3228: Thoughtworks on AI Agents, Explainability, and What's Next

[00:00:04] From the rise of AI agents to the growing influence of small language models or SLMs and open source AI, businesses are now navigating this rapidly shifting world of possibilities and challenges. And my guest today is Mike Mason. He's the Chief AI Officer at Thoughtworks. And together we're going to explore the next wave of AI advancements.

[00:00:30] And with a deep understanding of emerging AI trends, enterprise adoption challenges, and the shifting priorities of AI leaders, Mike is going to help break down how businesses can harness AI for real world success. While also addressing some of the rising concerns such as explainability, security, and compliance.

[00:00:55] So are we heading towards a future where AI agents become our true digital teammates? And will SLMs outperform their massive counterparts in privacy sensitive tasks? And how will open source AI reshape innovation while also ensuring transparency and legal safeguards? These are just a few of the areas we're going to explore today. So let's get Mike onto the show now. So thanks for joining me on the podcast today.

[00:01:25] Can you tell everyone listening a little about who you are and what you do? Sure. My name is Mike Mason. I'm the Chief AI Officer for Thoughtworks. We're about a 10,000 person digital services company where we actually build a whole lot of software for our clients. I'm the Chief AI Officer, but I've been there for 22 years now. So I've done a lot of the technology roles as a programmer and an architect and a technology strategy guy.

[00:01:51] And now I'm working on bringing bringing AI into everything that we do at the company, which obviously is a big deal because of all the all the all the stuff that's been happening in the space, especially the boom over the last two to three years. Yeah, it has been a crazy few years for the last two years. It's been all about generative AI got and I said last year it's going to be all about agentic AI in 2025. And true to form, AI agents have been a hot topic in the first three months of 2025.

[00:02:21] But I'm curious from what you're seeing, how do you see them evolving this year and what industries do you think will benefit most from their adoption? So I think the thing that we're really going to see this year is an evolution from very simplistic agents. There's a lot of what we call agent washing right now where people are labeling something as an agent when really it's just a fairly simple chat bot or a simple AI system.

[00:02:47] Whereas an agent, a true agent has some autonomy, can kind of do a multi usually do a multi step process to create some value for you, can make a bunch of decisions, maybe with a human in the loop or not.

[00:03:01] So what we're going to see is sort of this true rise of agentic systems that are able to handle much more complex tasks, multi step tasks, and then also something that we call multi agent systems, where you have a number of specific agents that are each handling kind of sort of a specific specialized task in kind of a chain or a grouping of agents.

[00:03:30] And where those agents are collaborating together to get something done. So that's those are the those are the two things that you're going to see. I think you're also going to start to see more agents integrated with human workflows kind of almost as teammates. You know, certainly I use I use ChatGPT quite a lot in my work and a number of other tools. It's fun to bounce them off each other. And I use them as as teammates.

[00:03:57] And when I'm programming, I have a you know, an AI coding partner that is that is helping me to think through what am I programming? What have I missed? What might I want to do next? Not just sort of the code generation thing that I think a lot of people get get excited about, but also just working there as a partner. So some of these business agents are going to know a lot about your organization.

[00:04:20] You know, if you're a customer service rep for for Telco or something like that, those agents are going to be able to tell you a lot about the customer that you're that you're trying to support. They're going to know a lot about your your products and services that your company offers, and they're going to be able to sort of partner with you on helping you do your task. I think customer service is sort of one of these very obvious ones. I think Klarna in particular has sort of made the most noise about this.

[00:04:49] Middle of last year, they said that they were reducing their human workforce from eighteen hundred people down to twelve hundred because they were using AI agents to respond to their customers. So Klarna is a buy now, pay later firm. And so you get lots of kind of customer queries coming in through that. And they've gone from that position to say, actually, we're not hiring any new people. We're going to scale our business using agentic AI approaches.

[00:05:20] And you said you enjoy playing with chat GPT. Have you tried the deep research? Yeah, I was playing that over the last few weeks. Phenomenal kind of results from that, isn't there? I have. Yeah. I mean, what I thought was interesting was I I asked a couple of questions in areas that I don't really know about. And of course, they claim that it has Ph.D. level responses. This this deep research thing.

[00:05:44] And I sort of posted a few things to my to my LinkedIn and one of my school friends who's actually now a math professor down in New Zealand said, actually, that is the level of quality I would expect from a first year Ph.D. So it's it's not complete hyperbole to say that, you know, the AI systems are starting to be able to produce very useful outputs with some quite deep.

[00:06:12] I'm going to say reasoning in air quotes because a lot of people will nitpick with that and say it's not actually reasoning. It looks like reasoning. And I would say if it looks sufficiently like reasoning, does it matter? I'll be splitting hairs at this point. But, you know, I should also point out open A.I. is not the only game in town here. So I was using Google Gemini the other day, which also has a deep research mode. And I was pretty impressed with the results from that. So I think that's another thing to bear in mind.

[00:06:41] OpenA.I. sort of does have a lot of the mind share at the moment. But it's a strong competition. You know, it's it's not clear who's going to win or even if there is necessarily a winner or there will always be sort of different models with different levels of capability to them. And what about anthropics, Claude as well? Do you play with that? And how do you do you use? I'm curious to use the three of them for different things or some perform better in other tasks than others.

[00:07:11] Yeah. So, I mean, they all have a different style. Right. So I actually find Claude to be somehow the friendliest one. And so the way I do it is I have a chat chat GPT or open A.I. kind of enterprise at work so we can create GPTs and we can share snippets with each other. And we've got kind of an enterprise data agreement with open A.I. That's one of the things around using A.I.

[00:07:37] You've got to be sure that you are not doing the wrong thing with your corporate data or your customer data or whatever it is. So you need to be sure about that. So for work, I use open A.I. and Gemini at Google because those are the two main ones that I use there. But from a personal perspective and for research and when coding, I use a lot of Anthropic stuff. It's also Anthropic Claude is really great for coding.

[00:08:07] It's a phenomenal tool for that. It's probably the world's leading coding A.I. And there's a whole bunch of tools that plug into it and then provide that. You know, we were talking about agents. They provide that kind of agent coding experience within my IDE. There's my developer environment that's running on my desktop. So it's doing stuff for me and also talking to the Anthropic backend servers to process stuff through.

[00:08:35] The other fun thing to do always is, you know, you get some output from one A.I., you paste it into another A.I. and say critique this. Right. And it will pick it full of holes. If you tell it to, it will it will find the problems. And I find that's actually quite useful if you're trying to do a bit of research and you're like, is this exactly right? And you can you can kind of get a double check on that. Yeah. Love it. Another trend that we're seeing more and more of is open source.

[00:09:01] A.I. seems to be gaining a lot of traction this year, especially with the need for greater transparency and obviously a lot of arguments around ethical data usage at the moment. So how do you see businesses balancing that openness with proprietary A.I. models? So I think there's a there's a number of different types of openness here. Right. And so there are the proprietary models where you don't get to see them at all. That's, you know, the open A.I. is the big, big Google models of the world.

[00:09:31] And there are more open than that models, which are open weight where they publish what's called the weights of the model. So that's that's, you know, when they say 10 billion parameters, the weights are the description of how those parameters are connected to each other in a neural network. And so that's these big multi gigabyte files that they publish.

[00:09:53] But they might not publish the code they use to train that model or as to your point, the training data for that model. So that's open weight. And then something that's the most open is open source, which I think is the thing that you're getting at where they will publish, you know, not just the result of the model, but also all of the code that was used for it and all of the training data.

[00:10:18] And so that gets you to like fully open source. And the reason that I sort of make sure of the distinction between those is that they're all useful for different things. Like a lot of the open weight models, like the, you know, the Lama models from Meta are very useful and still get you some benefits over those proprietary models, even though they're not fully open source.

[00:10:39] But if you're worried about sort of the ethics of it and, you know, what data are we using, then you probably do want a fully open source model. I think on this one, the companies that we work with, the clients that we work with, they take a range of stances on being concerned about intellectual property and ethics and all that kind of stuff.

[00:11:03] A lot of them are taking the approach. You know what? I'm using open AI on Microsoft Azure and Microsoft is indemnifying me against any kind of intellectual property concern. And I've got a relationship with Microsoft and I trust them. And they're a big enough gorilla with a, you know, an entire building full of lawyers to defend me. I'm going to, I'm going to say that's good enough for me. A lot of other organizations don't say that.

[00:11:30] And they say, no, I actually want to see all of the training data that went into that. We have a large financial client where their policy is never to use any cloud services because they're highly regulated and they're suspicious of, of cloud services. And so for them, we used an open source code generation model and then fine tuned it for their specific environment and their code bases and their coding standards.

[00:11:56] And then they deployed that or we helped them deploy that within their own data center running on hardware that they controlled. So that was sort of the other end of the spectrum, fully open source, mostly from a, from a slightly from the ethics perspective, but much more from the regulatory angle and worried about data leaks and all of that kind of thing.

[00:12:17] So, yeah, I mean, I think it's a, it's kind of a fluid, fluid situation and organizations are going to, the main thing, the main, main bit of advice actually is to be having the conversation about it, to go in eyes open and to like have that debate around where do we want to land on the, you know, the, the ethics, transparency, explainability, bias, all of those things.

[00:12:42] They're very important. And, and, and it's very important to be having that conversation within an organization. And what did you think of, I think it was Sam Altman that originally said everything online is fair game, hoovering up online data to train their data models and make chat GPT what it is now. And then a slight change of heart around IP without any sense of irony when DeepSeq arrived. It's, it's just an ongoing debate, isn't it?

[00:13:09] Well, world's smallest violin from me, frankly, on that, you know, the, the, certainly the artistic community's response to all of these AI models hoovering up even copyrighted images and using them, I think is very reasonable. You know, people are very concerned about it. And one thing I did see Scott Galloway, who I, who I love is a professor of marketing, but also, you know, does a whole bunch of stuff in the, in the online world.

[00:13:39] He was a South by Southwest last weekend and talking about the explosion of AI, but he also said there is very little these companies can do to protect their intellectual property from this kind of siphoning off of that, you know, IP value. Because it, I mean, you know, not only is it technically difficult to do, it's also the most sensible thing to do. If you're trying to build another model and you've got a super powerful one over here, you would want to take advantage of it, right?

[00:14:08] Like that's a pretty sensible thing to do. So, so I think, you know, at the moment, there's tons of investment dollars flowing into the, the, the AI gold rush. I think we are going to see a leveling out of that. I'm not sure if we're going to see any particular winners, but I think we're eventually going to get to the point where people realize, you know what? It's not the LLM itself. That is the thing that you can build a motor round.

[00:14:37] It's the next level up. It's the platform above that, or it's the application layer above that. Maybe it's agents, you know, maybe, um, uh, companies can use, uh, agentic systems to create that differentiation and advantage. Uh, maybe that's where we'll start to see some of that, uh, kind of moat building and, and, and, uh, bit more, a bit more of that. Yeah, completely agree. And, uh, explainable AI is something else also crucial for trust and regulatory compliance.

[00:15:06] So, uh, how far do you think we've come in making AI decision-making more transparent? And what challenges remain, do you think? Because there seem to be so many different stances. I know Europe were being, we're going to regulate here, but then I think UK and the US, they've gone more of, we're going to push forward and talk more about security than regulation. And there seems to be so much up for grabs.

[00:15:30] And I think over-regulation could slow things down and even leave, like the European Union getting lagged behind there. But I think this week I've even heard that they're maybe changing their mind as well. So what are you seeing here? How do you see it all panning out? I mean, so just, just briefly on regulation, I think it needs to be done very carefully. Like, you know, personally, personal perspective here. Yeah.

[00:15:54] I'm in favor of regulation that, that protects the average person because this stuff is super complex and, uh, difficult to understand. And we actually do need, you know, uh, a level of, of government oversight on that. Um, but the European situation, uh, if you look at some of the regulations that were passed from a technical perspective, they're borderline unimplementable, right? Like blanket statements, like the data will not be biased, right? It's not actually helpful, right?

[00:16:22] So there's, there's, there's also a question about like the mechanics of getting technical people involved in the regulatory process, which did happen. And I've spoken to someone who was involved in the European regulations, but by the time it had gone through the sausage machine and come out the other side as actually the, the fully baked regulations, it, the language in it was, was actually, uh, slightly problematic in terms of, of implementation.

[00:16:46] Um, I think on explainable AI, it's really important to sort of like talk about what do we mean by explainable, right? Like if I've got a neural network that is assessing people's mortgage applications and I trained it on mortgage default data and who, who pays their mortgage and who doesn't and who's high risk and who's not. I train it on all of that stuff. And then I put your information into it cause you're applying for a mortgage.

[00:17:10] Um, and I say to you, well, uh, the neural, uh, the neurons in the model, you know, fired at these levels because of these input things. And then this output node, it added up these five things and it came to more than 0.5. So you're high risk and we're not going to give you a mortgage. That's not a very good explanation, right? Like it is an explanation, but it's not, it's certainly not understandable to the average person. You've got no idea whether that's fair or not. I didn't tell you what kind of, kind of data I trained it on.

[00:17:40] So, uh, you, you know, and you, you probably, most people are not a neural network expert, you know? So I think, um, the, we have made some progress on explainable AI. There are frameworks to do it. There are ways of building models that are more explainable. Uh, but unfortunately there's also a trade-off with efficacy because as you make the model more explainable, it actually gets less good in its predicting power.

[00:18:05] So in that mortgage example, um, if you had a model that was very, very good at, uh, you know, predicting who was a good mortgage candidate and who wasn't, if you, if you tweak the model and use different, uh, model building techniques to make it more explainable, it's actually less good from a business perspective for doing the thing that it really needs to do for the business, which is determining, uh, mortgage risk. So, um, I think this is going to be an evolving landscape.

[00:18:34] Um, you know, one of the other things is not necessarily, uh, explaining model mechanics in detail, but having a technique of, um, uh, evaluating model output. So this is something that we're doing research on here at ThoughtWorks, um, where we are, you know, looking at large language model performance under different, um, input, uh, scenarios. When you ask it a question, are you getting the right answer every time?

[00:19:03] The thing here is, you know, if you, if you ask a large language model, a question, it can phrase its answer in multiple different ways and it would still be the same, uh, correct answer, but you somehow have to evaluate all of that. So this is a pretty deep area of, of active research right now. And I think we're going to see that evolving, um, you know, throughout the years, this is, this is not going to be a quickly solved problem.

[00:19:27] And something that we've all been talking about for the last two to three years is large language models, no avoiding it. But of course, now we're starting to talk about small language models, SLMs that are sometimes proving that bigger isn't always better. So where do you see SLMs outperforming large models and, and what does it mean for AI deployment? Would you say from what you're seeing here for small language models?

[00:19:52] I think it's important to be clear on what we mean by outperformance, because in general, a bigger model will, you know, on a benchmark, if we've got one of these standardized benchmarks, it will give a better result on that benchmark. Um, but a small language model, either, uh, you know, a distilled version of one of those big ones, or simply a model that has designed, been designed from the beginning to be a, you know, a smaller, you know, and it's funny. We talk about small 10 billion parameters is now a small language model.

[00:20:22] I'm like, okay, great. Um, but, uh, the reason those things are so interesting is because, uh, you can literally run them on your laptop. You know, um, I've got a, a fairly beefy Apple laptop, but I can run, I can run a deep seek model on, on that, you know, not the 671 billion parameter model that requires a Nvidia super cluster to, to run on, but I can run a cut down version of that.

[00:20:48] Um, and there are other organizations, um, building small language models, uh, a number of those coding models we talked about earlier, um, those open source ones, those are small language models. Um, they're really helpful because, uh, so from a, from a data security perspective, small language models are also great because you, uh, can control the hardware that they run on.

[00:21:09] Um, they're also cheaper to run, which is very helpful, uh, especially when you, when you think about some of these agentic systems that basically need to run a large language model kind of in a loop, um, as it works through its task. It maybe tries some things, especially in the coding space, it might try some stuff that doesn't exactly work. And then it has to kind of fix a problem with its work. So it's running that large language model in a fairly tight loop.

[00:21:36] And that can get pretty expensive if you're using, um, you know, one of the big, uh, online model providers. Uh, so that, that also helps. Uh, the other thing that I think, especially deep seek has pointed us towards is that, uh, fine tuning a model for your own specific purposes, uh, can be, uh, an important strategy. You know, can be a viable strategy previously.

[00:22:00] We thought that, you know, only if you had gobs and gobs of data and tons of money, would you need to fine tune a model? Um, or would it be cost effective to do so? But deep seek has kind of shown us, uh, that maybe, uh, maybe more people can be fine tuning models. And before you joined me on the podcast today, I was doing a little research on all things thought work. So one of the things that I came across was your global survey that is highlighting a divide in gen AI.

[00:22:27] So if anyone that's not seen this report, what would you say are the key differences between how leaders and emerging companies are approaching AI? Because we've been talking about it for a couple of years now. We're now seeing a level of maturity. We're focusing on things like ROI on those AI projects and measurable differences. But what are you seeing in the differences here? Yeah, sure. Thanks for asking about that. So we did a global survey.

[00:22:55] It was a thousand senior business decision makers. Um, and, uh, we looked at the difference between, um, uh, the leaders in gen AI adoption and, and the, and what we call the explorers. Uh, so for the, for the leaders, um, they're prioritizing, um, kind of a strategic integration of AI. Uh, they're emphasizing security, compliance, bias detection, um, and their main challenge.

[00:23:21] So for 50% of them is adapting gen AI to specialized tasks, uh, rather than worrying about all the data quality that they need to go into those AI, uh, systems. And they're investing heavily in, um, automated bias detection tools. Um, and they're, they're, they're thinking a lot about alternative data sources. So synthetic data, licensed data, um, stuff like that.

[00:23:49] And their key concerns are, uh, things like copyright violations. Um, for the explorers, uh, they're primarily concerned with sort of immediate operational efficiency. Um, and their main struggle is validating data quality. Uh, so, you know, that, that data quality is an interesting one to pick up on because that seems to be sort of the, the major difference between, uh, companies that are leading in gen AI adoption and those that are still lagging.

[00:24:18] Uh, and again, for those, uh, for those laggards, um, they're also concerned about data privacy violations. Um, 88% of them are concerned about that. Um, and, and much fewer, uh, only 60% of them are adopting alternative data sources. Uh, I think the other thing that, um, uh, came out, uh, that was interesting, uh, there are regional differences. So in the U S and Singapore, uh, companies were focused heavily on, uh, monitoring and logging.

[00:24:48] Um, those were the top things that they were looking at. Whereas in the UK and Germany, uh, companies were prioritizing, uh, regulatory compliance, which kind of gets back to your, your point earlier in the conversation about the regulatory landscape, uh, uh, between, you know, in different parts of the world. Yeah. Again, so many great points there.

[00:25:07] And it's easy to, to say that some of these businesses risk getting left behind because they're scared of the technology, but it is very serious matters that are holding some of those organizations back. And you mentioned two great ones there, security and compliance, a big concern around AI adoption. So for any cautious business leader listening, what practices should they, or what best practices should they be following to mitigate some of those risks and some of those concerns when they're scaling AI solutions?

[00:25:38] So I think the key thing is that rather than being reactive, an organization needs to be proactive. So you need to, um, integrate security, transparency, and a good governance process, like at every stage, um, rather than waiting until you have a problem.

[00:25:57] Um, I think the other thing that I would mention is, you know, the, the recent boom in generative AI has, has, uh, you know, shone a light on this, on this, uh, issue. But, uh, people who've been deploying machine learning systems and, and really data at scale systems of any kind over the last 20 years have had to look at some of these issues. And AI is just kind of, uh, you know, turning up the temperature on that.

[00:26:23] Uh, so a few specific suggestions, um, I, I, I would, uh, I would offer, uh, you know, one of them is to adopt secure by design practices, uh, at ThoughtWorks. We try to build security in at every stage of, of building a software or technology system rather than trying to bolt security on at the end. That's very difficult to do. Uh, so that's things like doing threat modeling and risk assessments throughout your process.

[00:26:48] Um, the, the next one is to, uh, establish and implement a robust AI governance mechanism. I mean, clear ownership and accountability structures across your organization. Uh, things like, uh, clear understanding, you know, if something is wrong in an AI system, whose job is it to fix that? And, and how, how quickly are they expected to fix it? Um, another thing that, that can be very useful is, is continuous monitoring and logging.

[00:27:16] So one of the things, uh, you know, you're not done when you put the system into production, you need to continue monitoring that and make sure that, uh, it continues to perform well, especially an AI system. And the, the model is not, you know, it's called drift that the model is not drifting away from where you had trained it and the, and the, the characteristics that you had trained into it. And has kind of, you know, wandered off on its own.

[00:27:40] It's not like the model will genuinely just, you know, randomly go in a different direction, but the environment in which it finds itself, uh, may change and that may reduce model performance. And so you need to kind of, uh, stay on top of that. Um, I think another thing that I would point out is that, uh, there are companies in the world that take a, um, privacy first stance. Uh, Apple, I think is the poster child for this kind of thing. Um, and from a branding perspective, you know, it's been incredibly good for them.

[00:28:09] So this isn't just about, uh, uh, uh, you know, something that you have to do to stay compliant. This can actually be a business advantage as well. If you are able to kind of work that into your branding and consumer confidence as, as a piece of your brand. And as we keep marching forward here, how do you see AI overall shaping business operations beyond automation, which is the big hot topic at the moment? And also what new roles or skill sets are going to be in demand?

[00:28:39] We hear a lot of those lazy, um, arguments about AI is going to take all the jobs, but of course, new roles are going to appear. So what, what are those new roles and new skill sets are going to be in most demand? Do you think? Uh, great question. I think, uh, you know, in the, in the, well, short to medium term, the new skill that we all need to get is collaborating with AI, you know, because AI isn't necessarily going to take your job, but someone who can use AI better than you, that's what you got to worry about.

[00:29:07] Um, and the, the move to AI augmenting everything that we do, I think is, is going to be a theme. So, um, you know, if you are a strategy consultant, um, you know, we have a few of those here at ThoughtWorks, uh, you know, if you're a strategy consultant, what does it mean to be an AI powered or AI assisted strategy consultant? And we need to start thinking about that.

[00:29:34] Um, AI workflow, uh, you know, if you think about collaboration with AI, um, I was using, uh, I was using Anthropic Claude the other day as, as we were talking about. And I asked it, um, Hey, I'm, I'm trying to, uh, do a, do a call with some of my school friends. One of them's in New Zealand, one of them's in the UK and one of them's in Canada. What would be a good time zone, uh, to do that?

[00:29:57] And Claude, rather than just giving me an answer, it built me a little JavaScript app on the fly where I could pick time zones and it, in my time zone. And it would tell me the other two time zones and color code them red or green, whether they were, uh, you know, day, day or night. Um, and so like, that's a, we can do that already, right?

[00:30:19] Like if we move to a stage where AI can dynamically generate user interfaces for us, depending on what we're doing, someone's going to need to design the AI system that does that. Um, I think, uh, you know, new skills and emphasize skills in, in AI ethics and risk management. We, we talked about that, um, uh, you know, uh, prompt engineering, uh, I think, uh, prompt engineering is a, I don't know. It's a funny term. I don't think it was ever engineering.

[00:30:49] It was more like a dark art. It was prompt voodoo. Um, but broadly speaking, that skill of being able to provide the right context to an AI and to actually train your humans in an organization to be able to provide context to an AI. Um, I think that's, uh, that's going to be an interesting role as well. You know, uh, AI disaster recovery, like what happens when your AI system has a meltdown and you need to pick up all the pieces? I don't know.

[00:31:18] I don't know how often this is going to happen. Um, but you, you know, I'm sure we will start to see problems that organizations blamed on, well, the AI did it. Well, someone's going to need to clean up after that. Wow. I know you're incredibly passionate about this space. We've covered everything today from AI agents, agentic AI.

[00:31:39] But if we were to dare to look even further ahead into a virtual crystal ball, what's the next big shift in AI that you think businesses should prepare for beyond 2025? Or is there something that just excites you or makes you want to jump out of bed in the morning that you're monitoring? Where do we go from here? So I think one of the things that we will see is, is something that we call domain specific large language models. So that's like industry specific models, um, you know, like a legal LLM.

[00:32:09] We already see those. Um, but even more, um, specific, uh, for an individual company's use case. So I'm expecting to start to see, uh, you know, uh, AI trained for a particular organization.

[00:32:23] So at ThoughtWorks, we have a number of AI tools that we provide to our people, which, um, we use that to embody, uh, kind of 30 years of our engineering excellence practices within an AI, uh, partner for, for use on teams. So I think we'll start to see, you know, more specific, um, AI, uh, I think that's interesting, uh, multimodal.

[00:32:49] So these AI systems that are able to kind of seamlessly switch between, um, uh, uh, language, uh, spoken language, uh, visuals, uh, text. I think that's very exciting. I think we'll see more video this year, uh, smaller models we talked about already, um, you know, Apple intelligence, possibly the, the, the, the, the, uh, clunkiest, uh, acronym creation, uh, in the, in the last couple of years. Um, but that is starting to roll out.

[00:33:17] So more and more people are going to get exposed to, uh, you know, being able to run, uh, AI on their device. Um, and as we spoke about multi-agent systems, I think the other thing that is happening, uh, you know, for the last couple of years, uh, organizations have been very much in that experimental mode. Uh, we are going to start to rapidly see, uh, much more pressure to show real return on investment, um, and actual business outcomes.

[00:33:44] Uh, we're already seeing that with our clients, but I think that's something that you're going to start to see scaling across organizations. And more and more of these, um, public stories about how AI has actually, uh, created real business value, uh, for organizations and is, and is doing so in exciting ways. Well, thank you so much for sitting down with me and sharing your invaluable insights today. But before I let you go, we're going to have a bit of fun with you and ask you to leave one final gift for everybody listening.

[00:34:13] We have a Spotify playlist where I asked my guests to leave a song to, and also an Amazon wishlist where they can leave a book that they'd recommend. What would you like to leave everyone listening with and why to know? Uh, so I would like to leave you the book. Uh, it's called the purpose code by Jordan Grumet. Um, and it's an interesting book about, uh, life's purpose. I think all of us, um, you know, we, we go through life. Um, we, you know, we're ambitious in our careers.

[00:34:43] We've got plans about what we want to do, but there's always a question of, you know, what is, what is the point to, to all of it? And I think that eventually, uh, shows up in, in, in people's heads. The book is really interesting because this has been written by, um, a, a hospice doctor who has actually helped many patients through, um, the end of their lives. Um, and in talking to those people, he's really tried to boil down, you know, why is it the purpose is so elusive, um, and, and difficult for, for, for people?

[00:35:12] Why is it that we sometimes feel, um, that we're not, you know, reaching our goals or, or whatever? And the key insight that he has in the book is that, um, there's both big P purpose and little P purpose. And big P purpose is kind of the, uh, you know, big life goals. I want to build this business. I want to become an astronaut, all that kind of stuff. Um, but sometimes those goals, you know, you don't achieve them and that can be really problematic, but there's also little P purpose in life.

[00:35:41] You know, the things that you do every week, like that you get personal satisfaction from, and that is important to you. And so for me, for example, you know, I have big P purpose goals in my role as the chief AI officer for a, for a multinational technology firm. But I also have little P purpose, you know, like when I get to coach people on their leadership journey through ThoughtWorks, I love that. When I get to talk to somebody, uh, on their podcast, like I'm doing today, I actually love doing that, you know?

[00:36:09] And those are kind of little P purpose for me, which is still, uh, incredibly important. And being able to identify those, I think is very helpful. So I haven't finished the book yet, but I'm loving it so far. Oh, also I'll get that added straight to, uh, the Amazon wishlist. And for anyone interested in learning more about everything we talked about today, more information about what you do at ThoughtWorks, the report that we referenced and everything in between, where would you like to point everyone listening as a starting point?

[00:36:38] Uh, well, so you can go to thoughtworks.com, um, and, uh, everything is, everything is on there. Awesome. Okay. I'll add links so people can find that information relatively easy, but we covered so much in what, 35, 40 minutes today. And I'd love to hear more from the listeners, see what they thought. If we missed anything, anything they'd like to add to the conversation and what they see next, but more than anything, thanks for starting the conversation today, Mike. Very happy to be here. Thank you.

[00:37:08] So the big takeaway, AI is no longer just about experimentation. It is now about execution. And as Mike pointed out, companies leading in AI adoption, they're no longer just looking for low hanging fruit or quick wins. They're now embedding AI into their strategic fabric, focusing on security, compliance, and explainability. But the big question is how will businesses balance innovation with responsibility?

[00:37:38] Will open source AI level the playing field or introduce new challenges? And how will smaller, more privacy conscious AI models change the way that enterprises handle their most sensitive data? And what do you think? Are we heading towards a future dominated by autonomous AI agents and decentralized AI models? Let's keep this conversation going.

[00:38:07] Slide into my DMs, LinkedIn, X, Instagram, just at Neil C. Hughes. And if you enjoyed today's episode, please subscribe to the show for more insights on AI innovation, the future of business, and so much more. I've already got a guest warmed up, ready for your listening pleasure tomorrow. So hopefully we can speak again then. Bye for now.