Turning AI Experiments Into Business Value With Newton

[00:00:04] What changes when the person listening to a business problem can also build the first version of its solution? Well, anyone who has watched the idea travel through meetings, requirements, documents and several teams, they will all recognise how much context can disappear along its journey.

[00:00:24] AI promises a much shorter route, but a convincing demo or presentation still leaves work before something belongs in production. And my guest today is Ben Grinnell, he's Chief AI and Capabilities Officer at Newton, a strategic delivery consultancy working across the private and public sectors.

[00:00:47] And today we're going to discuss how to put AI tools closer to the people who understand the work, all while keeping engineering judgment, security and trust firmly in the picture. Because Ben will explain today why starting with one useful problem, how that can create very real momentum.

[00:01:10] And we'll also talk about why measuring success through job cuts can very easily undermine adoption. And there's even a retirement story that suggests some people find relaxing considerably harder than working. I know I can relate to that one. So enough from me. Buckle up and hold on tight. I'm going to beam your ears all the way to the UK where you can sit down with myself and Ben right now.

[00:01:37] So a massive thank you for joining me on the podcast today. Can you tell everyone listening a little about who you are and what you do? Hi, thanks, Neil. Great to join you. So I'm Ben Grinnell. I'm the Chief AI and Capabilities Officer at Newton, which is about 700 people delivering strategic delivery work for our clients. A bit about my background. I started off founding a web consultancy back in the 90s, 94, because I needed to fund my university studies.

[00:02:07] That was when Mark Anderson was writing Netscape, so pretty early days. And then my first encounter with AI was when Kevin Warwick, who's Professor of Cybermen. He's at Warwick now, I think, bought the seven dwarfs to university. And they were sort of eight-nodial networks. And he showed me them and everyone was giving them human dwarf names, Dopey and Bashful. And I thought, that's eight nodes. What would eight trillion do? So I've been on the sort of AI journey for a long time.

[00:02:31] I did go and set in a self-drive car in 1996 when I was at the IFAC Control Conference in San Francisco as part of my PhD. And, yeah, I've been frustrated at how long it's taken to get those out into the world. And I started my career at Anson Deloitte's and then joined a startup consultancy with an aim to change the way to build consulting. I spent 21 years there, retired in late 2024.

[00:03:00] And I'd built a big network with the DevOps and the AI community at the time. And I retired and went around the world. And I was sitting in the Galapagos Isles in April last year, reading a draft manuscript of the Vibe Coding book that Gene Kim and Steve Yeggi wrote. And I just looked at it and thought, I can't be retired. It is such an exciting time to be in the workplace. So I came back. I was advising the board of Mutant around the AI strategy. And I just thought, it's going to take more than advising a board.

[00:03:29] So I went full-time into the role and have loved what you can achieve if you've got the backing of the sort of board and the leaders in the business. And it's been a fantastic roller coaster of a year since then. So brilliant fun. What a great story. Right from the heart of the browser wars to present day in AI retiring. Almost feels like that scene in The Godfather 3 where Al Pacino says, just when I think I'm out, they pull me back in. Yeah, absolutely. But it's funny because I've spoke to so many founders.

[00:03:58] A lot have had huge exits. And I think, are you going to sit on the beach and drink cocktails for the rest of your days? And they say, all that is great. But once you've got that travel bug out your system, you've seen everything you wanted to see. And in your case, the Galapagos Isles are absolutely amazing. But once you've done all that, you need a reason to get out of bed in the morning again, don't you really? Yeah. I mean, I had a 22-year-old bucket list goal there was to do the seven marathons on seven days. So I did that at the start of the year. And having crossed it off, I thought, well, now I haven't got any reason to get out and run anymore.

[00:04:28] So I've had to do some work. Love it. And obviously, fast forward to present day here. You're right in the heart of AI now. And I was reading that you argue that one of AI's biggest opportunities right now is actually putting tools directly into hands of people who understand the business problem. So what changes when those people can test and build solutions themselves? Because all too often, we've gone tech first, problem second. But what happens when you do it the other way around? Well, it's just been an evolution, hasn't it?

[00:04:58] I mean, if you look back 20 years, we were outsourcing a lot of that and offshoring it. And then we had the DevOps movement, which said, OK, well, let's make sure developers deploy stuff so we can get faster iterations. And I think 2007, was it? 10 deploys a day at Flickr. Famous conference with John Oldsport. And that always fascinated me. And I look at now and I think, OK, so now I can put one of my consultants in front of a client and I can have the same person listening to the problem as the person building the solution.

[00:05:27] And if I think about the effects of that, as you become more and more agile, less and less waterfall in your delivery, then the speed of the iterations is all that matters. So how fast does iterations be? Well, if you remove the handoff to the developer or especially when there's a contract involved in that handoff, then you've literally got someone building and experimenting with the client's thinking on the fly. And what that removes is you always lose a lot of context when you describe what you heard to another person.

[00:05:58] And people have different understandings as well. So it's a bit like Chinese whispers. So get rid of the Chinese whispers. So the person building the solution has the full context. There's a lot of cognitive load when you have to describe to another person as well. So you remove some of that cognitive load. And then the context, the cost of the coordination of one person's listening to the problem and one person's doing the work, that goes away too. So you end up with an ability to build the right thing faster. And in the world we're in, that's so important.

[00:06:26] So I think the thing you don't want to do is when you're building the right thing faster, get caught in there. Well, I've just really live coded this thing for the client. They can see it working. It's finished because obviously there's a ton of security, non-functional requirements and architecture that gets frequently lost in that.

[00:06:49] So there's definitely a need to make sure that I'm trying to make sure that I put the guardrails around those sort of client-facing developers so they can develop safely, which is exactly what I did in the DevOps world. I put guardrails around. I automated operations so the developers could do it. And again, that was an extension of the world where people had to build servers once and then we had the cloud and then we had DevOps. So it's just a continuation of that journey.

[00:07:15] But now it's reached a point where it's literally one person sitting in front of a person that's actually describing a problem that can solve it for them. And I must admit, I've been in enough stakeholder meetings with IT and people from around the business to know that traditional projects can lose important context as that idea moves from a business team and then onto the technology specialists and the project teams there.

[00:07:41] So for what you're seeing and hearing here, where does AI shorten that journey and where is specialist engineering still essential? Because those days of IT team taking, I don't know, 18 months to two years to deliver a project, those days are long. God, they need to move faster now, right? Yeah. And the biggest problem is the world's changing so fast. So if you do take 18 months to deliver a solution, even if you've got the requirements right at the start, it's definitely wrong by the time it comes out.

[00:08:09] So we have to be in a world where we deliver iterative value very, very fast. And the AI can speed up all bits of that. So I know we've seen Vibe coding and some of that work speed up the development cycle. And we have got tools like Blitzy that allow legacy code bases to be looked at and to bring that context in. So a lot of the early problems with Vibe coding have now been solved.

[00:08:35] But I was with a group in San Jose in April where we were writing some thought leadership papers on how you automate the testing bit of that work and how you can get models set up to challenge the code that's been written by a different model and make that bit faster as well. And whereas a year ago, maybe six months ago, you couldn't really ask the model about the architecture or to get the architecture right. Now you've got a very good architecture coach in a lot of those models.

[00:09:02] So I think the speed's brilliant. AI can speed up any element of that pipeline, but only if the person controlling it or directing it knows what questions to ask. And that's where it all goes wrong. If you don't know to ask the right questions, if you don't know to ask about what non-functional requirements should I consider, if you don't know to ask, how should I make this a modular architecture so it's easy to maintain, then you won't get that. Yeah.

[00:09:31] And looking back, if we go back, what, 10, 15 years, we had BYOD, then Shadow IT. Now we're hinting at Shadow AI. Just the thought of employees creating their own AI tools or workflows 10 years ago would have been unthinkable. But this is where we are. So what kind of guardrails are needed now so experimentation doesn't accidentally drift into introducing security data or even governance problems? There are some interesting parallels, aren't there?

[00:09:59] Because we had all employees creating Excel spreadsheets with macros, which are, I would say, more difficult to maintain than a lot of the solutions people are Vybe coding in Python now. Particularly access databases. So we had the Shadow IT sort of wars back then. I think we are a place where we can Vybe code things that are much more testable and maintainable than some of the previous versions.

[00:10:26] But we do need to make sure that we've got IT knowledge aware of them. I think for a long time, I've been working with clients to help democratize IT into the business and bring IT closer to the business. Because coming out of the days of offshoring a lot of that because IT is just a cost center, which was, we were in that world for a long time.

[00:10:51] If IT is business critical, and if it's not, I question which business you're in, but I think it's business critical for a lot of businesses now, then you want to keep some of that capability in-house and some of it very close to the business. So stop thinking it was a cost center. And if I'm in the IT function, I need to be working very closely with the business. So I'm closer to the profit and the revenue, because otherwise, I'm going to get stopped.

[00:11:19] A lot of the projects that start off in IT run quite well, and then there's a downturn in the market or something, and they get mothballed for a period of time. Or they go too big, and they build a lot of momentum, and then they can't keep that up. So I think having IT in the business with the people that know IT, actually working closely with their business counterparts and delivering value early and iteratively is the way we should have all been going for a long time.

[00:11:45] So this is just making that something that you have to do faster, because otherwise you're going to lose control of it in entirety. And just to bring to life some of what we're talking about here, especially for people listening, let's try and walk through something here. Why can starting with one narrow, very real problem, how can that create more momentum than launching a large company-wide AI transformation program where everyone tries to boil the ocean?

[00:12:11] Well, if we start with that one real problem, walk me through what could happen there. So the momentum is exactly the thing to focus on. So you can build a lot of momentum by starting a great big project, like let's do an enterprise data or enterprise AI project, but you won't be able to sustain that momentum.

[00:12:29] If you create a very narrow single use case where you can demonstrate something and it produces business value, so it impacts the P&L and people can see it and you can deliver that very quickly, then you create a smaller amount of momentum. But the thing that transformations need to do is they need to continually increase the momentum, otherwise they die. So I think starting small, learning fast and growing is the right way to go.

[00:12:56] And if you create business value, you create FOMO as well. So other parts of the business, everyone eventually is measured on business value. So if you're creating, as John Smart would say, better value, sooner, safer and happier, then a lot of people are going to be looking at that and going, well, how do I grab some of that? And that's where you get the pull, not the push. And in organisations, I think creating a pull for your services is massively important.

[00:13:25] And pushing against an organisation which has generally got too many big projects that it can possibly deliver is an uphill battle. So you have to create the pull and I think grow that momentum and keep it growing throughout the lifecycle. Your agents aren't producing accurate answers because they don't have a complete semantic understanding of your data. And Denodo is solving this and solving it through semantic consistency.

[00:13:55] Through semantic consistency, your agents can start making accurate predictions in real time. So see what else Denodo can do by visiting denodo.com to learn more. And I think, again, if we look back, tech projects traditionally failed when changing company culture, for example, was neglected and ensuring that people were ready to adopt any new solution.

[00:14:21] So when we fast forward to AI, what does successful adoption look like, especially when employees begin pulling AI through the organisation because it genuinely makes their work easier? Yeah, so I think we only really move at the speed of trust. So I think getting trust across the organisation in things we induce is absolutely critical. And we're not doing a good job of that in the news at the moment, especially this week.

[00:14:49] So I think if I look at the way AI is used in most big organisations at the moment, it starts off being something that the people use behind them. So they use it to enhance themselves and then present the AI to the world because that's the way chatbots have started. So initially, everyone is using it to make their back office, their personal back office more efficient.

[00:15:13] And I think the adoption path that really gets a company transformation is when it moves beyond that. So when people, which you couldn't do a year ago very effectively, but when you just start to create agents. And I think there's a load of different types that you could do. But I think the first one most people try is to create the expert agent.

[00:15:32] So if I'm the expert that everyone phones up to know everyone about the company, everyone about a specific part of knowledge, can I create a bot that they can then ask when I'm on holiday so that I don't get the call? So that's the sort of the personal expert bot. And then I think we're now moving into an age where you can have bots that show up as members of the team.

[00:15:55] So if you're a business that use Slack for your communications, then you've got dot in Slack that shows up and you can have it as a member of the team that the whole team can share an agent with the context. I think that changes the dynamic completely and massively drives adoption because then the whole team can see it happening and they've all got to adopt. Whereas when you've got individuals, you'll have a lot of leaders and laggards.

[00:16:21] And then you've got agents that go across the teams and agents that are chief of staff for people, which is another iteration, which I think is going to become very common in the workplace over the next six months. And then I think the adoption will get faster and the trust will get faster as well. But we've got to make sure that we're implementing the guardrails around that as we go, because there's a huge amount of power.

[00:16:45] I mean, if you look at the things we had recently where the bugs have always been there, but the AI has found them and made them more discoverable. In lots of large organizations, the data has always been there, but AI makes it really more findable. So on lots of SharePoint sites and organizations, the data that shouldn't be there. It's been there probably for 10 years.

[00:17:07] But when AI gives you ability to have such fast access over such a broad scope, then suddenly that's a much bigger problem than it was before. And that's exactly the same as the vulnerabilities that we're seeing out in the news at the moment. Corporate data and the trust that we have in the AI, we're going to have to make sure that we implement it sensibly and we put guardrails in place so we only give access to the things that we are comfortable with.

[00:17:37] And before you join me on the show today, I was doing a little research on you and I was reading how you said that some of the best AI implementations should actually remain almost invisible behind familiar systems. So do you have an example of technology changing substantially without forcing employees to change how they work and go somewhere else? Yeah. Yeah. I mean, that was based on my early experiments with AI and looking at how you can use it to re-platform an existing system.

[00:18:06] Yeah. So you can re-platform a system with all the bugs the original system had. And a lot of people say, well, why would you do that? Well, AI can do that. Humans can't because you just can't get developers to write those bugs in. But with AI, you can mirror what was originally there. And that came from my experience of trying to replace. I was the IT director for the home office, the border agency for 2007, 2010 when we were putting in the e-gates and things like that.

[00:18:33] And I was trying to replace one of the large legacy systems at the home office. And what I realized was that the system had been built and had evolved and it hadn't matched user requirements. So I had teams in lots of different cities, teams of caseworkers who had all worked out a different way of working around the system. Yeah. So everyone's got their own way of working around the bits of the system that don't work. Now, if you try and replace that system, one, you've got a massive technical engineering task to do.

[00:19:02] But two, the change management is almost impossible because whatever you do, it's wrong for everyone but one of those teams. Well, if you actually want to divorce the massive change management problem from the technical problem of this system's on a platform I can't support anymore, you could just get AI to rewrite it with all the bugs so that it performs exactly the same. So all those teams can carry on doing what they were doing. And then it becomes much easier when you've got a stable platform to do the other bit, which is the change management in probably a piecemeal way.

[00:19:29] So I think that's one of the use cases that was one of the earliest ones, which is now getting really advanced very quickly, putting a platform into a more, it might be something that's more defendable, more secure. It might be something that's in warranty or it might be just a coding language that you've got engineers actually know rather than one that everyone's retired. So lots of systems can benefit from that at the moment without having to wait for the change management effort that goes with them.

[00:19:55] And for any leader listening that is feeling somewhat inspired by what you're doing here, how should they decide which problems they should attack first and which belong with the frontline teams, which need a central AI function and which should still firmly remain with the traditional tech teams? A bit of a balancing act there, but any tips or advice on how to make that or make that decision a little easier?

[00:20:18] So I think I'd say it as an opportunity because a lot of companies have got a lot of projects running that are the traditional, they'll have a portfolio and there'll be the traditional IT projects, which might be I'm building a big enterprise service bus or I'm building a big ontology for all my data. So I've got one version of the truth, which you never quite get to. And there's a lot of white elephants in that space.

[00:20:46] But you've generally got three or four of those big corporate projects running that are struggling to get that momentum or maintain that momentum or get that business engagement. And then you've got a lot of projects in the business that might be doing the shadow IT or trying to do things faster and they can't get support. There's a wonderful opportunity with AI to work with the business delivering the value. So take some of the biggest problems that the CEO is keeping awake at night and say, OK, well, can we address that problem?

[00:21:16] And if we can do a fast iteration on that, but we can actually, because we can write in any particular code stack or language you like, because AI is brilliant at helping us with that. We can actually say, well, let's look at those big corporate projects and find which of those can we actually utilize to deliver this solution rather than building shadow IT. Let's make the back office IT project famous by using it to do this narrow slice.

[00:21:44] And then suddenly we found a way of joining up what we were trying to do here, which we were never going to get enough momentum around with what's the biggest problem in the business here. And the two are mutually beneficial. This one's not shadow IT and this one's no longer a big white elephant in the center, but it's actually driving value. And if you just carry on doing that with other vertical slices, but solving significant problems that impact the business and involving that and joining them up with the back office projects,

[00:22:12] then you've solved a lot of the organizational friction that you have with those competitions. They shouldn't be competing with each other. They should be helping each other. 100% with you. And I think if we look back, I think it was, what, 18 months ago, that infamous MIT study that found that lack of ROI from some of those big AI tech projects. There's been a big focus on securing ROI and measuring progress there, but knowing what to measure.

[00:22:42] We all know you can only improve what you measure, but knowing what to measure, that can prove difficult. So again, give leaders listening a valuable takeaway. What kind of measures will tell an organization that a small AI solution is actually delivering real value and ready to scale beyond its original two? I think it's a really difficult question, which is why you end up with studies like that one.

[00:23:05] And I didn't like that much because it actually said that, if I remember rightly, 95% of AI projects are not delivering a tangible difference to P&L. And if I was in an organization and one of my 20 experiments delivered a little value of P&L, I'd be quite pleased with that metric because if you're not experimenting at that rate, then you're probably not doing enough experimentation and innovation to really stay current in your market.

[00:23:33] So it's a headline grabber that it's not really what we should be measuring. I think when I look at the difficulty is because if you look at doing a study and doing a baseline, by the time you finish doing the baseline so you can show improvement in a mathematically proper way, the tools have all changed. So you never get a chance to do the same experiment again because all the tools have changed under you, which is why it's really difficult.

[00:24:01] But I think where we've got AI being used, we need to look at some of the measures that aren't just about efficiency. I think there's been a big focus on let's measure how many heads we can save. And I think that's the wrong way to do it.

[00:24:22] So if you're doing any transformation project that's focused, that gets a reputation for being about removing heads from the organization is going to get a lot of tissue rejection from the organization. Because at least for now, when you pit humans against AI, the humans win. So you will not get any adoption and they will set it up to fail.

[00:24:44] So we're describing the amazing new impact you can have if you use AI and how your job can be different before we're describing how few people we can have doing the job. So what is the next level job? And getting the people who are more keen to jump on that bandwagon before everyone jumps on them. Well, this is going to be a cost saving dragon.

[00:25:05] So we've really got to look at where is AI producing an increase in revenue, better services for our clients, more profitability or just removing the toil from some of our work so we can deliver those things ourselves. But it's got to be the positive measures rather than the ones that are generally perceived as negative because we won't get the adoption from doing that.

[00:25:27] And we are recording this at that time of the year where we're drinking pumpkin spice lattes, preparing our inboxes for Black Friday emails and Cyber Monday with one eye on 2027. So finally, what's keeping you busy at Newton? What can we expect next year and beyond? What are you working on? So we have taken all of our work throughs through the live coding hackathons.

[00:25:54] So everyone from the CEO to the junior person, just to make sure that everyone understands what we're working with. Because I think there's a need to be a bit hands on, whatever level you are in the organization, to be able to effectively work with this technology. So that's been one thing. We've also been looking very closely at the war for tokens, as you said. We had token maxing and now they're getting quite expensive.

[00:26:20] So we've got some sovereign compute. One, because tokens are expensive and we want to be able to at least run insurance ourselves. And two, because a lot of our work is in places where data shouldn't leave the UK. So we need to have that capability as well. And there's a, I'm predicting a big crunch in the inference market because there just isn't enough compute to go around. They can't build it fast in America at the moment.

[00:26:49] And I don't think we're doing any better in the UK. So I saw Microsoft pull out of a contract the other day, which set one of our main hopes of having a lot of infrastructure go down a bit. So I'm really keen that we protect our ability to use AI in a safe and secure way.

[00:27:08] So that's been a big focus as well as making sure that, you know, I've got the right things in place to make sure all of my workforce can continue to use it in a, not an unrestricted way, but a way that's giving our clients the biggest benefit.

[00:27:23] And one of the things I've done there is I've said, I do want people to have access to the latest models and they might be the most expensive, but I'm looking at the people who are using it most are already adopted and they can look at how can we make it more efficient? How can we use more efficient models? The people that are trying for the first time, I want them to be using the best model and getting the best experience because otherwise they won't become long-term adopters. I think that's a perfect moment to end on.

[00:27:50] And for anybody listening that just wants to find out more information about you, Newton, anything we talked about today, where would you like me to point everyone? Yeah, please. Newton's website, newtonimpact.com is a great place to learn about what we're doing and sharing with clients. And personally, I'm quite active on LinkedIn. So we're sharing a lot there about what I'm doing. Awesome. Well, I will have links to everything that you mentioned there and pop that in the show notes for people to check you out.

[00:28:17] I'd love to stay in touch with you and see how things evolve next year and see how many more marathons you're going to be running as well, maybe. But more than anything, just thank you for sitting there with me talking about your story. Really appreciate your time. Great. Thank you, Neil. That was great. I think Ben's advice today left me thinking about some of the handoffs that we just accept as part of working life. Every time someone explains a problem to another team, something can get lost. AI might shorten that journey.

[00:28:43] Provided the people use it understand when to ask for engineering help and how to protect the information involved. I also appreciated that challenge to how we measure progress. Saving time has value. But so does better services, stronger revenue, and giving people that extra room to do more useful and valuable work.

[00:29:08] And if employees hear a conversation or hear a rumor about using technology at the expense of reducing headcount, it's hardly surprising, isn't it, that enthusiasm will quickly disappear. So a big thank you to Ben for joining me today. And remember, you can learn more about the team at Newton Impact, at newtonimpact.com. And I'll add everything else to the show notes too.

[00:29:33] But over to you, what small business problem could you start with right now and solve with AI? And what evidence would persuade your colleagues to try it for themselves? I want to hear from you. TechTalksNetwork.com. Let me know. I'm afraid we have reached the end of this episode. I will be back again tomorrow with another one. But thanks for listening as always. Bye for now. Bye for now.