and what our wishes would be for how we would prepare for that.

• 52:24Speaker 01

  And we might want to retrain our people whose time we've just saved to be able to be effective in that future. Um, giving an example of Ikea reducing um a lot of the um customer, the basic customer service, I think, phone requests. Um, and this was a a client facing one, um, an external facing one that seems to have worked. Um, but they they reduced those um hours and that that need for that staff. What they did was they retrained them to do AI assisted um interior design, interior decorating design.

• 53:05Speaker 01

  So you have um a complete change now in what you can offer as a business and how you can do business. And you don't need to hire specialists because AI can upgrade many of your non-specialists into specialists. What's more important is uh valuing the people who carry the context of your business, uh, know how your business actually works, who know what your clients need, um, and upskilling them aggressively with AI into an interesting future. So that's where we start with all that efficient stuff.

• 53:40Speaker 01

  That is just the first part. Now we have to break up, break down whatever, whatever we, we are imagining, to break that down into what it would take to make it happen. And you want to ideally have a, a culture not of people saying no to you, but of people expressing, here's what it would take. Because again, when we go and verify those, those individual pieces, it may be that it is cheaper or easier than we realize. And what it takes when we are enabling individual workers, giving individual workers tools with which to work, what it takes is we have to figure out how to train them.

• 54:24Speaker 01

  We have to figure out, um, if something goes wrong, if there are issues, what are the escalation paths? Um, when should they override what the tool is using? When shouldn't they? What guardrails we're putting in place? Um, when it is things involving, um, I don't know, quality control on Warhammer figurines? That's a very different thing from, you know, using AI assisted tools for, um, surgery, right? These are different. Um, cutting a human and, and, uh, cutting a plastic, um, all, um, no, no disrespect meant to the Warhammer community. These are different things. So different applications. And so it, it really, it becomes

• 55:10Speaker 01

  very individual as what you are building. If it is, again, something that is an internal system and it is not actually connected to any external systems, that it's going to have different security requirements to things that are connected to the outside. Um, we see now that, uh, generative AI is now the top, uh, according to a, a very recent report, um, top data exfiltration risk. It's now been email, um, as, as the, the way in which

• 55:44Speaker 01

  corporate secrets will, will leave the company. And then the more, um, directly customer facing it is, the more you have to plan for all kinds of things going wrong. So this is, this is a setting up of, um, guardrails and control structures before we're even talking about vendor selection and actually building the dark thing.

• 56:07Speaker 01

  And then depending on what it is, either there it it's an obvious vendor and a and a partnership will they where they will

• 56:16Speaker 01

  do a lot of the advising from here, because that's what they're supposed to do.

• 56:21Speaker 01

  Or it will be a case of evaluating and choosing among many, and that's its own journey. Or you may find that it doesn't exist externally. It makes a lot of sense for you to put it in internally. You might have to be able to own yourself from scratch. And inwards do that, then you have to attract the talent to do it. There's a whole lot of balancing and and weighing there, while always remembering that this system is not going to exist in isolation. Even if it is the most back office application, it's still going to exist in the context of what your people already do and other systems. And people are not used to software that is probabilistic or used to deterministic. Like if your spreadsheet

• 57:05Speaker 01

  gives you some issue, you typically think that the problem is some human typed in something wrong, the formula, the data. You don't think that every now and then the summation function is just an average and that just happens. And again, not always, not all the time, but eventually mistakes do come up and you need a completely different way of thinking about how to deal with that. You need a completely different set of trainings for your AI users.

• 57:36Geoff Nielson?

  So that's a that's a point that's come up now a few times, Cassie, and so I want to double click on that a little bit because the the notion that with AI in general, with any of this generative stuff, it's not black and white. It's not right and wrong. It's probabilistic. And and you know, you've said that that has sort of, I guess, cultural and mindset implications for, you know, everyone it touches in the organization. And and it requires a new way of thinking. Can you share a little bit more like, what is that new way of thinking? How how do do leaders need to be communicating that? What are like the right skills and mindsets to actually, you know, thrive in in that world?

• 58:19Speaker 01

  Yeah, so there's a there's an answer that works for right now. And there's an answer that'll probably work for a nearish future.

• 58:29Speaker 01

  Because I'm of the the opinion that we will culturally all adjust to being more tolerant of mistakes. And so in situations where mistakes are, uh, let's say, life threatening, then we will have um a lot more guardrails, safety nets, and so on than we usually do. But for the rest of everything, I think we might actually develop more of a sense of humor. We will just have situations where the chatbot does something funny. And, you know, I I can sort of imagine a future in fifty years' time where instead of that just being such a

• 59:11Speaker 01

  everyone is amazed that a corporate system could do that, people would laugh it off, understanding that as a as a civilization, we get a lot out of um being able to do more and automate more. But the price for this is the occasional mistake. And we we've got to we've got to accept that. And we will we'll know, we'll we'll expect it, and for that reason, we won't um be alarmed.

• 59:42Speaker 01

  Now, um in the near term,

• 59:47Speaker 01

  uh, and it's funny, so I'm a recovering statistician, I I think proba probabilistic thinking has been beaten into me, um, you know, I I've been it for more than half my life, so um it's actually quite strange to me to think deterministically, um. But I'm always I'm I'm always a little bit cautious when things are important. It's there is a sort of trust issue, maybe, that we statisticians and recovering statisticians have that the rest of us could adopt. And the worst way to have trust issues is to be paranoid about everything all the time.

• 1:00:31Speaker 01

  The best way to have trust issues is to first have a filter. Which is your um of your core skills as a decision maker. Is to have a filter based on the importance of what's at stake. You know, the the reasonable best case and the reasonable worst case should um inform how much cognitive effort you put into something. And so there are so many situations in life where we know if something goes wrong, we're just shrug it off. There is not that much reason to overthink and check and obsess and worry about. But when you um. I mean, uh, life is probabilistic, also not deterministic, but but I think.

• 1:01:17Speaker 01

  And as an example. When I'm thinking about my travel plans. I'm always amazed by friends who expect things to actually go the way they're supposed to. Right? I'm amazed by this. It blows my mind. I had been seeing fifty plus possibilities of how things could go. Adding time, subtracting time, you know, where which levers I can press, which ones I can't. Um, how do I set myself up with more optionality? I that I just do that like breathing. And some people are like, huh, the flight's delayed.

• 1:01:52Speaker 01

  What a surprise that this kind of thing happens um. We are just gonna have to start applying this. These trust issues a little bit to our tools, same thing as out there. How do we? We have to start thinking about. Let's imagine that mistakes are possible. Let's just really hold that thought.

• 1:02:15Speaker 01

  Sometimes it will do something weird. Let's just imagine that that's true. Where can we apply this? Where it's fine that that's the case.

• 1:02:24Speaker 01

  And maybe that's subtle, maybe that's very obvious, I don't know for listeners. But I think that there's, there's so much, um.

• 1:02:34Speaker 01

  There's so much framing around how we use AI that suggests that we put it in first and then we worry about the quality of it afterwards. Instead, so you have a fallible tool, just like so you have a fallible human. And you know that humans are fallible. You begin by designing the guardrails and control structures before you even get there. You, you deploy the humans to where it's acceptable to have mistakes.

• 1:03:05Speaker 01

  And you know, sometimes mistakes happen. And so you set everything up with the expectation that mistakes are possible. And just do that same thing with machines. You don't have to deploy AI to every which where. Yeah, it actually takes some thinking from leadership of, let's say this thing completely messes up. Where in my business would that be fine?

• 1:03:32Geoff Nielson?

  It's, yeah, it's, it's really, really interesting. And I, I don't know if this is a word you like or hate, but I, I keep coming back to the word architecture and architecting, like actually understanding the entire system here, the entire series of processes here and thinking intelligently about how to redesign it, you know, with governance, with guardrails, with, you know, an understanding of where the value is, what probabilistically could happen and just like, as you said before, to quote you, like doing the work here. And I, my, my sense here is one of the bigger challenges for organizations is when you're talking about this across the scale of, of an entire enterprise, we're, we're talking about a transformation here, right?

• 1:04:14Geoff Nielson?

  And, and, and how do you, how do you do that? Like, how do you muster, I guess, kind of the organizational will? And I guess also the political will where you're willing to let, like, how, how do you let people make the right decisions and who are those people? If I'm not asking too big a question with that.

• 1:04:36Speaker 01

  Well, that's, and is it just one again? Or how, how many questions are in there? Quite a, quite a bunch.

• 1:04:45Geoff Nielson?

  Take, take that wherever you want, Cassie.

• 1:04:47Speaker 01

  No, look, um.

• 1:04:50Speaker 01

  Transformation is hard. It doesn't matter if it's AI transformation or any other kind of transformation. And I think that the, the more technically minded you are, the more you're just like, oh, it'll transform, you'll install the latest version and everything will be great. And the more you've spent time with people, you do understand that, that the organizational will and the political will are very real beasts that are very hard to wrangle. And if there was a magic

• 1:05:25Speaker 01

  spell that would make you win at that every time, right? That, that will be the, the number one best seller for the rest of time.

• 1:05:35Speaker 01

  Unfortunately, that is exactly tricky because every time that you have something working a particular way, there is a, a whole ecosystem that that feeds. And when you disrupt that ecosystem, you have to compassionately manage that change and figure out what's going to happen with all the various people and their skills and their relationships and their knowledge. And we are a species that creates technology that builds upon its technology and makes more interesting technology. Um, you know, starting with some, some sticks we rubbed together all the way through to now, and we've managed to somehow manage that change and occasionally it hasn't gone great, but

• 1:06:20Speaker 01

  um, takes good leadership and good leadership. Again, if it was just that formula, we could teach it to everyone, wouldn't the world be beautiful? But, you know, that's not how this works. What we, I guess, have to understand is that the human side of it is going to be at least as difficult as the technological side. And then to come back to something, cause I, I, oh, callbacks, something you said way at the beginning. About, um, is AI readiness preparing your data? You phrased it like that.

• 1:06:57Speaker 01

  And now you also said about architecture will be knowing how all the systems work. I want to take these two concepts and combine them with another concept to to truly make this whole thing sound dismal, so that leaders know why they have to back up.

• 1:07:15Speaker 01

  So the the there are different channels of AIs, I like to think of them. And sometimes it's easy to think you're on one channel, having a conversation with someone else on that same channel, and actually you're having you're talking completely postponed. So to to kind of come back to basics. There's all the AI stuff, that's the theory, that's your researchers publishing papers. If you're an executive, if you need that, you already know you need that. If you don't know why you need that, skip until you know. Then there are a few other channels and you might be on one of them. And this is data, generative AI and agents.

• 1:07:55Speaker 01

  The data thing has been what we've been doing, that we've called AI for a long time now, or machine learning, where we find patterns in data. We turn them into recipes that a machine will follow, the a substitute for human written code, human written instructions. And what we're doing is we're essentially expressing our wishes with data, with examples instead of instructions. And now we let's ask ourselves, why? Why would we do this? Instructions are how we get control. If we know what instructions we wrote down, we know exactly what's going to happen next, we get a nice deterministic solution. Why, for goodness sake, would we give that up? Sounds mad.

• 1:08:36Speaker 01

  The reason we give it up is that there is some thing, when it's important, when when it's lazy and it doesn't touch anyone, we can do what we like. When it's important, when the business relies on it, when the when we need return and investment. The reason we do we use data for automation is that we cannot come up with the instructions. That's why something about the instructions is hard.

• 1:08:59Speaker 01

  Uh, if it's that we were wrong about them and we had to check and we did some science and now we can write them out again, that's not AI. Right, that's good research. Nice, nice science. If we are physically incapable of writing those instructions.

• 1:09:16Speaker 01

  That's AI. Use data in that situation. And the reason we're physically incapable is that it is so complex. The solution is so complex. Through our complex problem. It doesn't fit in our heads. So even that, that data piece on that channel. What we see is that data gives us the gift of memory. That is what we are doing. We're, we need the gift of memory to deal with complexity. We need machine memory. For vast complexity that overwhelms humans, that makes us not able to write down instructions. And so the gift of memory is a powerful thing, but we also know that we sign up for everything complexity implies. All its opportunities and all its threats. Including when things are complex, you don't know how they're going to go wrong.

• 1:09:59Speaker 01

  So that's why you start to think about paranoid things like control structures and security and guardrails and all this. Generative AI, the next channel, is when someone else, not you, someone else. Apply to that automation with data principle to language.

• 1:10:21Speaker 01

  And. Or video or physics or something like that, some sensory.

• 1:10:28Speaker 01

  And now, uh, but focusing on language, because that's the one that that really um. Is, I think. Valuable for folks to, to connect with. Language is the universal interface for human collaboration. It's how we collaborate with one another. And so if we solve language understanding. Any of us can communicate with a machine.

• 1:10:54Speaker 01

  So companies like Google, Anthropic, OpenAI, that lot. Providers of the foundation models, they did the paradigm in the previous channel. Applied that to solving language. And then they offer tools to developers and to end users. It's like language, here you go. And on this channel is all the what could we do with language? If we could all speak without learning some unnatural language like Python, C plus plus, the rest of it first. What could we ask? So this is the gift of language. Sits on top of the gift of memory. Language is pretty complex. What could possibly go wrong? We're sitting with built on data. There's complexity here. What could possibly go wrong with language? Well, our unnatural languages are pretty stiff and constrained, but at least we know what they're gonna do.

• 1:11:39Speaker 01

  Our natural languages, our mother tongue, is filled with ambiguity. Half the time, we don't know what we're saying. So when we automate that and we're not watching and we're not our own human in the loop, we automate based on, you know, something that sounded good at the time, that's your genie story. With potentially a very unskilled wisher not even knowing what they're saying. And at scale, kind of terrifying. And then the agentic piece, that is the gift of action on top of the gifts of language and memory.

• 1:12:11Speaker 01

  And the gift of action is only a gift if the action is good and useful and something uh one that makes your life better as opposed to worse. And in the agentic paradigm, you realize that there was a very, very important piece in that genie story. It's not just the genie and the wisher, it's also the lamp. The lamp is really important. What is the, the actual control structure and security? How do you prevent an unauthorized wisher making requests on a powerful genie?

• 1:12:42Speaker 01

  How, how do you deal with all this stuff?

• 1:12:46Speaker 01

  So that's what we've got. Now you're like, what the hell, where the hell is she going? I'm bringing this back. First, you might not need anything to do with getting your data in order for doing AI. That might not be part of your AI infrastructure. You might just be wanting to ride what it means to be able to play with language.

• 1:13:05Speaker 01

  And plug language understanding into your business processes. That gives you rather powerful ways of translating things. All kinds of things, if you have the creativity to see it. Or you could plug that into the gift of action as well and have autonomous actions taken. Then you've got to be very careful about what kind of wisher and what kind of lamp you're dealing with. And if everybody is just talking about the genie, you're missing a big, big piece. But all of this stuff, this is why, Jeff, you're like, wah. All this stuff is about, is built on top of complexity. And so now you can't say things like, I'm gonna be an architect and I'm gonna know all the, all of the, all of the, all. There's gonna be so much you're not gonna know.

• 1:13:47Speaker 01

  You have, there, all, all of these engines are complexity based. That's why we start talking about this is probabilistic now. Um, some of it might even be deterministic if you could understand it, but it's too complex for the human mind. It may as well be probabilistic. When it's generative, it's literally probabilistic. So it's very, very complicated and these complicated systems that can go wrong in complicated ways. Are going to be put in first one. Now you've got one complicated system, you've got to try to understand all the architecture around it. But once you've got a bunch of these, plus a bunch of augmented employees, workers, who are all augmented in their own way.

• 1:14:31Speaker 01

  That is a lot of complexity all moving around. And so this idea that you would, you would understand the all of it. Uh, that does, that doesn't compute. What you're going to have to get really, really good at is understanding your piece and the edges where it connects to the universe. And then it's going to be a game of trust, trusting other humans who are, for lack of a better way of putting it, gardening their own gardens and how that connects to you. And still being tolerant of, there can be errors and leakages. And the more that could hurt people, the stronger again, the lamp, the guardrails and the rest of it need. So that's what it comes down to.

• 1:15:16Speaker 01

  Um, the wisher is much more important than the genie and the lamp might be more important than both of them. If the lamp just shuts everything down, um, when we need that.

• 1:15:29Geoff Nielson?

  So first of all, I feel like I have to say thank you for landing that plane because that was a, that one was a wild ride. And you had me for a minute and then, oh, there, we've landed the plane. Um, and I, and I, I love the story that you told and I, and I love the way you kind of framed the role of complexity and some of the implications. But I wanted. I want to maybe take that up a level and, you know, apologies because I feel like I'm going to give you another multi-part question. But when you think about this world with increased complexity, where it's, and we can start with an organization, but when we've got increased complexity, where we've got people who can't see it all anymore, they're tending to their own garden. And, you know, as you said, there's, there's control implications here, right?

• 1:16:13Geoff Nielson?

  We're giving up control in the name of being able to do more, to tackle more complexities. What are, what are the implications of that in terms of organizational leadership? You know, in the, in the scope of the future of work and the future of the organization? And then maybe if you're feeling bold, what, what are some of the implications of that? Do you think more broadly societally or outside of the organization?

• 1:16:38Speaker 01

  Well, I mean, look, that, that cuts to the heart of it, right? But, but the idea that we will eventually, as a, as a species, and we do already, like if you, if you think about it, how do you know that a skyscraper that you, you know, you go up to the 50th floor? How do you know that that all works? I don't, I know nothing about this, but somehow there it is, right? We, we do end up having to collaborate with others who all hold their, their piece, their bit of something much greater than us. And occasionally, there

• 1:17:19Speaker 01

  are mistakes and problems. And we really all do try our best. So, well, not all of us, some of us don't try our best. We can talk about work slop anytime you like, the, the AI kind and the regular kind. But, you know, we try our best to put systems in place that, that limit the, the speed of the damage of things going wrong. And we, we do give up control already. It's just that we are adding it, it's going to be another order of magnitude.

• 1:17:54Speaker 01

  And it's going to be systems that are fundamentally probabilistic. So that's why I said earlier, as a society, we are going to have to be more tolerant of mistakes, more aware of the possibility of mistakes, um, less gullible. I think that we're going to go through some, um, growing pains, for example, now with, um,

• 1:18:18Speaker 01

  with propaganda applications, um, people who don't, for example, know that

• 1:18:24Speaker 01

  a video could be entirely, um, fake and AI generated and it might look real, but it's not evidence of anything, right? We just need to grow as a, um, as a species to now realize that we, we exist in a future, in a present, where that's technologically possible. And so that is no longer our way of forming trust. And maybe trust will have to move back towards, uh, trusting humans, trusting individuals, individuals that we know. Um, individuals that we know to do their jobs well, to, uh, take care of their piece of the, the, um, conflict, complex sphere that they are in charge of, knowing that there is no perfect way to do this.

• 1:19:07Speaker 01

  But you know, the one thing that I, I just keep coming back to, I hate, I, I hate, um, making predictions, this is a statistician talking, so you've got to love the irony. But I hate making predictions for the future, particularly when, when there's so much change. What can you really predict? How do we know what the world's got to look like? But my one controversial prediction that I, I feel pretty okay with is that we're all going to have more to do, not less. Just because of that sheer complexity and all the different ways that it's, everything can fit together. I mean, imagine, imagine a future where all the kids are all on individual learning and education journeys.

• 1:19:49Speaker 01

  Uh, workers are fundamentally chimeric and can, um, pull in new skills with AI as needed and put those skills down. Just, you know, rattling the concepts of job ladders and, you know, bursting the, the boundaries of their roles. And taking on new challenges each week as needed. And then you have technologies that are going to put all kinds of, um,

• 1:20:19Speaker 01

  Probabilistic things again, probabilistic automations at scale. All these pieces all moving around. How do we coordinate all that? How do we make sense? How do we plan? How do we regulate?

• 1:20:37Speaker 01

  There is going to be so much, um, you know, uh, what am I trying to say? Uh, the, the duck's little feet under the water are going to be, have to be swimming very, very quickly to catch up in this. I don't see us all, um, going back to our n- natural state of grace, naked on this event or just pulling a, a dropdown menu from the sky. And, you know, ba- back to, back to our, uh, hunter gatherer roots. Um, I, we're just, we're gonna have a lot of work, a lot of work ahead of us.

• 1:21:08Geoff Nielson?

  How do you, how do you see that work being distributed? Because one of the, one of the points of fear that, that, you know, I hear a lot about these days. And I, I love, by the way, that you brought the conversation to one about trust, because I think that's kind of an operative, operative word in all of this. How do you see that, that orchestration, that work, that planning being distributed? Because the, the fear that I hear about is this is going to be overly concentrated in big tech or in a few, you know, small firms that through the platforms that they operate have outsized power. And, you know, people fall less and like people fall, th- they become less important and less valuable in this new economy.

• 1:21:50Geoff Nielson?

  I- is that a concern to you or do you see it being more decentralized than that?

• 1:21:55Speaker 01

  Um, I guess I, when I put my economist hat on, um, I think to myself about barriers to entry and which parts will have high barriers. And the, the hardware and hardware innovation, um, those parts, I think, have very high barriers. And I'm not sure whether it's going to be possible for many, many, many players to compete, right? And see that that's, um, uh,

• 1:22:30Speaker 01

  certainly a centralizing, of course. But then we, if we think about things like, um, What would be things keeping consumers loyal to, let's say, platforms, apps, etc. It's often laziness. It's not that it's the best thing. It's just maybe it's hard to move to another one.

• 1:22:57Speaker 01

  Uh, maybe you've given all your data away already and the system knows you really, really well. So a competing product would be difficult to stand up. Right? You could see those types of of barriers and those types of capture. But technology is also going to make things pretty easy. Everything a little bit easier for everybody. And

• 1:23:22Speaker 01

  I don't know to what extent that will uh uh counter um natural inertia. But I do think that it will be easier for anybody to um make the attempt, I guess, make the attempt, at least in the digital space, to offer something and offer something um nuanced and personalized and to be matched. Right now that we've got pretty bad matching, like, um, you know, maybe somebody could might be designing the most perfect custom T-shirts uh for you, Jeff, and not somebody might be based today in, I don't know where they they might be living, maybe they're in Jakarta.

• 1:24:05Speaker 01

  There's not a great system to connect you with them and with their art and with their um very individual offering to you. But I'm I imagine that um providers of um that matchmaking, that will be a high barrier to entry thing, that will make a lot of money. But um all entrance into that, I think, have an opportunity to to um participate and participate in interesting ways in the economy that uh they wouldn't have before. And there'll be a lot more um very small businesses that that start because of this, realize with a very few people, you could do some unique, interesting things.

• 1:24:47Speaker 01

  And um while we will see more personalization from behemoth corporations again, um managing all the humans and the humans then being clear on what precisely they provide. Um, there will be a sort of, uh, how shall I say this? I'm trying to find the word.

• 1:25:09Speaker 01

  It will be hard for the humans to track and take responsibility for very different or very custom offerings, even if the system itself could do that. So I still expect that at the edges, you get the interesting stuff. And if the barriers are lower there, you could have a lot more entrance at the edges. If you have good matchmaking, you can have a very interesting digital and physical and services.

• 1:25:36Geoff Nielson?

  Cassie, this has been super interesting and super insightful. I really appreciate everything you've shared today. You've certainly given me a lot to think about. Thanks so much for joining.

• 1:25:45Speaker 01

  Thank you for having me.

• 1:25:47Geoff Nielson?

  If you work in IT, Infotech Research Group is a name you need to know. No matter what your needs are, Infotech has you covered. AI strategy: covered. Disaster recovery: covered. Vendor negotiation: covered. Infotech supports you with the best practice research and a team of analysts standing by, ready to help you tackle your toughest challenges. Check it out at the link below and don't forget to like and subscribe.