Making Industrial AI Deliver Real Operational Value With IFS

[00:00:00] - [Speaker 0]
AI agents are only as strong as the data that they're given. When provided with an outdated dataset, your agents could end up doing more harm than good, but not with Denodo. With an AI data layer built within your platform, your agents are provided with real time data changes. So with Denodo, your agents can finally make the right business decisions. Simply visit donodo.com to learn more.

[00:00:31] - [Speaker 0]
What happens when an AI recommendation can change a work order, dispatch a technician, or affect a wind turbine? Well, today, I'm joined by the chief AI officer at IFS. We're gonna examine what industrial AI must get right before it moves from pilot to production. My guess will explain why business processes, operational logic, and decisions about what AI can see and determine success long after the model demonstration ends. Yeah.

[00:01:05] - [Speaker 0]
We'll cover everything from a digital worker projected to return 90,000 technician hours, sovereignty, operational control, staged autonomy, agent identities, and who remains accountable when automated actions run the risk of causing real world harm. We're gonna cover a lot today. So if your AI strategy measures pilots instead of uptime, cost, safety, and turnaround time, today's conversation will hopefully offer a much sharper test for whether the technology is delivering real value where work happens. But enough for me. Let me introduce you to my guest now.

[00:01:47] - [Speaker 0]
So thank you for joining me on the podcast today, Bob. Can you tell everyone listening a little about who you are and what you do?

[00:01:55] - [Speaker 1]
Absolutely, Neil. Pleasure to be here. So my name's Bob Decaux. I'm the chief AI officer at IFS, which is a global enterprise software company. I've been in the role around seven and a half years, and and my background is in complexity science.

[00:02:09] - [Speaker 1]
So that's why I did my PhD. And then I'd to say I did did multi agent systems and agents before they were fashionable. But IFS is enterprise software that sits behind, you know, a lot of the physical safety critical infrastructure that we see airports, wind farms, oil rigs. So we think about, you know, industrial AI, you know, how we apply AI and embed intelligence into those real world systems. And we like to think of it as where AI becomes real through, you know, work orders, schedules, the parts, connecting the decisions to to the real operational outcomes.

[00:02:44] - [Speaker 1]
So that's that's me and what I do.

[00:02:46] - [Speaker 0]
That's one of the things I love about what you do here because every day on this podcast, I try and get people thinking differently about the impacts of technology. And Some of the industrial areas, which I know you serve, people don't automatically associate with tech. And we have spent what the last few years talking about AI models getting more powerful, but but many businesses are discovering that capability alone doesn't get them from pilot production. So when we're talking about industrial AI, what separates those that are successfully scaling it from those accumulating experiments that never deliver measurable value? We've all seen the horror stories, but what is the secret to those that are succeeding, and what are you seeing succeeding there?

[00:03:29] - [Speaker 1]
Yeah. So, look, I I think you mentioned it. Right? It's the scale point. I think it's it's easier than ever now to to just get get started.

[00:03:38] - [Speaker 1]
Right? We see that with large language models. You you can just start start building things very quickly, and and you have plenty of lab experiments. But the the difficulty is getting it reliable, governed, and and usable in in production, and that's that's where the gap is. I think something that is not anywhere near as much as it used to be is like a data cleanliness problem, right?

[00:03:58] - [Speaker 1]
I think that used to be the issue maybe a couple of years ago, but, you know, language models really help us, you know, get through that, a more easily than we used to, you know, taking all that messy data that companies have across their whole data enterprise and bringing it together. So it's still an issue, but but not as big one. But I think the real moat for the companies that are scaling well is that they understand their business, their processes, and they know how to expose it to AI. So, you know, the processes, the metadata, the logic that they've got in how they operate to actually get things done. And they're the ones that are that are really pulling ahead rather than just having the cleanest data or the the tidiest data warehouse.

[00:04:41] - [Speaker 1]
And, you know, a lot of that is is kind of operational and architectural discipline. Right? It's it's almost the the unglamorous work that's gone on behind the scenes to deliberately decide what should you show to AI? What decisions should it make? And what do you want to keep away from the AI?

[00:04:59] - [Speaker 1]
Where are the things that you can't afford it to to get wrong or misinterpret or paraphrase? And and that whole idea of, like, surfacing your business to an AI to control, I think that's where we see the companies being very successful, the ones that scale.

[00:05:15] - [Speaker 0]
And industrial AI feels very different from just asking a chatbot to summarize a document in an office because decisions can actually affect factories, equipment, supply chains, and people in the physical world. So where are you seeing AI produce the most meaningful operational outcomes out there? And and what could maybe other industries learn from some of these deployments, do you think?

[00:05:38] - [Speaker 1]
Yes. So look, as you say, think in in the industrial deployments, often subjectivity is just not an option. Right? If you're if you're going AI to write a nursery, right? Mean, you you can be wide open if you need it to, you know, order an exact replacement bearing for a wind turbine gearbox.

[00:05:54] - [Speaker 1]
Right? It has to get it spot on. So, you know, we we we see the the businesses that that are being successful taking advantage of, you know, the data often off their their assets so that that will be, you know, things out in the field, their, you know, their compressors, their turbines, but being able to turn that into to meaningful actions. So not just being able to make predictions on what might happen, understanding the consequences of that down the stream, automatically raising the work orders, getting technicians out there. So the success is really embedding AI across that whole business process thread, if you like, rather than a sort of set of individual pieces all along the way.

[00:06:36] - [Speaker 1]
You know, a concrete example, customer of ours, Kodiak Gas in North America, and they've deployed a digital worker agent to handle material replenishment, right? How they manage their stock more effectively and put that in front of their field technicians. So instead of them having to spend a lot of time on phone calls, remembering part numbers, technicians can just talk to it in plain language and it will go and find out, you know, what it needs to order, go and get hold of the parts. They project that to save them around 3,000,000 a year in ROI, ninety thousand hours to the technicians who can be doing higher value work. You know, those are the sorts of initiatives that we're seeing being very valuable.

[00:07:19] - [Speaker 1]
And I think it transfers to other industries in terms of the focus, right? Don't just think you put AI over everything without changing your processes and it's magically going to fix it. Apply AI very specifically to the tasks at hand along your business processes. Be very deliberate about the kind of the boundaries it operates in and that that's how you'd be successful whatever you're trying to do.

[00:07:44] - [Speaker 0]
And another big, talking point at the moment is AI sovereignty, which is a term that can quickly become tangled up with geography and where technology providers are quartered. But what does meaningful AI sovereignty what's that actually look like for an enterprise, and how much control should organizations be retaining over their data, their models, decisions, and ultimately infrastructure?

[00:08:09] - [Speaker 1]
Yeah. So, look, I think you're absolutely right. It's not just a a data residency or a compliance point. It's very much a control point for your operations. Right?

[00:08:19] - [Speaker 1]
It's about control. It's about resilience to change, and it and it is, of course, around your data boundaries. It's not about being completely independent from global technology providers. I don't think anyone can afford to build their entire stack alone in a world that's moving as fast as it is. Need to keep up with the specialists with the level of investment that's being made in AI.

[00:08:43] - [Speaker 1]
But you have to retain that control over how you apply the technology. I mean, can think about it, of the duck that is serene on the surface and paddling away furiously under the water, right? If you think of that as the fast changing AI technology layer, you need to keep those models interchangeable. You need to keep flexibility there. Different providers, you need to think about running things, you know, locally on premise using open source models as backup, maybe as a new option, small language models instead of large language models.

[00:09:17] - [Speaker 1]
That's the really fast moving piece that's changing and you don't want to tie yourself in. Companies don't want to tie themselves into things there. But on the surface, things move a lot more slowly. The actual application of these things, you don't wanna be changing all the time. When a new model comes out, you can't just rip up a process you put in last week and put a new one in.

[00:09:38] - [Speaker 1]
Right? It has to be audited, tested, all those things. So that speed of movement of the technology remains separate. And I think sovereignty is just keeping control of of how you manage that that piece below the surface. It's certainly not putting a wall around your border.

[00:09:55] - [Speaker 1]
It's it's owning your data, your decisions, and your ability to operate under under different circumstances.

[00:10:02] - [Speaker 0]
And before you join me on the podcast today, I was reading how you believe that The UK has an opportunity to become particularly strong in applying AI within complex and regulated industries. So tell me a little bit more about why you think The UK has an advantage here. And, also, what do you think needs to happen for that opportunity to translate into businesses, jobs, productivity, and and better economic value rather than just more AI ambition?

[00:10:29] - [Speaker 1]
Yeah. I I think it's a really important sort of framing question, first of all, all about what the ambition should be. Right? And so, you know, the I think if you if you want to try and compete on the on the frontier model as having the biggest clusters, the most compute, the most capital invested, that's not a race that The UK is is going to be able to win. Right?

[00:10:52] - [Speaker 1]
And so I think, you know, targeting the ambition to where The UK has the advantage is really important. You you have to do some of that investment, and we'll we'll touch on that in a minute. Right? But I think that the opportunity is is absolutely focusing on applied AI, right? And leveraging the world class industries, aerospace, life sciences, financial services that have a couple of inherent advantages.

[00:11:16] - [Speaker 1]
They sit on deep specialized data. It's very hard to obtain. Right? And there's not in these highly general, huge frontier models that exist. Right?

[00:11:28] - [Speaker 1]
And there's also decades of hard won domain expertise and trusts applying these problems into the industries that you can't magically make appear overnight. Right? And that's the layer that no one on the outside can reach. Right? The frontier labs, all the new AI startups have fantastic technology, reaching into that data and the hard run experiences is the hard part.

[00:11:57] - [Speaker 1]
And that is that's where The UK can really a distance a difference, sorry. So applying AI to the hard real world regulated problems, I think is what it should be doing. And there's a huge amount of value there, right? I mean, I think the IMF had pointed out £47,000,000,000 of potential annual productivity boost for the next decade by deploying AI across different industrial sectors in The UK. And the take up at the moment is still pretty small, around 18% of UK businesses are reporting AI adoption.

[00:12:33] - [Speaker 1]
Number is obviously always going to be growing over the time, but there's a huge amount of upside there. The government recognises this. It's invested, you know, 500,000,000 in the sovereign AI unit to really sort of kick start that in terms of its AI policy. But I think a key to be successful is to not just take that 47,000,000,000 as an abstract figure, start mapping it sector by sector and get specific. So, you know, in manufacturing, where can AI help?

[00:13:01] - [Speaker 1]
Optimising processes, predictive maintenance in construction, making your supply chain more efficient and robust in energy, you know, balancing the grids, smart asset monitoring. You can really get quite deep into the use cases that are going to be valuable for the industrial companies against the operational metrics, the KPIs that make them perform. So I think that's how you you target the value. But but The UK is really sitting in a in a great position for applied AI because of the data and domain expertise.

[00:13:34] - [Speaker 0]
And, of course, there is a big difference between demonstrating AI to a boardroom and persuading an engineer or field technician to, hey. Trust our recommendation when something so important is at stake. There could be a bit of friction there thinking the suits are are are thrusting this technology and this big project upon us. Nobody talks to us, etcetera. So that does need to be managed.

[00:13:56] - [Speaker 0]
So what have you learned about earning trust in the field, and how should businesses maybe design human oversight without reducing AI to something employees just simply ignore?

[00:14:07] - [Speaker 1]
Yeah. It's it's a great question and one I know you've you've explored a lot on on the podcast with the guests. Right? And, you know, I think a lot of that has been focusing on the so the AI model aspects. Right?

[00:14:18] - [Speaker 1]
How they how they scale or do they scale? Do they work at all the edge cases, which is obviously really important. I'll try and sort of frame it through a slightly different lens focusing on on the industrial version, if you like. And it comes down to two things you mentioned, right? The oversight in industrial settings and building the trust with humans.

[00:14:40] - [Speaker 1]
So the industrial oversight is, you know, for a lot of the industries, there's decades of accumulated accountability in the whole infrastructure that they operate with, with sign offs, dual controls, the way you have to report incidents, because human failure and machine failure just can't be tolerated in in a safety critical operation in the same way that it can be for sort of, you know, a finance process, for example. And AI hasn't earned the rights yet to avoid that accountability. I mean, not even close, right? And that oversight is the price of admission, if you like, for AI to play in a regulated world. But it's hard to do, right?

[00:15:24] - [Speaker 1]
And that's something that we focus on at IFS is how we can take the all the new technology that's coming, but apply it in a way into these regulated industries with all that oversight. So that it operates in a way that these industries can be comfortable with because that oversight is not going to go away. But the human factor is critical, right? And it's really a change management exercise around building trust. We see this with the scheduling and optimization product that we have.

[00:15:58] - [Speaker 1]
So what that does is it is basically trying to help companies with large field operations get people to the right place at the right time, right skills, the traveling salesman problem at large scale, if you like. And typically, this has been quite a manual process. You you have dispatches in the office who are managing a workforce of people around the country, moving around. And it's not that hard to show that if you apply AI to do that, can get better results. But building up the trust with the dispatcher, with the people in the field, that takes time because they've got all this experience of how it should work.

[00:16:39] - [Speaker 1]
And they don't just want to be told that it's better. They want to see and understand the outcomes, right? They want to be able to explain the decisions that it's making. So that trust is really earned through AI over time and it's not earned overnight. You know, I like to think of it as sort of nudge behavior or staged autonomy, what we call it, where we kind of release a little bit at a time.

[00:17:01] - [Speaker 1]
So you move from manual to, you know, more of a chat interface where people can ask questions and then maybe we serve up proactive recommendations to them on the AI telling them, you know, there'll be a good idea to do this. And then you start moving towards more of a full automation and getting those skilled people to focus on the higher value work. So it's not something you can rush through, but I think that's how we think about building trust in the field.

[00:17:28] - [Speaker 0]
And when doing research on you, one of the things I like is your argument that businesses should actually be thinking about AI agents as a workforce rather than another category of software. So if we follow that kind of idea through, should every agent have a an identity, a defined role, permission, separation of duties, and separation of duties, and somebody ultimately accountable for all those actions. And and what would that operating model look like in practice, in the area that you're working at the moment?

[00:17:59] - [Speaker 1]
Yeah. So I think it was something that Alex Bayvey raised on your on your podcast. I listened to that one where he was talking about the agentic identity and and, you know, having a control plane and agents and not just another piece of software. And I think from the from the IT point of view, that's absolutely right. But again, if we think about, you know, industrial operations, then it becomes much more important because of that sort of human oversight element that there is a sponsor for what the agents are doing.

[00:18:30] - [Speaker 1]
There is the sign off. There is the audit trail. This just doesn't live inside a security tool or, you know, some some IT tool. You have to treat this as a, you know, another worker within within your enterprise, which, as you say, you know, does mean an identity, role based access, separation of duties, and being able to see every decision that's been made. I think the important thing is that that doesn't just start when the agents are out in the real world.

[00:18:59] - [Speaker 1]
When, you know, if we go back to what I said earlier about, you know, what makes organizations successful with industrial AI, it's deciding what the agent should do and what they shouldn't do. And that's decisions that you need to make in advance. So you need to make sure you're giving the agents the right right tools, the right data, the right context and the right scope to be successful. In the same way as you couldn't blame a new hire to your organization for failing a job where you hadn't trained them or given them the right equipment, Right? I think it's we should be thinking of agents in in the same way.

[00:19:38] - [Speaker 1]
So that runtime governance is obviously still important. You need to know what agents are doing. You gotta have the feedback loop. You've gotta have the ability to to stop them doing things they shouldn't. But a lot of this is is about setup and setting yourself up as a business for agents to be successful as well as just an IT problem.

[00:19:58] - [Speaker 0]
And I guess that leads us to one of the hardest questions in autonomous AI right now, and that is the the concern that if an agent was to make a decision that, I don't know, causes financial loss, operational disruption, or potentially physical harm, where should accountability sit here? And our business is moving quickly enough to to answer that question before giving agents greater autonomy. What are you seeing here around this conversation?

[00:20:23] - [Speaker 1]
I think what we're seeing is is definitely uncertainty as you say. Right? It's a very hard question, but I think the technology is is moving so fast that the accountability that has to sit with the organizations that are deploying the agents, they're the ones that ultimately are trying to get the value from this and they're the ones that need to provide the context and the governance and the rules for which it's going to operate, even, you know, as we've got that duck paddling furiously underwater, you know, that that that governance can't change. Right? So I I think with agents, once you're in the field, right, if if they if they make a mistake and there's a bad output, it's not just the same as like a bad answer from a chatbot.

[00:21:12] - [Speaker 1]
It's a bad action in the field. As you say, that might be mission critical. It might cause physical harm potentially. Right? And that's why the accountability has to to rest with the organization that that's putting those agents out there.

[00:21:28] - [Speaker 1]
But if we go back if we go back to that trust point about trying to build the trust with the humans, I think once that trust is built, you have to watch the problem going the other way, which is they start believing the machine can do everything, and then they don't have the accountability that's necessary. Right? So I I still believe firmly that, you know, the way that organizations maintain accountability for the agents is through that human oversight. But the humans have to be diligent to always apply that oversight even when they think an agent is doing its job really, really well in in the same way that they'd have to be diligent around a human. Because in industry, doing it right 999 times and getting it wrong the thousands time, could mean something catastrophic.

[00:22:19] - [Speaker 0]
And at IFS, you have a reputation as a company that is leading the way with a lot of this technology, and I've looked to inspire a leader listening to maybe move forward more comfortably and confidently. And there is a lot of pressure on leaders today to have an AI strategy. So if you were advising someone listening today, what should they stop obsessing about, and what should they start measuring to determine whether AI is genuinely improving how their organization is operating. Any tips, advice or anything you pass down to those people listening?

[00:22:51] - [Speaker 1]
Yeah, absolutely. I think it's it's a really good point. And actually, you know, when when we started thinking about putting AI into the software at IFS, we always wanted to focus on outcomes and making it as easy as possible for for users to be able to use AI almost as if they didn't know. But as you say now, AI has become such a ubiquitous entity that everyone needs an AI strategy or at least thinks they do. I think, you know, there are there are some key things I'd say, right?

[00:23:22] - [Speaker 1]
Don't worry too much about which frontier model and the benchmark scores. For a lot of the use cases, that's not the important thing, right? It's actually about how you're going to apply AI into the problems that you're going to solve. Start thinking about what's defensible for your business and your operations as a leader. Right?

[00:23:44] - [Speaker 1]
So your data, your domain expertise, we've talked about any problems that are really specific to your operation that a generic model can't solve, that's going to be valuable. Well, that's going to require more effort from you. So, you know, I think they're the things to focus on. And then absolutely don't measure AI activity in some sort of generic nebulous form, like how many pilots you're running, right? It's it's got to be about changes to specific operational numbers that are relative to your business, which for our customers in in industry is uptime, costs, turnaround of parts, time saved for technicians, you know, the real operational metrics that are tied to their processes.

[00:24:27] - [Speaker 1]
And, you know, you've had plenty of people talk about, you know, the ROI shift from proof of concept to proof of value. I think that the extra nuance with the industrial version of that is the downside of getting it wrong just shows up as safety or physical outcomes that no company has. Right? So that's just an additional sorry. No company has.

[00:24:47] - [Speaker 1]
No company wants to have. So that's just an additional layer that that leaders need to factor in as well as just the AI performances. You know, what effect could this have if it if it all goes wrong? Because it's not just gonna affect your budget line.

[00:25:02] - [Speaker 0]
Well, I think that's a great moment to end on. We've covered a lot today from building trust in AI, AI agents as a workforce, accountability around that. And I know it's a busy year for you guys. You've got the, IFS Unleash event coming up, and, hopefully, I will get to meet you in person there. But for anyone listening, wants to find out more about the event, anything we talked about today, or just keep up to speed with some of the announcements.

[00:25:27] - [Speaker 0]
Where do you like me to point everyone?

[00:25:28] - [Speaker 1]
Yeah. Absolutely. Thanks, Neil. Yes. The IFS Unleashed event is is in October.

[00:25:32] - [Speaker 1]
You can find all this stuff at the ifs.ai website and learn about, you know, how IFS is is applying all the ideas I've talked about into industry through, you know, the AI in our software through our loops, So Gentic platform and through, our accelerated time to value, organization called Nexus Black. So ifs.ai for that and, you know, carry on the conversation with me, you know, LinkedIn I think is is the best place to find me to hear about what I what I'm talking about and I'm always happy to engage on these topics, passionate to talk about it and you know thank you very much for giving me the chance to do so today now.

[00:26:10] - [Speaker 0]
Well, thank you. And as I said, we've covered so much with other areas we covered. It was industrial AI, real world productivity, AI sovereignty beyond the hype. Anyone that's going to I f IFS or at least, please get in touch with either of us. I think we'd all love to meet you in person and carry on this conversation today.

[00:26:28] - [Speaker 0]
I'll have links to everything that you mentioned there too, Bob, as well. But more than anything, thank you for taking the time to sit down with me, sharing your insights today. Really appreciate your time.

[00:26:38] - [Speaker 1]
Thanks, Neil.

[00:26:39] - [Speaker 0]
I think Bob left us with a direct challenge today, didn't he? Stop measuring AI through pilot counts, benchmark scores, and start measuring the operational number that the business actually cares about. And an industrial setting, that could mean equipment uptime, maintenance costs, technician hours, turnaround time, or safety performance. And trust, most importantly, must be earned in stages. So begin with questions, move to recommendations, and only grant autonomy after workers can understand the decisions and begin to see those very reliable outcomes.

[00:27:21] - [Speaker 0]
And agents, they need identities, permissions, defined roles, sponsors, audit trails, and a human who remains accountable even after the technology has performed well 999 times before. Gotta keep an eye on Drift. So my thanks to Bob for joining me today. Remember, you can learn more about IFS. It's industrial AI work and IFS Unleashed.

[00:27:47] - [Speaker 0]
Can find all the links in the, show notes. And remember, you can meet me at IFS Unleashed if you're going to. But over to you before I go, which operational metric would prove that your AI investment is producing value. Write it down. Let me know.

[00:28:04] - [Speaker 0]
Send me an audio message over at Tech Talks Network. I'd love to hear from you. But that's it for today. Thanks for listening as always. Bye for now.