LH #120 Agent Provocateur with Dr. Ashwin Mehta

This podcast is supported by World of Learning Summit. This leading L and D event returns to Olympia, London on the 3rd to 4th of February 2026. Two days to learn, discover, interact and network with the best in the business. Register now. It's free@learnevents.com this podcast is supported by Synthesia. Training Videos shouldn't take weeks to make Synthesia lets you turn text into professional multilingual training videos in minutes. It's fast, scalable and so simp well. Anyone can use it. You can write an email, you can make a video.

Hey there, my name is Alex and I'll be guiding you through this training session.

70% of Fortune 100 companies use Synthesia to modernize onboarding, simplify compliance and scale global training. Podcast listeners get an exclusive 30 day free trial. Go to synthesia IO. Greatminds don't miss it. This offer won't last.

People will talk about hallucinations as if it's a thing, right? It's only a thing. When you go directly to an application, ask it for an output, and take the first output that you get and walk away with it. If you're not scrutinizing what's happened with that output, if you're not scrutinizing the references, you're not going to read the references, then to some degree you're going to get errors.

Welcome to the Learning Hack, a podcast about the people and technologies that are creating the future of learning. I'm Joe Helm.

Now guess what? Learning is cool. Learning is cool. I'm learnding. Learning is fun and knowledge is power. Knowledge, education.

Learning. We hear a lot about AI agents nowadays. Recent research from Don Taylor and aigler Vinauskeiter suggests that some people talk about using AI agents when they actually mean AI assistants. And a lot of people in conference keynotes will throw around the word agentic, and sometimes it's clear they don't really know what they're talking about. So I wanted to talk to somebody who really does know what they're talking about on AI and bring a bit of clarity to this whole area of AI agents for learning hack facts.

Dr. Ashwin Mehtar is Founder and CEO of Metadology A and also Chief AI Strategist at the Learning and Performance Institute. With a background spanning data science, psychology and digital transformation, he advises global Enterprises on integrating AI into workforce development and capability frameworks. A frequent conference speaker and thought Leader on Agentic AI, Dr. Mehtar is known for his practical insight into how intelligent systems can augment human creativity and learning. He previously held senior roles in ed tech and corporate learning innovation across Europe and Asia.

The dream we've been sold as learning professionals is that we can get hands on with Genai and do clever things with agents that previously we wouldn't have thought of attempting without hiring in some coders or reaching out to it. But is that just a dream? This conversation with Ash should really help to clarify your thinking on those points. Certainly did for me. What are the vistas that AI agents can realistically open up for you? What's available to you now or in the short to medium term? And when and how are you going to have to make use of specialist expertise? Hello, Dr. Mehta. It's a great pleasure to have you on the learning hack, Ash. I should say we know each other pretty well. Welcome to the show.

Yeah, thanks for having me.

So, to start with, can you tell us a bit about your role at the Learning and Performance Institute? What does a chief AI strategist actually do?

Yes, so I'm the CEO of Methodology, which is an AI and tech consulting firm. And as part of that I do some fractional work for different clients. And the LPI is one of those. What I do with the LPI is partly, it's demystifying the AI revolution for the learning industry, the LPI member base, and making sure that the offerings that are going out to the LPI member base are grounded in what I would usually say is some kind of thought leadership in the AI space, which is my specialty.

And what's your own professional journey? How did your background lead you into this fascinating intersection of learning and AI?

Yeah, so there's, there's the journey into learning, which is. I started off in chemistry and I was moving from chemistry to chemical consulting to business resilience in the medical research space. And I started working in Sub Saharan Africa and I was looking at business resilience broadly and part of that is workforce resilience. And that got me into the digital learning space because using elearning back then in the area that I was working, which was the Gambia, that was difficult. It was difficult because of unstable power and broadband. It was difficult because of lack of computer labs and digital skills and device availability and all of, all of these things. So I got a very holistic introduction into digital learning. We started on the elearning journey and the LMS journey and sort of social, social networks and that kind of thing. And after that I wanted to pursue that more. So I did a, an MBA and a PhD in technology adoption and that is statistical Modeling of human behavior at the technology interface. And there I was looking at structural equation model. I was trying out different techniques for analyzing data and looking at patterns in data. And I stumbled across machine learning and neural networks back then, which were not that far removed from what I was doing in the structural equation modeling space.

What year would this be that we're talking about?

So this, I started these projects in the early 2010s.

Okay, right.

So I continued probably for six or seven years, not only doing the academic stuff, but also embedding and implementing tech technologies in various countries in Sub Saharan Africa as well as the uk. And that gave me a fairly broad taste for digital learning. And from then I moved into big four consulting and I was leading learning, tech innovation, things like that for various clients. So I was client facing. And after that I went into the pharma sector and did the same, the same thing, learning tech, digital learning and looking at data and things like that. But throughout I was maintaining this thread of touching on tech. So the tech side of things was evolving. We had not only the machine learning, neural network side of AI, but we also had things like blockchain were, you know, coming to the forefront, digital twins, the Internet of things. So I was doing quite exciting projects in various spaces like that. And then the generative AI revolution came upon us. And it wasn't too much of an extension from what I had been doing. So I got into that a little bit. But I got into that from a practitioner point of view as well, because I was leading an AI team at the time. So that allowed me to do quite a wide variety of things with what was emerging in the generative AI space. So back then you had not only the ChatGPTs of this world, but you also had things like avatar tools were coming out, wrapper tools were coming out as well. It was possible to do some data analysis in the skill space. So there were quite a few interesting projects that we could get off the ground using AI. I did that for a couple of years and then I started my company, which is an AI and tech consulting business end to end, looking at not only strategy, but also governance, tech development and data science.

Little bit of jargon there to unpack. Rapper, you're not talking about hip hop there. What's wrapper?

So wrappers are where a company will launch a new product, but it's basically chatgpt or something similar with maybe a front end on it. So it's not doing much in the space.

So another thing that we hear a lot of about at the Moment is agents. I remember we had a kind of brief discussion about this at Learning Live. For listeners who might not be familiar, what exactly is an AI agent? I mean, most of us will know this, but some people are not entirely sure. Even some people who use the language perhaps. And also, is there a difference when people talk about agentic AI, is it basically the same thing or when I get into talking about that area, yeah.

So the key word is agency. Do you have tech in front of you that can do something in a way that is autonomous? If you can answer that question, then you can start to make the distinction between assistants and agents. If you're using AI in workflows, you're using AI in software tools that can take action, that can make decisions in a way that does not require human oversight. So this might be where your decisions can be rule based or they can be effective code structures or algorithms that will take decisions and then take actions in applications. That's where you move towards this area of agents. And the typical example in the super agent space might be that you want to book a multi city tour. You say, I'm going to these five places, book me the hotels, book me the flights, here's my budget, here's my credit card, go and make it happen and it will just book all of that stuff for you. Now generally speaking, people aren't using those things yet. Many reasons. Capability of the tech is one, but trust is another. You know, how much do you trust that these things are going to go well? So this is still an emerging space. What most folks are doing is looking at assistants. So that is where you have a version of a large language model capability. So this might be a copilot of GPT at Gemini or a GEM or something like this.

Yeah.

Where you say here's a structure of a task, here's a structure of an activity. You're going to get some kind of input from me or from somebody else. And with that input I want you to take a decision based on these guardrails and maybe take an action in a particular set of applications that's much more realistic and that's probably in the semi agentic space. The capability of the tech and the capability, when we think about tech capability, it's not only what the tech can do, but what it can do when it's applied. And we have to think about governance, we have to think about IT infrastructure and systems in companies before we can start to think about can we really execute something that is agentic. So my view on this is most companies are using probably assistants, if they're even there yet. And the reason I say are they even there yet? Is that many of the conversations I will have in the learning and development space and in the human capital space more broadly, maybe even in the marketing space, we have people who are having a one to one interaction with an application. So a person goes to ChatGPT, a person goes to Copilot, puts in a prompt response, it could be voice, it could be text, asks for some kind of output and gets that output. And that's a one to one relationship between a person and an application. We really need to be thinking about moving towards workflows if we're going to get something that's repeatable, so that is based on input data that happens regularly. Every day something comes in. Every week there is a request. If we think about the learning space, every week there might be a request from a subject matter expert for some digital learning. What do you do with that in a structured way? And that starts to move down the agentic workflow conversation route. And that's very different because at that point you move away from a person having a one to one interaction with an application towards multiple instances of those reasoning models, being able to take actions within a chain of events, which is a process.

And to what extent is the technology there? I mean obviously the users aren't quite there, but it's the technology there. I mean my own kind of experience, this is fairly limited because I'm not a software engineer, but I have projects set up in ChatGPT for regularly doing stuff with the podcast. I've asked it about whether it has links to certain of the platforms I use, like the one we're recording on now, Riverside, and it doesn't. Do we need the provider OpenAI in this case to have integrations with external platforms? I mean I have heard that you can kind of make this stuff happen via things like Zapier and other kind of third party intermediaries, as it were. But where exactly are we as far as the kind of non software engineer, you know, domestic user of ChatGPT is concerned? Who wants to get into more agentic stuff?

Yeah, it's a good question. And I think the tech is broadly there in the consumer space. So in the consumer space you have platforms like make and N8N which are being used fairly widely in the enterprise space. I think Microsoft, Google, all of these OpenAI, they have some emerging capability in this space. So Google has Node, for example, which is released in the US but not in the UK yet. The Copilot studio space also has capability here and OpenAI have recently released the Agent software development kit. Anthropic have a software development kit as well. I think they released last week with Sonnet 4.5. So there are capabilities here where you can start to link different tools together. Now the capability on the agent side, on the software development side. So if you think about these big software companies, their capability is one part of the whole landscape. We also need to think about the capability of third party tools to integrate. And the integration is a handshake between two different applications. So not only do you need the platforms, the Microsofts and the OpenAI's of this world to have the integration capability with their tools, their models, et cetera, but you also need third party tools to be able to do this. If you think about. So you just mentioned the platform that we're recording on, which is Riverside, they don't have connections. I use this platform. The connections just aren't there. So when you record something like this, you then need to export the different variations that you're going to use, put them somewhere, Google Drive, OneDrive, something like that, and then try to retrieve those into a workflow. If you want to automate your social media, for example, you want to automate moving them into the Instagrams of this world. Now most of these platforms, think about Google, think about Meta, they have maybe even LinkedIn, they have developer portals where you can go and register an app and you can start to post things to your feeds. The avatar space. I think most of the tools that the L and D market is using tend to be around content creation. So you have avatars, you have images, you have videos, you talk about things.

Like Synthesia for instance.

Yeah, Synthesia, Colossian R1 hey gen. So I use hey Gen. The API is very, very easy to use. I don't necessarily use Synthesia because that's. I think Synthesia and Colossian are more focused on having learning designers or developers work one to one with the tool. And while every company needs its own niche and market, of course for me I'm pretty much not going to do anything in the AI space unless I can have an automation or workflow do it for me.

Right, Okay. I mean it is the case. I've noticed that all platforms want to capture your entire workflow. And as we move forward, this has all been, I mean it's not like kind of the way the Internet was initially developed by defense and government money. This is all being done by companies and they kind of want to keep you within their universe, as it were. You know, whether it's kind of Microsoft, Apple, whoever, no judgment. I quite understand why they want to do that. But does that dynamic slow the development of agentic AI, do you feel?

Not so much the development, because think about who's developing it. It's the big companies that are going to be developing these kinds of things.

Maybe I should make a distinction here between kind of consumer use because the growth of things like ChatGPT have got us all used to this idea of oh, you don't need to be a software engineer, you don't to be a developer. I can do this just by talking to this device. But it seems to me when you get to what you're talking about with agents, you're talking about third party intermediaries, APIs, it's beginning to get out of the kind of consumer use sort of universe into one where you actually need to get someone involved who's an expert like yourself to do it.

So there are two ways to think about this. And honestly, John, nothing annoys me more than seeing on the media. And when I say media, I mean every platform, every social media, all over the Internet. These, these kinds of messages where you say, you know, no experience necessary, no skills necessary, you can click a button and you can achieve magic, right?

Yeah, yeah.

I don't think in my tender 40 something years, I don't think I've seen anything in my life that didn't get better or didn't require a little bit of work. Once you do the work, you get a better outcome. And this idea that you don't need any skills, you don't need experience and you don't really need to put in any work if you want to get magic. I think this whole idea is basically ruining society and ruining work hit a nerve here. There's an expectation that, you know, I can just buy something in, click a button, there you go, magic is done. And, but let's think about that for a second because there's this issue of democratization in L and D where we say we can put these hands in the, we can put these tools in the hands of experts. And by putting the tools in the hands of experts, subject matter experts, not learning designers, not graphic designers, not anybody who's worked in creative media, but maybe someone who's working in the business and we're going to get something that's amazing, pedagogically sound, visually impactful. That kind of idea seems to remind me, it reminds me of the days of eLearning about 15, 20 years ago where you'd say, everyone can produce this stuff. They have all of these really easy to use authoring tools and you get a bunch of elearning courses which were basically bullet points on a slide with the next button.

Yeah, SMEs can build courses was the.

Yeah. And. And I think we need to divorce ourselves from the idea that we can get magical outcomes without having to put any work in. And this takes us back to what you were saying just a few minutes ago, where for me, a reproducible and robust workflow that gives you structured outputs is always going to be better than somebody working one to one trying to produce a single output. It's about scale. But when we start to think about scale, we have the conversation around accuracy, we have the conversation around hallucinations. And this is something that we see nearly all over the place. You've been to many conferences and so have I. People will talk about hallucinations as if it's a thing. Right now, it's only a thing. When you go directly to an application, ask it for an output and take the first output that you get and walk away with it. If you're not scrutinizing what happened with that output, if you're not scrutinizing the references, you're not going to read the references, then to some degree you're going to get errors. And there have been a couple of cases recently where errors have been in client outputs with big companies in the bin and the media. And there's a little bit of, there's a part of me that says, you know, put the work in, use the tools, of course, but put the work in to make sure that your outputs are sound. There are plenty of parameters that can be tweaked. When you're looking at workflows and you're using reasoning capability. So sampling temperature is one of those things where you can say, you know, if you're going to retrieve from these documents that I've just given you, dear Agent, dear AI entity, retrieve from these documents for me, but make sure you only retrieve what's in these documents and make sure that it is accurate. Yeah, you can tweak parameters to make that happen. But this kind of thinking, what I'm talking about, where you have third party tools, you have assistants, you're tweaking the parameters, you're maybe even fine tuning your model. These kinds of things don't tend to happen in enterprises. They don't tend to happen in certainly in L and D, where the predominant focus for L and D is to take a point solution from the market, a thing that does one thing rather than a general purpose tool and then try and make that output stick in the organization. So it starts to become tool driven design rather than design driven, which is to me it's kind of the wrong way to think about it, but the transformation space in L and D and pretty much every other sector as well, so marketing for example, sales, the transformation space is nearly always going to be about can you retrieve your company data into something that has agenta capability and make that somehow employee or learner facing so that they can effectively self serve.

If you're looking to connect with the top minds in learning and talent, there's no better place than an IVENTIVE event that's I V E N T I V. Since 2009, they've hosted nearly 300 interactive conversational events exclusively for senior leaders like Global Hitler Learning. I've been a sponsor myself and found these events invaluable. Less about sitting and listening, more about reading. Real impactful conversations. It's an unmatched opportunity to collaborate with leading chief learning officers on the challenges they're tackling today. Join the conversation visit iventive.com learninghack that's iventive.com learninghack it's what you're saying, in a way, without necessarily saying it, is that we, we do need specialist skills to be involved in this, to make it work properly. In other words, as well as I do things that people say at conferences. One thing we've heard is that generative AI is the revenge of the arts graduates. AI used to be a really math C thing. Well, no, it's not now. It's an event in language, not an event in mathematics. But I think you're kind of indicating that really we may need to sort of get that person down the hall who knows about this stuff properly to come and make it work for you. You're not going to be able to do it just by kind of, you know, talking to ChatGPT.

So having some code knowledge is going to be very beneficial. But it's not the only pillar of skills that folks need in the future. So if you think about what I've just said, you have the agentic world, the ability to structure workflows, processes, and with those things, both of those things together need the ability to retrieve data. So you've got effectively those pillars. And those pillars need to serve some kind of human outcome in the L and D space, of course, or in the HR human capital Space we're looking at how do we make the people side of a business better, generally speaking. So that gives us kind of the pillars that we need to look at. We need to look at tech. So this is the AI workflows, the agenta capability and all of these things are broadly speaking tech. We need to look at our ability to use data or to retrieve data within the tech space and then we need to think about what the people focused outcomes need to be. So that gives us three pillars of skills. And whether we're talking about HR broadly or we're talking about learning and development, those three pillars of skills are going to be those three things. So data, tech and people. Now my fear in this space is that when you have industries like L and D historically have been people focused and have not really been tech or data focused, data comes up as a challenge. And it has come up as a challenge certainly for the last 10 years that I remember where folks will say, well we don't know how to measure the ROI of our interventions, we don't know how to scrutinize the data to make data driven decisions. We're struggling under the scrutiny of proving the value of what it is that we do. And to cope with that we're going to measure what happens within the department. And that's broadly speaking not good enough anymore. Folks need to be looking at what the business impacts are of the things that they are spending money on. And that requires a data focus. And that's difficult for this particular industry generally it's not difficult for other industries, but here it's very difficult. And I could. So I've worked in a lot of different industries now we think about health and safety industry. You know, they're very, very used to looking at rolling averages of accident statistics and looking at near misses and looking at risk profiles and things like that. Look at business continuity, they're very used to looking at time based recoveries, they're looking at impact analyses and they're used to doing that. You look at marketing, they're used to looking at A B testing and they're used to looking at engagement statistics and thinking about the top of the funnel. You look at sales, they're used to looking at deal sizes and sales cycles and all of these things that that matter to the business. And then you come to L and D and they say, well how many people did we get on this course? How many people clicked through this particular elearning and were they happy with the sandwiches when they went on this away day? And that's broadly not good enough in the data space.

Yeah.

So my, my fear is when I say that we have three pillars. We have data, we have tech and we have people, we're going to get a lot of the industries that will say, I will focus on the people because this other stuff is difficult. And there are two implications for that. So the first implication is that we're going to get a flood of supply when it comes to things like coaching, mentoring, the people side of life. And that's also going to start to infringe on other disciplines. So change management, for example, I've seen a lot recently on LinkedIn where they'll say, well, change management, L and D people need to focus on change management. They need to have these skills around managing change. And my immediate thought is really, do you. Because there are change management experts out there and is this just a case again of people who have an instructional design background not being secure in what they do and starting to bleed out into other disciplines and therefore do a lot of different things badly? So there's going to be a supply side implication where you've got lots of the industry pivot towards people, things where they don't really have the skills. And then the other impact is going to be that the data and the tech side of things is going to be effectively devolved to people who have data and tech skills. So this is IT departments, supply chains, maybe even the consulting side of things. Which means that there's a loss of control in that space from the L and D sphere, right?

Yeah. There are things I could pick up on that there some learning. People say that all learning is a change. So it's right that learning people are involved in change and change management methodologies and the literature behind that. But it is interesting what you say. I want to talk a little bit more about how L and D changes later, but first of all, why, While we're on the subject of agentic AI and agents, so I had a word with Ed Monk, who you know well, he said you told him after working with him for a week that you had 10 direct reports and none of them were human. Could you explain what these AI agents do for you and how you've gone about setting them up?

Yeah. So I started this year with the glorious ambition of having a headcount of 50 people in my business with no.

People, no people at all.

That was the ambition. I now have a person, I've hired a person. But the thing that I was looking at at the time was I had a couple of agencies working for me in my business and I was looking at examining what it was that they were doing. Now a lot of what they were doing was process based throughput. So if we partitioned the activities of a particular role and anybody out in the industry, you can go and do this for your, your roles, whether it's someone who's doing learning, needs analysis or design. These are rule based or process based activities the folks are doing. And when you think about that, how much of the activities that one role does can be partitioned into a reproducible and repeatable workflow where maybe the inputs vary and that makes the outputs vary, but the process of getting from inputs to outputs is basically the same every time. And that's what we call a workflow. How many workflows do you need to do the activities of one particular role? Now over this year I've built lots and lots of these and broadly Speaking, it's between 8 and 10 workflows to capture most of the automatable, or not only automatable, but things that can be done with AI. Most of those activities in most roles can be partitioned into about 8 or 10 workflows and thereby you have agentic workflows doing most of the activity of the person. Now what does that mean? The implication of that is that there are things in marketing, in sales, in. I don't have a learning department on my business, but marketing, sales analysis, for example, deck structuring, you know, making decks, making PowerPoints, doing basic research on some of these things. A lot of these things can be done with a workflow to the extent of about 80 or 90% requires human oversight. But that's therefore why I have myself and I have recruited a person who is kind of unique in the industry because her role is working and overseeing an AI team. So that's how I've structured my business to be AI first. And I can do that because my business is less than three years old. Companies that are more than three years old, they have to cope with legacy systems, legacy skills, structures and all of these things which are slightly different. But that's the approach that I've taken and I'm not alone in taking this position. Many big consulting firms are doing the same thing.

All right, and how does that work in practice? I mean, do you have like an ecosystem of different applications and products, platforms that you use for each workflow and do you have a kind of overarching control platform or app for that stuff or do you directly code it? How do you go about doing it?

So it's a mixture. So I use. I'm fully agnostic in the platforms that I use. So I think the distinction between my company and an enterprise is that enterprises tend to be locked into a particular environment, Microsoft or Google, for example, so they only really use the capabilities of that one platform. For me, I'm agnostic, so I will use a wide variety of platforms, not only in the workflow space, but also in the data space. And this allows me to retrieve the wealth of information that I've already generated myself, but also to retrieve those in the reproducible way. Now, of course you've asked about coding. There is some coding requirements for some of these things. Not only potentially understanding things like Python, but also understanding JSON. I think everybody who's working in prompting at the moment should understand JSON format outputs and go and research that and look it up. But if you can do that, you start to really hone in on the reproducibility of things. And then there's also trusted sources and being able to validate those sources, which requires quality control. So I think between these handful of things, there are platforms in the middleware space, there are data platforms, of course, which we all probably know about. So AWS is one, Google has a cloud, Microsoft has a cloud, and there are other more niche technologies that will allow you to do things like this, Pinecone, Supabase, et cetera. Anybody who's looking to code, you know, try and pick up Python, think about Visual Studio, of course. And then in the workflow space there are a wide variety of platforms again, so I work directly in Azure. Sometimes we have the Amazon capabilities as well, we have Google Cloud capabilities. And then there are consumer platforms, if you want to get started in some of these areas. So make N8N sapien these kinds of things.

So it's clear from what we're saying that as teams begin to integrate AI into their operations, especially agentic AI workflow, re engineering will become an increasingly important area of consultancy. You talked there about difference between you starting off your company from scratch and an enterprise which has a kind of established department that does things with roles and people and allocated skills and so on. And that if you're going to bring AI into it at some point that has to be re engineered and it might mean changing quite radically who does what and how it gets done. Do you agree that this is going to become an increasingly important area of consultancy? Great for people like you. What kind of changes should organizations be preparing for?

So this is going to be a big thing over the next Couple of years. Because I guess you've heard of multi agent frameworks. So the idea that you have lots of agents that are doing different things as you just described, you know who, what's the hierarchy of control there? And having master agents that decide which agentic workflows are going to be activated depending on what things. That's basically the principle of agent framework.

So you can have agents within agents, as it were.

So think of it like an organizational hierarchy. You might have a director agent that has multiple manager agents working for it. Each of those manager agents might have workflows that invoke third party tools. So you might have a content development function, you might have, as I was saying, a marketing function or a sales function that do different things. And these form a hierarchy and that's a hierarchy of agents. Now the obvious implication is that we need to be thinking about our people structures in a way that mirrors or matches or has synergy with the agent structures. Because the future of most organizations is going to be some combination of these things working together. And the skill in the future is not only thinking about what process re engineering looks like, but also thinking about what organizational restructuring looks like in the context of the technology that we have.

And is there a danger with this, that as all these workflows get re engineered that the system, the organization itself becomes a bit rigid? I mean, because if you want to introduce change into that system, and God knows there's been a lot of change in marketing in the time that I've been doing it, then the whole thing has to be unpicked and started again from scratch or radically changed, doesn't it? And then that's another re engineering of the engineering which people might not have the money or the will to do. Does the organization become more rigid?

So it depends on how you configure it. Now every organization is going to configure itself differently. If you think about people hierarchies, even if we forget the world of tech for a second. Has anybody watching this? Have you ever worked in an organization that was just a people hierarchy where everything was rigid? You had multiple levels of authorization for every decision that needed to be made, you had to go to this other department for approval, that other departments had to be involved.

No, because there's masses of tacit knowledge in an organization that keep it running. I mean, this is partly my point is that with technology, this type, everything has to be explicit and kind of written down. And that there is a danger that tacit knowledge disappears. Is there?

So potentially it depends on how you structure your organization. So I have, as I'm sure you have, over many, many years, witnessed a lot of organizational restructuring that were, that was based on basically flipping a coin. And you know, you lose a lot of good people because people who have options and have skills will go somewhere else. Now, when you're doing this kind of restructuring, how do you optimize your human, human hierarchies for things like tacit knowledge? You probably want to do that alongside something that is a knowledge management or a knowledge capture exercise, because you don't want to lose that stuff. And we're generally speaking, not that good at capturing tacit knowled. We can do that more and more as we start to move into the, let's call it the agentic space. But I'm really talking about voice agents where you will have a conversation back and forth with something that's effectively a chatbot in that space. The potential for knowledge management and knowledge capture, data capture from conversations is huge. So as we move into that space, we're going to really start to see an increase in how much of this tacit knowledge can be captured in databases and a decrease in the reliance on folks who have that tacit knowledge, which means that the risk is to some degree mitigated. But also responsible organizational management should capture some of this tacit knowledge. When you're thinking about restructuring, you're not thinking about restructuring for headcount. You're thinking about optimizing for your organization for the future. The different things.

Okay, that brings us neatly along to what I'd like to do now is to go through some of the negative attitudes towards AI that are coming out all the time and kind of evolving in some quite interesting ways. How do you help people address common concerns such as this in your consulting work? And let's take a few of the one by one. Number one, it's just a cost cutting exercise to replace humans.

So sometimes it is. Now, I've witnessed this a few times out in various case studies that I have not been involved in, just to be absolutely clear. But I've seen organizations who will say, right, we're implementing AI, so we're getting, you know, we're cutting 7,000 jobs, 8,300 jobs, whatever the number is. And they'll go and do that. And when you look under the hood, what they're actually doing is they're offshoring and they're using AI as kind of an excuse, a panacea, to say this is justified because we can change what we do. On the flip side, you do have companies. One of the big tech providers, I can't remember who it was, implemented an HR chatbot and they were able to reduce a number of their roles. Is that, is that a cost cutting exercise? Is it a headcount reduction exercise? Or were they able to get the value out of roles that are no longer needed? Because we really need to scrutinize the roles. If there are roles out there where value is not being created, this is the nature of business. If roles are there where value is not being created, either because there's a new way to do those things or, and this is important, right. Or those roles were never creating value in the first place, but the scrutiny wasn't there to figure that out. And because of the presence of AI, now businesses are looking at those roles with a different eye. So both of those scenarios could be true. And if value's not being created in a role that it's legitimate for a business to remove those roles. Is that a cost cutting exercise or is that an optimization exercise?

Interesting. Number B. It's filling the world with AI slop.

So there's a lot of garbage in the market a lot when it comes to the supply chain and there's a lot of garbage when it comes to production. So think about. I've given this story before, I'll give it again. I was in Kenya on, I think it was Kenya. I was working there for a little bit and I went on safari because why not when I'm there? And we had a bus of 25 people and we were going around the safari park and we found a rhino and 25 people all with phones, taking, I don't know, 20, 30 pictures of this rhino over and over again. Two tours a day, seven days a week. How many pictures do we need in our data sphere of this one particular rhino from every angle? We probably don't, but we have right now in the AI space we have lots and lots of people's. I think there's a phrase about monkeys and typewriters and you know, it's writing.

King Lear or Hamlet.

Yeah. So we have that kind of scenario where you have, in every organization folks are going to a co pilot, a GPT, a Gemini, something like that, and asking basically the same handful of questions and getting a response. And most of those responses are either think about normal distribution, many of those, the responses are going to be poorly prompted. The responses are not going to be great. Folks are going to be using AI at whatever their skill level is. So we have a lot of production that's happening. A lot of that production is not going to be of superb quality. We have a lot of supply chain that is working on wrapper technology where you're not getting anything that's really outperforming the GPTs and the Geminis of this world. But maybe it's got your logo on it and that makes you feel that you need to go and splash out and get whatever response you're getting. So from both of these angles there is a reduction in quality of what you have. But really there are a handful of big models, there are a handful of big technologies and optimizing for your use case is probably going to be the best way to get get around this idea of slop or garbage.

I think you've answered the question there, but I think what you're saying is you need a bit of quality control there, optimize it so that it creates the value. But don't you just end up with people saying, well we need a load of images here and they don't really have to be that brilliant. So you do end up filling the world with AI slop and everybody just tolerates it. This reminds me of the debate in music about AI bands. There was an AI band released on Spotify recently and everybody got very up in arms about it, was complaining about it and I had a listen to it and I thought, well this is bland, boring rubbish. But most of the stuff I listen to like this is bland boring rubbish anyway. So what's the difference if it's cheaper? Personal prejudice. Next one. It reinforces existence. This is a favorite of L and D. It reinforces existing biases as non fairnesses.

Yeah, maybe. Does it, I mean, depends on what we're talking about. Right. If it's the favorite of L and D, then what are they really talking about? Are they talking about image generation? If we go moving slightly outside into the HR space and we talk about ATS systems and we might be talking about hiring decisions that are based on the demographics that you've always hired in your particular business, then of course there is likely to be bias there. There is also of course bias in model training. You know, we have the. If we go back to November 22, GPT was trained on the Internet up to 2021, which was produced mostly where probably in America. Right. So we have a reinforcement of cultural bias. We have probably a reinforcement of the language bias as well. If you use a translation app at the moment for your courses, your multi language courses, do you know if that translation is happening by going from the English Base or do you know if it's happening somewhere else? Right. What's the language route that's happening in your translations? It's worth figuring it out. But you have all these biases in the system. Of course, the challenge to LMD folks is how much do you mitigate that in your use of the tools. So you're prompting your structures, your instructions, your parameter control versus how much does it really matter anyway if you're asking for a picture of a guy in a blue suit in a corporate boardroom because you want to put an image in rise course, frankly, how impactful is that anyway? So it's worth thinking about this in the context of what you're actually doing.

Okay. And lastly, if we build AGI, It'll kill Us All. And it's almost the title of a recent book by a very interesting writer whose name I'm afraid I can't pronounce or remember. But you know, the scenario, it's, you know, it started with Nick Bostrom, I think, you know, super Intelligence will Kill Us All.

Yeah, it's kind of the Terminator scenario, right? If it, if it, if we have something that's generally intelligent, will it decide that human beings are better, the planet is better off without human beings or something like that, right?

Yeah.

Well, there's two challenges to that, right. One is maybe we as human beings should just try and try and be better, more valuable, right. And as a species, do we really do everything that we can do to be the exemplars, the people who really should exist?

That's existential. Well, no, we don't. And we're probably not going to get any better. I mean, we're probably getting worse, rather better at that kind of stuff. So it still leaves the question, you know, given the imperfectness of humanity, is this leaving us open to a kind of curly quotes, superior intelligence to just decide it'll wipe us out?

So there's a big if there, right? And the big if is, is, are we capable in our technology space of producing something that is sentient, something that can take action, something that is agentic to the extent that it could make decisions and decide on things that are against humanity and you know, Asimov's Rules, probably we would be coding against that. But if we did reach a point where we wanted something that was free to make all of these decisions, that's probably going to be. I don't want put a timescale on it, but I'm not sure I'll be here for that. Right. Because the tech isn't it's not moving quickly enough. Despite the hype, the tech isn't moving quickly enough to produce sentience. And the natural language processing part of this is around. We've gone from a space where we're able to predict the next word in a sentence, which is nlp, and where we've produced sentences that make it look and feel like a person, person wrote it. And that can be moved into the audio space, it can move into the visual space as well. Going from that to something that is generally intelligent enough to make decisions for the broad base of humanity. To me, that's a massive leap and I don't think I'm going to see it.

Yeah, that's a lot of skepticism there, which is in a way reassuring. And I think I'm kind of with you that. Different question entirely. You've said before that L and D isn't a single job, but an ecosystem. How does AI reshape that ecosystem? So looking at the L and D function particularly, where do you think the biggest opportunities and the potential pitfalls lie for learning professionals in this world?

So the biggest opportunities in this space are around optimizing for humans, optimizing data for humans, to put it more succinctly. And that is something that we could have effectively agentig middleware doing. Most of the things that L and D historically have done around data curation. So if you think about L and D as a kind of data curation reformatting function, you take data from somewhere in the organization, you reformat it, you summarize, maybe you put some images or whatever it is you visualize, but effectively you reprocess that data so it's palatable for humans. AI basically does all of that stuff now, which means most of that function is not required, which is why there's a lot of angst in the industry, because most people were basically doing that stuff. But if you think about it more holistically, data curation, optimizing that for people, what data needs to come in? How clean is that data? How on point is that data in terms of messaging, in terms of accuracy before you feed the AI engine? I think there's a huge opportunity there that nobody's really doing in any organizations. This isn't really part of the IT space, it's not part of the HR space. So to me that's a massive opportunity that that kind of data curation piece. The second part of that ecosystem, of course, is the tech space. How do you optimize tech ecosystems for people? IT folks know how to Optimize tech ecosystems for other factors, for value, for cost, for security, for simplification, and all of this stuff, integrations, mcp, API. Right, all of these things. But optimizing those for people, human hierarchies and multi agent frameworks, et cetera, that kind of conversation. I think there's a space there for L and D folks more broadly HR probably than L and D. And the third is this optimization of the human experience. So. So for too long. I'm gonna get on my soapbox for a second. I think for too long L and D folks have been satisfied with producing poor quality eLearning courses. AI could do that stuff for you if that's what you really want to do. You want to produce a thousand videos, there are tools that will go and do that. You want to produce a bunch of elearning courses, there are tools that will go and do that. You can structure workflows. That stuff doesn't need to be done. And that's the stuff that doesn't add value whether you're in house or whether you're going to an agency and buying elearning capability. This stuff is just a dead market. It doesn't need to be done. What does need to be done is this is the stuff that's always been talked about at conferences. How do you structure learning experiences that really change behavior to have an enduring and lasting effect on people, on their worldview, on their capability to do things? This is the concept of development. And in that space, I've not seen anybody operating in that space for the 14 years I've been in this industry. This is the. Who's the next Hitchcock of storytelling? Right? Who's the next Hideo Kojima of game making? Right? These are big, big questions. And if you can have an experience, a learning experience that actually impacts somebody, and I don't mean that they went to a course and they said, yes, I'm going to do these three things afterwards. Right? That's not the same thing. If you can architect experiences that are impactful for behavior, I think that space is still a massive opportunity for L and D folks who claim to do this stuff for a living and yet spend most of their time on lower value stuff. So if we can pull ourselves out of those weeds as an industry and refocus on people experiences using all of the tools that we have to create those experiences. But it requires something that's creative or something that is in the storytelling space, something that is the human, in the human behavior space. And I think that that opportunity is huge. But I think the skill gap is also huge. So that's going to require a lot of work to get to that point.

Strong message there. And finally, after all this serious talk and strong messaging about AI, what's your go to Karaoke favorite?

Karaoke favorite. So I've probably got three, right. First one is Warm Up Software song. As you can see, I. I sing, I've sang for many years. Warm up song to test my voice is always superstition. Stevie Wonder.

Yeah.

Then I'll play around with a few kind of funk stuff. Play that funky music is a good one. Stretches the voice a little bit.

Play that funky music. White boy.

Yeah, I gloss over the white boy.

Thing, but anyway, you'd mess that up. Who's that by? Was it Wild Cherry?

Wild Cherry, yeah.

Yeah.

Then I usually do. I usually do a little bit of just random singing with people. A couple of Disney songs. Right. You know, Aladdin's a good one. And then I try and finish on the final countdown as a duet with somebody. So, John, the next time we're karaoke in, that's what we're gonna do.

Well, I've gotta make some kind of cheesy link now about the final countdown as we come towards the end of our time together, which for me has been absolutely fascinating. Thank you very much, Ash.

Thanks for having me.

That's all on the Learning Hack podcast for this time. Many thanks to our guests and to our sponsors. The Learning Hack is completely independent and transparently funded by sponsorship. It is produced by John Helmer and hosted by him and also me, Mariana, a Synthesia, a AI avatar. If you want to help others find us, please like follow, rate, review and subscribe wherever you get your podcasts or on YouTube. Until next time, stay curious, learning people.

Now I finally get it.

The Learning Hack podcast is supported by Learning News, the learning sector's news wire. Rob and his team are good friends of the podcast and we really value the help and advice we've had from them and they do a great job. For the very latest news from from around the learning sector, for interviews with learning leaders, the latest from Learning sector vendors and features on workplace learning, go to learningnews.com.