If an AI pilot works, what evidence would persuade you to fund it across the business? Three years after our conversation about Digital Transformation Payday, I welcome Tim Bottke back to Tech Talks Daily to discuss PREVAIL, his book on the decisions leaders make before anyone selects a model. Tim's argument is that by the time a project disappoints, the questions that might have changed its direction can be months behind it.

We talk about what happens when the CTO wants room to experiment, the commercial team wants a better customer experience, and the CFO expects lower costs. Those are all reasonable goals, but they lead to different projects and different ways of measuring success. Tim describes how he used company filings to study patterns in what businesses say about AI over time. The filings can show how the story changes. They can't tell us on their own whether a quiet company abandoned a project or simply stopped talking about it.
I also asked Tim what a CFO should expect to see before approving an AI budget. His answer includes a baseline, a comparison with simpler options and the costs that appear once a pilot is running across the business, including model usage, maintenance and retraining. We discuss his idea of "latte time," a reminder that saving a few minutes for each employee leaves a separate question about what the business or its customers gain from the time saved.
The conversation turns to AI agents and what meaningful human control looks like. Tim argues that leaders should decide where an agent can act, which cases should be escalated and who has enough time and information to intervene. We finish with three decisions a team can make before its next funding meeting: define the problem, name the person responsible and decide what will happen after launch. I'd love to hear your experience. Which of those decisions has made the biggest difference to an AI project in your business?
Useful LInks

[00:00:00] The leading issue of agentic AI in businesses right now is ensuring agents act with compliance guidelines. And Denodo applies guardrails across your entire data estate. By aligning your company's data infrastructure under one system, these guardrails perform consistently across your platform. So start scaling your business and start with Denodo. Simply visit
[00:00:25] Denodo.com to learn more. If your AI pilot works, what would convince your CFO to pay for it across the business? The last time I spoke to my guest was three years ago. We were then talking about the digital transformation payday. But he returns today to talk about his latest book called Prevail.
[00:00:50] And it is a book about management decisions that shape an AI initiative before a model is even chosen. Yeah, we're going to talk about the human leadership problems, not just the tech. And today we'll discuss what happens when the finance, the technology and commercial teams all think that they're solving a different problem. And Tim will explain why a pilot budget
[00:01:14] can hide the costs of running and maintaining AI, and why a simpler tool deserves a place in the comparison. And there's also a very useful distinction here between getting people to use a tool and improving the work that they do with it. No good saving time if you're only going to use that extra time sat in a coffee shop. So if you're about to approve another pilot, my guest today will
[00:01:41] offer some questions to take along into that meeting. But enough for me, let me reintroduce you to him now. So a massive warm welcome back to the show. We last spoke three years ago. But for everyone that missed that conversation, can you remind them with a little about who you are and what you do? I need it. Great to be back again. Cannot believe it's three years really. I'm Tim. I'm a senior
[00:02:05] strategy consultant, more than 20 years in boardrooms helping clients make important technology decisions, strategic decisions. And more often than not seeing how they then, after decision, go quite not delivering returns, etc. That's why I wrote the book Digital Transformation Payday, what we talked about three years ago, linking value to digital transformation. And guess what? These days,
[00:02:31] it's all about AI, but it hasn't really changed much. But still, I've been asked many times by readers, can you do something about AI in the same mindset? Maybe a bit less academic, much more pragmatic. And that's what I did. It's called Prevail coming out in January. And that's one of the reasons I was excited to get you on the podcast today. Because as you said, you've been working on this book called Prevail. It's not out until January, but I'm going to try and
[00:02:57] get a few little teasers out here before the book can get in everybody's hand. And from my understanding, it draws on over the 18,000 company filings from 978 listed businesses. But I've got to ask to begin with, what did those findings reveal about the difference between AI programs that are actually creating measurable value and moving the needle inside an organization from those that just remain stuck at the level of ambition or the dreaded pilot purgatory?
[00:03:27] Very good question. So the book is about the experience over the years, what happened in these boardrooms. But look, I'm a bit of an analytical person. So I just didn't just want to write down opinions, but I wanted some facts. So I reused the methodology from Digital Transformation Payday, now on steroids, with all the technology you can do these days, to find out what companies really say about AI. And guess what? They never say we messed up our AI program. Why should they?
[00:03:57] Certainly not in public filings. So what they do is they start excited, we're experimenting, and we are looking into, etc. But when you follow the reports over time, you can see different patterns. They repeat the same crap again and again. They just go silent. And so there are all these different patterns which we can now analyze with natural language processing and the new technology.
[00:04:24] And I also actually had to read what it found, because one of the theses of the book is, it's all about humans, not technology to succeed. So I couldn't then just let technology analyze the reports. So I found three and a half thousand talking about AI over five years. They were then sorted, etc. And as I said, the patterns are really interesting.
[00:04:50] So people start talking about technology without having the problem defined. They start losing the story about AI over time. So all these different patterns you can see, but never really admittance of pure failure, obviously. And it feels like there's somewhat of a paradox in many of these initiatives, because as you said there, humans, not technology, are what are needed to make a project a success.
[00:05:16] But when an AI initiative fails, it's the technology that often solely gets the blame. But on that side of things, what kind of management decisions are possibly setting failure in motion before a model is even trained? That's exactly what I'm writing about. So the actual problem of why these initiatives fail actually starts way before someone ever selects a technology. And it's leadership issues. So things
[00:05:42] like, have we actually defined the problem which we want to solve, which is the major question which has often not been asked at the right moment in time, because people were carried away by an amazing pilot. And let's say it's true, all these technologies are amazing. But pilot and scaling is a very different story. So when you do not define the problem, you just are taken away by all these nice things
[00:06:08] technology could do, but it will not do them for you at scale. Then you do not have the data ready. No one is asking, OK, the demo, the MVP is amazing. But what about our major data source? Then there are no numbers available. So you cannot compare what you do to what you have done before. You cannot measure success. And all these kind of management layers, I describe them in a building in my book. All these building layers, they actually make sure that if you do a mistake at the
[00:06:37] foundation, the building cracks no matter how far up you go. And I suspect over the last, what, three years, we've all seen examples of organizations that have gone AI first, problem second, or have shiny new technology that is in search of a problem. But refreshingly, you suggest asking teams, what problem does this actually solve? But on the flip side of that, what does a convincing answer sound like? And what should, or what does the pause before
[00:07:05] the answer tell a leader? What kind of things have you seen here? Look, when you're in the room with the right people, the CFO, the AI lead, someone from the business side, et cetera. So if you ask the problem and they all give different answers, that's already a clear sign. The CFO wants to experiment. The CMO or the commercial officer wants to improve customer experience. The CFO wants to save cost. The vendor wants to earn money as much as possible. So when you ask the
[00:07:34] problem, you do not get a convincing answer that's already a clear sign or no answer. That also happens. A convincing answer can have many different dimensions. If you are more on the efficiency side of things, which is only one aspect of AI, you need to have named clearly the process you want to fix. You need to have a proper baseline you can measure against. You need to be sure you can scale
[00:07:56] it. And you need to be sure that the cost you incur over time, you know, the tokenomics term, which is now used all these times that actually tokens can be fairly expensive if they're used in the wrong way. So all these need to be set up in place at scale. That is very important because as before for robotics, process automation or for digital transformation, a little pilot is cool and fun. But if it doesn't
[00:08:24] scale across the organization, the CFO would just not see it in the spreadsheets, you know, and I've had this happening before. The CFO telling me saying amazing return, you know, eight months payback for five FTE efficiency. I have how many Tim? Oh, 10,000. Okay, so what does it feel? So the second part, because this is the cost part, and the second part, which is now coming more and
[00:08:51] more into the scene is, do you really rethink your organization in a way that AI is giving you a competitive advantage? New revenue streams, new something, whatever a competitor cannot replicate. And then we are back to what we discussed also last time on digital transformation payday. You need to play to win in the space. And if everyone has the same agent, the same AI based
[00:09:16] on the same data, no one gains advantage other than the vendors implementing. So you just get a new state in the market where no one has any advantage. So that's the second part, the much harder to define what problem does it solve. Because if your problem is you want to compete and beat competition, you cannot just say AI is the solution. It has to be more than that. It has to be a combination of different factors, including AI, where you can be better than competition, at least for some time
[00:09:47] until they can catch up. But what evidence should a CFO expect before approving an AI budget? And how can teams maybe better define and measure a result that finance will actually get on board and can actually trust and believe in? I think that the simple trick is that it needs to be a longer term business case. So it doesn't make sense just to look at the pilot and not think about how you can scale it. Because a pilot will always
[00:10:15] work. You know, you have to be to have a very bad team to make sure that a pilot is not working if you want to make it work and prove what a good idea it is. But then the scaling of this idea to from the single pick process to a bigger one where it really changes how the firm operates or how it competes against the market in the market. That's the tricky thing. So you need to have a cost baseline running
[00:10:43] down for a couple of years. You need to have clear breakpoints. You need to have all the costs in, including token costs, including maintenance, including retraining of the models. And you need to make sure that you have a comparison, a test group, because I've seen cases where the AI was so amazing that people forgot that a simple deterministic model at a fraction of the cost would lead to more or less the same results. So you need to make sure that you're actually comparing
[00:11:12] versus your baseline and not just that the delta versus the baseline is that it now includes AI, which is no goal by itself. It's just a tool. Just to bring to life some of what we're talking about here and help it resonate with leaders listening all around the world. Can you maybe walk me through an example of that infamous tool first, problem, second mistake, and just show how its cost can compound as that program grows?
[00:11:39] I think there are so many examples out there these days. And I think any personal assistant type of AI from all the vendors we have out there in the market is a good example. So all firms introducing them, everyone who is trying to experiment is being trained in them, etc. And then what people say is, now we have the tool. How do we measure that we are not actually producing what I call latter time?
[00:12:03] Latte time means, okay, the time you gain by using AI is used to have a cafe latte in the cafeteria. So it doesn't really change anything other than it's more comfort or producing additional data, etc. and producing latter time. And then organizations, what they actually then count is adoption, you know, how many people are using it. But that's useless. Just because you and I use it,
[00:12:32] it doesn't mean we're either more efficient. And it doesn't mean on the other side that we are more successful than the market have better products, etc. And you need to measure on that side of the story, not by just adoption. So that just a tool doesn't help anything. You always have to define first, what problem does it solve? Otherwise, you'll end up nowhere.
[00:12:57] And we are recording this towards the end of 2026. And the big talking point of the year has been all things agentic AI and agents that can act across several systems and do so at machine speed. But as everyone's rushing into this, and we're now talking about not hundreds, but thousands or hundreds of thousands of agents out there. I'm curious, what kind of gaps in ownership, permissions, accountability, or process? Are you seeing exposed most quickly? Are you seeing anything here?
[00:13:28] Yeah, a lot. And that's something I am still struggling. And I think everyone on the side of my clients and in all markets we're struggling with. So it's scary that, remember all the layers I defined of the building from the ground floor where you define the problem, from the data you have or either you don't, from the governance, all these layers in historical management, you could just go layer by layer and
[00:13:54] say, okay, I fix this layer, then I fix this and I fix this. It's still hard. Almost no one is doing it properly, but it can be done. But when you have an agent or many agents, these are CS elevators, you know, running through the building at machine speed. So you have no time to really think once the aid, the elevator and the agent is running through the organization, then to fix all these
[00:14:19] different layers, you need to fix them first. And that's really scary. And then what do you do? Do you, do you say, okay, I put a human in the loop, you know, this famous term, what does it mean? We're happy to talk about this more, but agentic, not defined properly on the governance side, on the permission side, etc. is really scary. Because in the end, the manager will be responsible.
[00:14:44] There have been cases where firms then said, but it was the agent's fault. That doesn't work. You know, if you set an agent free and it does harm to human to anything, it's the management to blame. And the management will go and will be punished for it. So with all the excitement I have about agents, these things have to be fixed first. And I have been to 20 plus tech conferences around the world. I can imagine.
[00:15:14] One of the phrases that I hear even more than agentic AI is that phrase you used a moment ago, human in the loop. And it feels like it's often used as almost reassurance midway through the keynote there. But what does meaningful human control look like in practice and more than just a checkbox to offer reassurance? I just wanted to say the one thing it shouldn't be is that somehow in the process, a pop-up in front of whoever the human is gives them the option to
[00:15:44] say yes or no. Yeah. Without proper time to think about it, without all the data, with too many of these cases running by them in too little time, et cetera. So for me, human in the loop is not like the last resort somewhere in the process to find someone to blame in terms of as a human. But the human needs to be in the loop before, you know, to define what the boundaries are, to check all the extreme cases,
[00:16:12] and then to have escalation paths in cases where there is doubt, where people, the humans in the loop, they're not like in a roller coaster loop, but they have time to really think about and decide. It's so easily said, we just run on agentic speed, we just put a human in there somewhere. But if humans cannot cope with the speed, it doesn't help anything. So it has to be the human has to be in
[00:16:38] front to define what boundaries are there and to make sure what are the edge cases where you wouldn't want any machine to decide all by itself. By the way, the machine will not decide. It's a probabilistic, not a deterministic decision in the end. And I always try and give people this thing a few valuable takeaways. So if we do have a board or a program leader that's listening to us today, and they want to improve their AI initiative on a Monday
[00:17:07] morning, what three actions should come first? And how should those choices maybe affect and impact the people that are expected to use the system? Any advice that you would offer there or a takeaway? It's very simple. Number one, which by the way, from the data I've looked into, is also the one thing, if it hasn't been done properly, the market punishes. Yeah. So if you do not define this number one, the market will know and your share price will suffer. That's
[00:17:34] what the data says. So guess what? Number one is ask, what problem do we solve? Just as this question, but you can then go as deep as you want. Do you have a proper case? Do you really differentiate in the market? Does it really scale? Like all these kinds of sub questions of what problem does it solve? They are the most important part. The second, if you decide or come to the
[00:17:59] conclusion that you found something which is worth the money on it to solve a problem, which is worth solving, then you need to make sure that you have the governance right for the program, which all the dimensions we can go into detail for every day written out in the book. But most importantly, someone needs to be in charge, not a committee, not an AI group, whatever. So the case which solves an
[00:18:24] important problem needs to be managed by a manager who can fix all the problems on the way, the conflicts between IT and commercial, any escalation issues, the decision to stop things without endangering the own career, but to stop it because it doesn't work. Because AI programs are not standard step-by-step programs, they're more experiments. And you can either scale them or not. But if you cannot, then it
[00:18:52] doesn't make sense to continue. And third, wanting people forget and all the excitement, you need to have a plan what you do once you've launched. Because after go live or after the first initial launch, there's still a long journey ahead. Data is changing, you need to retrain, cost might change. That's what many people have seen when the vendors increase their token cost, not as such by token, but by the
[00:19:21] volumes people would use to solve a problem which hasn't been properly defined. So these three things, very simple. What's the problem? Who's in charge? And make sure you have a longer term plan for scaling, not just piloting. These are the three things people should look at. But you find tons of recipes, questions, etc. in the book, checklists. But I think if you start with these three, you're already on a good
[00:19:45] track. Fantastic advice. And I would imagine writing a book about technology now, it must be incredibly difficult and challenging because the speed of technological change. I mean, recording this in early November, the book's not out until January. Is the book completed? Are you still adding some finishing touches before it's released? No, it's all completed. That was the most
[00:20:10] encouraging result of having the first draft. Guess what I would have happened if I had found out that technology is the problem. Because then how it's running so fast that by the time you finish a sentence, it's already wrong. So fortunately, the problem is on the human side and we are slow. So what is written in the book will be true for years, no matter what technology you put behind. Remember the payday,
[00:20:38] digital transformation payday had some of the same messages. So that's encouraging. So the book, while it takes a bit of time, it's already out in the market, I'm having many discussions, and it will still be true. Because it doesn't really matter what large language model you implement. If you do these things wrong before you implement, it will not help. So the big the book release there, when is it out? Where can people get hold of it? And how can they
[00:21:08] connect with you? Where can they find out more information? Look, you can prevail is available in any bookstore for pre-order already. So you can already order it. If you really want to read something early, I can encourage you to go prevail the book.com the web page. We are still recruiting launch team members and they get special benefits, including more or less reading access to the book in two days,
[00:21:33] plus additional gadgets. And that's what you can find in the launch team section of prevail the book.com. And you can reach me on LinkedIn, or via the contact information on that launch page. Awesome. Well, I will add a link to the website for the book there and your LinkedIn. I encourage people to get in touch and continue this conversation we started today. And indeed, the book's coming out
[00:21:59] in January. It'd be great to get you back on in the new year when we're happy to. Yeah, when we've all got that magical feeling of wanting to make a big change in the new year and do things differently. So I'll get you back on in early January. We'll talk about the book as it comes out. But more than anything, thank you for giving us a sneak peek today. Really appreciate your time. Neil, so happy to be back and looking forward to January. Thank you so much. One of the many things I'm taking away from this
[00:22:25] conversation was Tim's advice to decide who can stop a project before asking anyone to expand it. And it's that person that needs authority to resolve disagreements. Look at the evidence and say when the result doesn't justify the next round of spending. And I think that same thinking applies to AI agents. Give people the time, information and authority to intervene. And a screen asking someone to approve
[00:22:52] a decision is only useful if they can meaningfully understand what they're being asked to approve. So a big thank you to Tim for returning to Tech Talks Daily today. And you'll find Prevail and the resources that Tim mentioned in the show notes. I'd love to hear, though, how your team handles all the problems that we talked about today. And what evidence would make you change your mind about
[00:23:15] an AI project that you've already invested in. As always, techtalksnetwork.com. You'll find 4,000 interviews. You can leave me an audio message or find out more information on how you can work with me. Lots to do there. But I've used far too much of your time today. I'll be back again tomorrow with another guest. Thanks for listening. Bye for now.

