520 — Crossing the 'AI Divide'

Hello, and welcome to the MindTools L&D Podcast, a weekly show about work performance and learning. I'm Ross Dickey, and I wanted to start this week by thanking Julie Dartson for guest hosting a series of conversations around AI over the last three weeks with Alicia Sanchez, Green Pagano, Marcus Bernhardt, and Samuel Salter. If you missed any of those episodes, I'd highly recommend that you check them out. Julie's series focused on the behaviors associated with AI, but today we're taking a step back from that, ask what return organizations can expect to get in their investment in AI, and what they need to do internally to maximize that return. You're with me to do that. I'm joined by Dr. Anna V from our Insights team. Hello, Anna. Hi, yeah. And we're also joined by Rodrigo Balanios, our global AI strategy officer. Hello, Rodrigo. Hi, Ross. Thank you for having me. Good to have you with us. So, Anna, I'll start with you. You've been doing some analysis recently and defined for what you call the AI divide. Can you just explain what the AI divide is and what the divide costs? Yeah, of course. So, I feel like the most obvious way usually to describe the AI divide is AI adopters essentially versus non-adopters companies that have AI versus all those that are not. But for us, we think that's a more interesting issue could be actually happening within the companies that have adopted AI. For instance, the U.S. Census data shows us that something about 18% of U.S. businesses are using AI at the moment, which doesn't sound like, which maybe sounds like more or less. Yeah, I know. That doesn't sound like a lot, but it's almost one fifths, right? So, it's okay. Yeah. And I think if you think about it, like I think I just assume that my work is similar to a lot of other people's work, but like there's lots of professions that very reasonably would not expect somebody to be interacting with an LLM or any kind of AI for that or a different thing. Yeah, or even a desktop device. Yeah, exactly. Yeah. I have my haircut a few weeks ago. I'd be quite surprised if my hairdresser used to. I've been made for inspiration or something like that. Yeah. Yeah. So, yeah, it's a little bit lower than maybe you'd expect, but that kind of makes sense. But the thing is, out of those 18%, nearly two thirds basically say that they actually haven't changed anything to actually support the use of AI. So, they didn't introduce any new workflows, they didn't introduce any trainings, they essentially did nothing but purchase a tool. And even when we look at a pretty, at a sector we'd expect to be pretty advanced. The information sector is just under 10% of companies that both adopted AI and also introduced some training on AI. So, I guess for us, the divide is essentially between access to AI and the ability to actually use it. And like you said, this divide has a cost or maybe even two types, of course, actually. So, one cost is financial. Obviously, the thing we care most about. And we have, for example, there was an interesting research done by KPMG who modeled two futures for the 2030. Essentially, in one, the US adopts AI very quickly and also upskills the workflows on it. But in the other, the US kind of adopts it quickly again, but doesn't do anything about the skills and the gap between the two is estimated in terms of kind of return and investment to be around 1.3 trillion dollars, which now that's a big number compared to 18% essentially. And this is the price of the divided kind of the national level. It doesn't just cover training, of course. But when we look at the employer level, so we look into how many people will realistically get upskilled through university education, how many will get reached by workplace training and so on. And essentially looking at how much of that gap employers can realistically close through workplace training. In our central scenario, good applied training with decent employee reach could unlock about 177 billion dollars by 2030 again, pretty decent. So that's a financial cost. But I think for us, what's also interesting is a human cost because this divide causes low anxiety. For example, there is a survey from 1026 that found that something like 80% of employees worry that they could be blamed or even fired when AI mistake. And as a result, like a third of them also say that they've held back from using AI. So all of this is kind of blocking adoption and just providing a license. It's not going to fix things. I think they're really interesting, particularly the human cost. I think on the technology side, I mean, this is something I guess that we've had conversations about in this podcast before is that kind of simply purchasing co-pilot licenses and upvading support for training or governance around that is not going to have the sort that you want it to. L&D audiences will be familiar with this or HR audiences from purchasing some sort of like technology platform. It's not the platform is useful. It's only useful if you kind of do all the work that sits around that to actually get people to use it. If you embed it in your workflows and and that sort of thing, oh yeah, it's essentially like, you know, giving someone a dream membership right. So access is necessary, but doesn't create its own habit of like going to the gym, how to use equipment, the confidence of doing things, you know, all of this needs to be learned and supported. Otherwise, it's a waste of money. I joined my first gym in summer. So, you know, that's for an example, Constable. Yeah, before you said that, I was feeling very, very exposed. It's universal experience. Yeah, it's a universal experience. I listened to this podcast. This is a few years ago now. I think it was a plan and money episode where it talked about the business model of gyms and how effectively the people who purchase memberships and don't go along regularly or subsidizing the cost for everyone else. So the people who use it regularly, their cost is effectively subsidized by the people who purchase a membership with good intentions and then go for maybe a couple of weeks and then stop going, but don't cancel their memberships because they've gone the most profitable. Yeah, there's the ideal customers yet. Yeah. And we also have, so there's others that we have on this page on our website that we can link to which summarizes the analysis that you've done claims that at least 75% of the responsibility for building AI capability sits with employers. What's behind that claim? You sort of touched on universities briefly in your previous answer, but yeah, do you want to just expand on that and where the, where the 75% number comes from? Yeah, of course. So in terms of the universities, it's kind of, I feel like it sometimes feels like, oh, well, you know, we can train people where they come to get trained, you know, university education, stuff like that, but realistically, the percentage of new workforce that kind of enters the labor market every year only constitutes about something like 1% of the US workforce. So that's a really, really small share of them. So you can't really rely on university-based education to and expect that your employees will be prepared just through the university. So that leaves quite a lot to the workplace-based training, but again, like how much of it should be done by the workers, how much of it should be sponsored by a same employer. And the number itself is 75%. So we, there is an interesting kind of data again that comes from OECD. So it shows that among US workers who went through a significant technology change at work and received training to adapt to this change. So 75% of them had that training paid for by the employer. That's our main anchor of this figure. It is obviously a neverage, which means that the actual number will vary. It's inevitable. Some employers will fund less training, some will fund more. Considering how massive overshift AI is causing, one would hope that the employers would aim to at least stay at the average level, rather than fall behind it. So that's kind of where the number itself comes from. I also, I mean, we also feel like it makes somewhat an intuitive sense because most people learning AI, like we said, they are already at work. And the time, the tools and so on, all of that will belong to the employer. And while it may be tempting to make an argument that while developing your skills, your own responsibility, your motivation, you know, the basic versions of the AI tools are free, the tutorial soundlines. So, you know, just get on with it. And some people do, of course. But behavioral science, of course, tells us why that isn't enough. And the reason is that most of what drives behaviors, of course, contextual, especially for organizations, people can of course learn prompting, for example, their own. But they can't create for themselves the context applied. So it's such as access to company data, clarity on what's allowed, for example, permission to change how the work that all depends on the employer. And when the employer leaves kind of this vacuum, people fill it themselves and they sometimes kind of fill it with staff that employer might not be very happy with. In cybersecurity research, for example, when guidance is unclear, people essentially build their own kind of personal security models, basically a mental model of what's secure and what's secure according to them and they might often be you know quite wrong about that. And AI is the same, you know, there's some data such as something like 40 percent of employees or more put sensitive company information into public AI tools obviously without improvements. It's not bad and it's a lack of guidance. And finally, I guess learning new ways of working takes mental space and requires a safe room for experiments. Again, we've been doing quite a lot of different kinds of research on cybersecurity in AI this year and we find that feeling safe to admit mistakes and ask questions is strongly linked to a whole lot of better cybersecurity behaviors, including staying safe with AI tools. So essentially individuals are responsible for their effort but employers as a ones responsible for the conditions that make this effort worse doing. Yeah, and I mean, you said in your first response that actually, I mean, adoption overall is maybe lower than we would expect it to be or then you might think it would be if you're going solely based on your LinkedIn feed or what you're reading and then use. But then all of those companies that have adopted it a lot are not providing any kind of training or support to use these tools effectively. Rodrigo, for those that are providing training, I've effective our kind of most upskilling initiatives and and why when they do feel is that what is that happen? That's a very good question because basically buying that technology or offering a short course on how to use AI is not going to be enough right because this has to be treated as a transformational change in processes and culture. What you just mentioned Anna is very important because that safety about X, I can explore with this technology and that's that's very important. That's a cultural thing. You know, this that I can do experiments, you know, I can propose new ways of doing things and that's not coming from a simple course on how to prompt, you know, now, now, even technically, now it's becoming much more sophisticated than just prompting. That was it's a three-year old term, I would say. Now, we have more complex approaches such as context engineering agents, you know, loops, goals and all this much. It's amazing how you how you can basically make the AI work for you and do amazing things, right? But in order to to be there and what is happening now is the training, this is something that we have been living for the last in my case 50 years. Training is the last thing to do, right? It's often is the final step and if we want as companies, as organizations to achieve those levels of promises from this technology, we need to treat it differently because this is like not having electricity, and now we have electricity, right? It's not even like the year, it's more a boolean or the the change is huge if you if we use this technology in the right way. And most of the blockers from the human element is are coming from lack of confidence. Judgment, because now in order to use these tools consistently, you need to judge in that judgment is a human skill that is not going to be covered by the AI at all. The AI has this nature that is probabilistic. I can ask the same thing to close into judge GPT and have very similar, sorry, say different approaches to that question. And the only one that will be judging that is is myself, right? So that's why it's important having this confidence judgment. And of course, have it somehow to use it because if we if we train people a very simple training, they are going to learn how to write emails. Okay, that's that's good. Potentially you're going to save a couple of hours a week, but it's not that efficiency is not becoming an unorganizational game, right? Potentially I will have I can live I can I can close my computer, you know, my love off like an hour earlier, but that's not not going to mean like workplace change, right? And another blocker is that you have technology, but the surrounding is not changing. So it's the I have the same process in that same boss, no place for experimentation or exploration. And I think that that that's producing a gap as well. If we don't change what is around that technology, right? And I would say the left thing that I can think about now is that learning is not enough, you need to apply it. And this technology and this confidence only comes after you explore apply, try and and you need some time for that as well, not to get the gains right? And and and please it's not I think I think we need to drastically change because we have been used to apply this to do a new CRM that is a discrete approach. I mean, how would you approach how to train on a CRM? This is multi-barrelable, right? And it's what need to change is how we approach problems at different levels individual teams and organization. I find myself I think in the discourse there's sort of these kind of two different visions maybe kind of they can be represented by these different research projects. So there's this project I think last year or maybe it was earlier this year. I remember that I was also so quickly. There was I think it was AI 2027, which was finding out like a trajectory for AI politically, socially, economically by 2027. And it was kind of a pessimistic, there's a high chance that this is going to end up in disaster. And things are going to accelerate so fast. And then the other side of that debate you've got this paper AI is normal technology by Arvind Narayanan and Saiyashpur, which basically says that yes, that technology is very capable, but the diffusion of that technology within organizations, the human messy bit, which is kind of what we've been talking about where like how comfortable do you feel using this technology? What is your manager role model? What do you feel permission to do? It's that kind of like messy human bit that will actually slow things down. Even if they technology is capable of incredible things, which it already is, I think, diffusing that and getting those benefits within organization. There's quite a lot of messy work that has to happen alongside that and simply kind of implementing a technology is not going to to get you to the way. If we're thinking about another sort of data point that we uncovered or are sort of finding from the analysis was that the design matters more than the hours that you spend in training. So after a certain threshold, you get sort of diminishing returns I might have time you're spending and training. What does good training look like? I would say that is the connection with real life, you know, with real use cases because it's and in order to make it relevant, really relevant for people has to be connected with what you do every day. And then support them to practice and improve. That would say in a line what a good program should be given to people in a few summary, like clear irrelevant use cases and that's a personalization part of the program. And then if you you provide like generic use cases, you know, it's going to be much more difficult for for everyone in the audience to really connect and engage. And like like I say, okay, like this hard, this is a moment of this is how can do that. And then we need practice and then we need to feel that mechanisms, you know, to make us reflect what was good, what was not good, it can be a proof. And but I think that now we did the same technology can help us to teach a technology because one of the most powerful use cases that we have been seeing using AI in the learning technology space is personalizing the experience. This is a dream we have had in the space for many years and we have doing a lot of things to try to do that and now it's possible. So using the same technology to teach how to use AI in personal analysis for each person depending on their role or the daily task now is possible. So let's use it. Just use the hammer to, you know, the same tool to teach how to use the tool. Yeah, I wonder, I mean, I think part of it is just like as you said, training is kind of always the last thing that's thought about not just when it comes to AI. I think we see this in lots of different spheres within organizations. And I think there's also often this view of training as something that, you know, you do once and then, well, that'll do the job. We've provided lots of information and then people's behavior will change. And it's really that. And people, I do wonder like to why extent because of the nature of this technology, there's an expectation that people just figure it out because you can kind of, you can kind of work with chat GPT or Claude or Gemini to ask the chat box of the LMS how to get the most use out of them. And so in a way that like implementing an LMS, you know, the LMS cannot teach you how to use it. You can actually kind of have productive conversations. So I don't, I don't, maybe that's being overly generous. But yeah, the assumption there is that like that people will learn the technology you're using the time. Yeah, exactly. I think that's true, but not for everyone. Not everyone is self-learner. Yeah. And not anyone is comfortable, you know, doing that by themselves. And not anyone has like that safety net to try, make mistakes, try again, you know. And what could happen is that you could be creating like pockets of experts within the company because you have some people that are more capable of experimenting, you know. And if you leave it, you put the technology there, there will be some people that are going to learn how to use the hammer by themselves and others need guide, they need guidelines. And that's what is happening now because you have power users and people writing emails with AI, you know, which is okay. I mean, writing things with AI, but you have to, you have to guide the thing, right? Yeah. And but that's what's happening if you leave it, just leave it there, right? And sorry, and in the levels of expectations for organizations in terms of ROI are huge, right? But that's where we see that, you know, that we are not getting those levels of expectations, but we need to support people together. Yeah. And maybe the only thing that's being communicated in some organizations is the organization has high expectations about the ROI that you're good to get. And that's where you then have that kind of human cost that Anna was talking about at the beginning where people are anxious about how they're using the technology through using it correctly. If the technology is maybe going to displace them in their roles as well, you know, there's kind of all of this stuff that if it's not addressed will obviously impede adoption or it will encourage adoption in ways that are maybe unsafe or kind of not aligned with the organization's values. So yeah, I mean, question for both of you as a sort of final thought on this, if a leader has already bought the tech, an organization has already adopted AI or purchased AI licenses, even for a small group of users, but they've done nothing else. What would you tell them to do? What's the first thing that they could do to get started? Because I think it can feel like a kind of purchasing licenses of the easy bit. It's like, where do you start after that? And I'll maybe come to you first. Okay. I mean, I'm very biased because we do it all the time. Okay, you don't know what I'm going to say. Stop laughing. I'm very biased. I'm going to say diagnose before we prescribe, essentially. I didn't think that's what you were going to say. Well, now I'm changing my assets. Just do it. Just roll it out. Tell them to just do it. No. Okay. Well, before you roll out, that's course you bought or whatever you did, AI you bought. Find out what your people are actually already doing and feeling about it right now. You know, who's using what, for what purposes, how often, and I guess most importantly, do people feel able to use it as they worried about getting drunk or being blamed as they worried about being displaced, like you said, because that's another thing to target. Are there any structural barriers? Because this kind of, if you skip the diagnosis, essentially you're assuming the problem just sits with you and please, you know, the luck skills. So you need to train them. And often the diagnosis, the diagnosis would show that there is potentially a skill gap, but there is also a different problem with the organization instead. For example, there's no time. Maybe they are unclear rules. Maybe like you mentioned, managers not sporting it. Maybe it's a culture where mistakes get punished quite severely. For instance, if the barrier is no time and you just start bombarding people with content to go through, that probably wouldn't immediately improve things. Or, for example, if people are already using unapproved tools, and you know, they're putting stuff into them, the most urgent step probably would be to, you know, first provide clear guidance on what's safe and what's not safe, and then it's, you know, encourage them to use tools more. And it doesn't even have to be, you know, like a six-month project, even though it like it to be. I feel like it, you know, a lot could be achieved within a few weeks, something like that. Longitudinal study is probably amazing. Exactly. 30 years, and you can confident that everything's fine. No, just a few weeks, you can do like a short pulse survey, maybe look at some data coming out of the existing tools in terms of usage. Maybe chat to, you know, a few managers of different levels. See what's going on. And then kind of diagnosis first is how that leader, hypothetical leader we are talking about would find out which part of the problem is theirs, essentially, and can deal with that. So that's what I think. Unexpectedly. I would say that I would potentially do two approaches. One is a very practical one that will be, okay, select a team in a set of tasks and ask them to, okay, let's apply it to this. And let's map what concerns we have, how we can change this task, this full process. I mean, it's divided process in different tasks and see where we can apply AI to improve this. And set some expectations and give them time to practice, you know, and then review what change, and then try to extrapolate that to other just cases in the company. Another approach could be just go and ask, you know, different teams and individuals within the organization. What could be the potential use cases from your perspective in your team? In teams in domains, there's infinite wisdom about their function and their expertise. And we can use that, you know, to get like a lot of use cases that we can centrally then assign like a priority in terms of level of investment and parity for the business. And then, you know, have like, okay, this is going to be the five AI first pray we're going to do during the next three months. No, it's a very practical way. And I think that the most important thing here is to share the successes because this is a cultural change, right, a change management thing. We need to share, we need to celebrate. And that's how you are going to change a culture and change a way of doing things. It's not only processes because everything ends up if people are doing the thing or not, right? And if we are doing it correctly, that's what my point of view. Nice, thank you both. We have a page on our website kind of summarizing some of the findings of analysis that has conducted and we will pop a link to that in the show notes. We will be right back for what I learned this week after a short break. So if you've been listening to this conversation and you're looking to build AI skills within your organization and realize some of the value we were talking about, then we can help. Our AI skills academy starts with a conversation. Your people describe their role, objectives and concerns and the entire program is then tailored to their unique needs. The platform generates use cases so they can immediately get started. At the remix of live sessions with experts, self-paced content and sandbox activities for practice, they build the skills they need and change the way they think about AI at work. If you have just launched AI licenses in your organization, don't let them go to waste. Get in touch with us and we can take you through how the AI skills academy helps your people save an average five and a half hours per week. To find out more, visit minetools-canew.com or drop me a note on LinkedIn and I'll get back to you. That link again was minetools-canew.com. Okay, so we'll turn out our regular feature what I learned this week, where we each share something we've picked up over the last seven days. Rodrigo, what is your name? Wow, I would say there's many things per day, but last thing that I did just in AI again is to just go to create a demo for a customer and make glue to work overnight for doing that. That's usually, it's a team effort of around, I would say, 20 apps and glue did, you know, four to five hours. That was amazing because I was doing other stuff during my day. I guess if it was working overnight. Yeah, exactly. Yeah, exactly. So it's, it's, it's, it's amazing because now we can personalize demos and for customers, you know, at a fraction of the time, you know, and be faster, but that's, sorry, all my learning is, is coming from AI stuff, but that is your, that is your job. That is your job, so yeah, yeah, Anna, what were you? Well, I'm sure when first now, because my learning was not at all as semantically appropriate as Rodrigo's friends. Neither is mine, so don't worry. Well, good, good, well, I haven't accomplished an insane intellectual fit of learning how to stop on my new bicycle. Don't, don't clap all at once. I know it's cool, I know it's impressive. I basically go to a new bicycle and, you know, how they say, you never forget how to cycle. This is true. You do forget how to get on and off the bicycle. You really do forget that. So I went to a cycle and, you know, I can start and I can go. And then I'm like, oh, I'll have to stop at some point, oh, yeah, literally, I was like, okay, how can we avoid stopping? We have to stop at some point. So, you know, you kind of press on the brakes and you fall sideways. So I literally was standing outside next to my bike with my phone out, watching a video, how to stop on a bicycle, the different process. So I was like, frx is saying, and I did learn. Is it a road bike? Like, is it one of the ones where you have to like pedal backwards and like use the brakes at the same time? Okay, I don't know about pedaling backwards. It sounds horrifying. It's a hybrid bike. I can press the brakes. Yeah. So I have to like press with them, kind of lean a little bit backwards if I'm going fast. And then it's like a whole process of one foot going down and as a food coming off the pedal, you kind of require a little bit to elegantly stop. That's not a normal bike. It's. Okay, that's not a normal bike. Yeah, I see. All right, right. Yeah. Yeah. Well, congratulations, I think. Thank you. I'm proud, personally. My way I learned this week is, I was listening to a podcast from 99% invisible. You have referenced that multiple times on this podcast. Great show. Yeah. The episode is called Doctoral Esperanto, which was the pseudonym of LL, LL, Xamenhof, who was the creator of the language Esperanto, the universal language, which, you know, the goal of it was to foster world peace. The idea is that if you had a simple language, everybody could learn and everyone spoke the same language. There'd be less misunderstanding and, you know, solve the biggest problem that the world faces. The thing that I learned specifically was that the Vatican apparently still issues three radio broadcasts per week in Esperanto, which maybe I'll dig out an episode and drop a link to that in the show, in case you're interested in picking it up. I think you will then go also offers Esperanto as one of their languages. So that's why I'm in this week. Nothing to do. Now I'm curious about the engagement levels of the Vatican. Vatican. Yeah. How many, how many listeners? Exactly. Yeah. Yeah. That is a great question. I mean, maybe a few more after there's a day's episode. I'll see. But that's it. That's all for this week. If you'd like to share your thoughts on anything we said on the show, you might as well make it in for more from us. Head to mindtools-canada.com. Thanks for listening and bye for now.