How Ensono is Building AI Resilience Beyond a Single Model

[00:00:03] Everyone is talking about the latest AI models, but what happens when one that your business depends on suddenly becomes unavailable? This is a question that shifts the conversation away from AI capabilities and towards something much more fundamental. Resilience. And in this episode today, I'm going to be joined by the Chief Strategy Officer at Ensono.

[00:00:30] And he will explain why infrastructure, governance and operational readiness, how all these things are collectively becoming equally as important as the AI itself. And we'll also discuss why the businesses that prepare now will be in a much stronger position for whatever comes next. Great conversation, this one, on a subject that I don't think we talk about enough. And on that note, let me officially introduce you to him now.

[00:01:00] So thank you for joining me on the podcast today. Can you tell everyone listening a little about who you are and what you do? Sure. My name is Neil C. Hughes. I'm Chief Strategy Officer at Ensono. I like to joke that that's not a real job, but I somehow kept very busy. Ensono is a managed service provider specializing in both legacy technology and modern. We sort of rebooted the company about 10, 11 years ago. We started at about 42 clients and 600 people of the carve-out that we did.

[00:01:30] And now 10 years on, 250 clients, some of the biggest companies in the world, over a billion dollars in revenue. And they all tend to have... We support large companies with mission-critical applications and sort of try and help them bridge the gap between legacy and modern. And that's kind of Ensono in a nutshell. So myself, I've been in technology over 30 years, lived and worked on three continents, visited 50 countries.

[00:01:56] Not particularly smart, but tons of experience and a lot of it on the client-facing side. I was one of the kind of founders of Ensono that kicked that off, was board member and COO, and now in this strategy job that's not a real job. Oh, absolutely love it. And your passion for the industry really shines through as well. I mean, before we even started recording today, we had a 10-minute conversation covering everything from cloud spend to tokenomics and everything in between.

[00:02:23] And fast forward to present day here, recent discussions around AI model availability. They're continuing to raise quite an interesting question. And that's one of the reasons I was excited to get you on the podcast today. And the question I'd love to ask you here is, should organizations be worrying less about AI sovereignty, which is a massive topic, and more about operational dependency on a single provider? What are you seeing? What are you hearing? And what are your views on this?

[00:02:52] A lot of passion, a lot of views on this. And in the time it takes to record this, the situation will probably have changed if things are moving so fast. The sovereignty issue is an interesting one and one we observed with our own clients for a while. I think people were very, very concerned about their data making its way into training models or being exposed in some way because it wasn't well understood.

[00:03:14] And then with the advent of RAG model and grounding AI agents in your own data inside of a gateway like a port key, that got to be understood fairly quickly. It's still, I'm not saying everybody's comfortable with every model and every country that the model is built in. But people have figured out that you can sort of, the model goes out in the world and learns how to speak English or French or Chinese.

[00:03:38] And then you can let it in your house to give it access to your personal data, but never let it leave the house again with your data. And that got under control fairly well. There are still some concerns, obviously, about sovereignty. But I think the data protection issue or concern has been, I'll say, somewhat solved. But what is now emerging is what you're describing.

[00:03:59] And what I've been telling my team on this is, and we obviously have a vested interest in this, but any technology goes through similar cycles in that your technologists sort of play around with it in their personal life. Sort of as a hobbyist, you can call that client server or Linux or the internet or cloud. And then it sort of makes its way into your business in the form of experimentation, either authorized or unauthorized. And then it starts to show some promise.

[00:04:25] And then you adopt that technology to underpin your service or the product that you're providing. And then you become dependent on it to deliver that product or service you're delivering. And when that last phase of the adoption curve gets hit, you are now dependent on it to run production, to ship your product, to route your trucks, to build your widget, to provide your service, to sell books on the internet. Whatever you're depending on it for, now that the thing you're dependent on needs to be governed.

[00:04:53] And governed can be such a boring word. But like we talked about in the opening, like FinOps is the greatest example. Like, all right, what am I spending on this thing that I'm dependent on? Is it the most optimal use of my money? But that's a flag. The FinOps is always kind of first. And that's your clue that, oh, there could also be security issues I need to look after. Oh, and reliability. Oh, and disaster recovery. What happens if this goes down? So that is happening much faster with this cycle than it ever has before.

[00:05:22] And I have clients and I have my own business now is dependent on AI to provide the service at some level. So boring old ITSM or IT service management comes into play. Are you monitoring it? Are you keeping track of it? Do you have tickets? Can you do change quickly? Are you agile? What do you do if a major incident happens? How do you respond? How do you get everyone on the phone? Do you have backup? All that stuff is starting to happen now at a rate I've never seen before.

[00:05:47] And just to hammer home what we're talking about here, especially for somebody listening who maybe has an AI model that their business relies on. What happens when that suddenly becomes unavailable? What should a resilient IT strategy look like? And how can organizations avoid that single point of failure? And I appreciate there's about four or five questions in there. But I think it's something that most organizations are just not thinking about at this current moment. No, completely. And full disclosure here, my first statement is going to be utterly self-serving.

[00:06:17] Well, you should definitely have a managed service provider to run AI managed services for you. And to do the boring stuff, you should outsource that which does not differentiate you and insource that which does. So the development of the model more than likely is something wildly differentiating for you, for this theoretical company we're discussing. You did something really clever that allows you to make it cheaper to ship your lumber to your distribution locations or something like that.

[00:06:44] But the running of that model and the monitoring of it and the looking after it and the cost optimizing of it may be not so differentiating. So in terms of like making sure it's going to support your business, either have an MSP, which of course I would say, or have an internal shop that formally looks after that. It's documented. It's an obligation. It's in your company's goals, et cetera, et cetera, et cetera. And you've got some sort of board or executive governance to make sure, oh, we're now dependent on something.

[00:07:13] Is that something being looked after? So again, the old boring ITSM stuff, even on this kind of fancy new technology, we're flying around in jetpacks, the same processes need to be adhered to. And then the dependency issue, that's a really interesting one.

[00:07:30] And this, some stuff happening like literally real time last week with some of the changes, I don't want to advertise for any particular company, but some of the changes Microsoft has made to their co-work and co-pilot, all of a sudden like, oh, that's actually pretty good. Whereas three months ago, it was pretty bad. So we, you know, we internally and some of my clients had built their sort of own AI models to do things like, oh, hey, help my salespeople produce a pitch deck that's grounded in my data, but can make a PowerPoint.

[00:08:00] Well, now all of a sudden embedded in the Microsoft suite is the ability to do this. And since Microsoft suite and they have something called Microsoft graph has access to, you know, oh, all my PowerPoints and all my instant messages and all my emails, et cetera, like, oh, they have a big competitive advantage. But then being used for something like a co-work versus Claude or others who have come out with similar things. So all changing really quickly.

[00:08:26] And there's an old adage in IT that systems should be loosely coupled and tightly integrated. And that has never been more true than now, because you want to take advantage of these things, but you don't want to lock yourself in for the next four years or find something that's hard to unwind. So loosely coupling things so they can be undone when the next cool thing comes around or something superior comes around, but obviously integrated so that it's useful.

[00:08:53] And I think over the last three years, many organizations race to be part of the AI gold rush, adopt the latest AI models. And we've mentioned that perceived boring side of IT that gave it that reputation as being a business blocker rather than a business enabler. But I mean, this belts and braces approach to IT has always been there for a good reason.

[00:09:13] So are they, from what you're seeing, are businesses paying enough attention to that underlying infrastructure, the governance and resilience that's needed to support these investments? What are you seeing and hearing here? Uh, no, absolutely not. Definitely not happening. Uh, because right now the trade-off is agility or, uh, um, stability a little bit. Yeah.

[00:09:36] And even, um, again, I won't name names, but some of my, some of my clients that were there at, at our annual advisory council meeting, which we have kind of every main person. And then a couple of virtual ones in between a couple of companies that you would raise your eyebrows said those exact words. Like we're going to trade off, uh, a little bit of, um, stability for agility. Then these are older, you know, companies who do very important things. Uh, but they understand they don't want to, they don't want to miss the boat.

[00:10:01] Uh, now this is within reason, of course, like they're not going to compromise the security of their institution or things like that. But there are a few clues we called out, uh, three really good examples to say that, no, that is not happening. One is, uh, it was a pretty widely publicized story. I think a company called Pocket OS who makes software for rental car companies. And they were kind of using Claude and a number of agents to kind of run their whole software stack. And Claude, uh, deleted their entire customer database.

[00:10:28] I think it made the assessment of, oh, well, this is so buggy. The best way to clean it up is to just delete it and all the backups. And then when they asked Claude, like, what are you doing? Claude came back with literally, I am so sorry. I violated every principle you gave me. This is a direct quote. You can Google this. Um, now they've recovered and they figured it out, but they, they had built a very slick software development organization, but that was very dependent on agents.

[00:10:54] And quite obviously didn't have the, um, governance in place to make sure those agents didn't do principal things. Uh, the other big example out there, uh, more on the financial side is Uber who consumed their entire token budget for agents in the first three months of the year. So FinOps is always the clue that the rest of these governance items are going to be needed. And I'm sometimes hesitant to use that word governance because it sounds so boring, but you got to know how much you're spending on stuff and what you're getting out of it.

[00:11:22] You have to know, Hey, if this breaks at three o'clock in the morning and I can't ship product, who comes on the call to fix it? Do I have a call out method? Do I have DR? Do I have backups of my customer database, for example, better, better, uh, not ephemeral in some way. So that is starting to happen again at a rate I've never seen before, but the adoption curve was so fast to your point. Um, it's, I'm actually amazed. We're not seeing more issues to be honest.

[00:11:48] Well, I'm so glad you shared that story because I think legacy technology and indeed legacy thinking is often frowned upon and blamed for just about everything these days. But one of the things I found refreshing as I, before you joined me today, I was reading how you argued that actually legacy systems are often the reason businesses continue to operate successfully. So how was AI changed that conversation around modernization without dismissing those decades of business knowledge embedded in existing platforms?

[00:12:17] Yeah, the, I think the answer to that question is going to be utterly AI is going to change it utterly. Uh, but in a way I don't think people completely understand. Uh, I think people are amazed when they hear how much legacy tech is still out there because what's written about and talked about is the, are the modern technologies. But the majority of everything we do walk around and see is still run by quote unquote legacy systems of some kind. What I didn't mention in the opening about in Sono is we specialize actually in mainframe management and mainframe modernization.

[00:12:47] And, uh, I think people are amazed to hear what those exist, those still exist, right? Yeah, they still exist. And if, if you turned off every one of them right now, there would be a zombie apocalypse that every airline, every flight that goes off every ATM transaction, every insurance claim, every banking transaction, not everyone, uh, our own, the U S social security system in the IRS, uh, NASA. And, uh, the companies still are very dependent on the mainframes.

[00:13:13] And actually IBM's launch of their new mainframe, the Z 17 was by far the most successful in history and happens to have AI embedded in the processors and chips of that. So you can do inferencing on the mainframe now. So all very wonderful. Uh, but the, the common problems or the challenges I should say that people face with legacy systems are one, the code that the programs that run on those legacy systems, uh, are written in the languages that they're written on and understanding and unwinding that code.

[00:13:42] Two is agility and the ability to change and move fast. Never more important than it is now. Uh, three, the ability to use modern dev tools. And four is tied to that, the ability to use modern developers who want those modern dev tools. All four of those can be corrected or augmented with AI without taking the very risky and expensive step of migration off of a legacy system. So we fight this all the time.

[00:14:10] I think people misunderstand the word modernization and they interpret it as migration. Migration is very much one motion under my, under modernization. The modernization can entail a much wider span of outcomes. So, uh, the modernization in place is cheaper, lower risk, and a much faster return on investment. So now I can use AI tools to do code discovery and document all my code and understand what it does. Fixes problem one agility.

[00:14:38] I can write modern, uh, uh, IDEs and modern, uh, um, uh, DevOps practices and procedures that can drop code onto legacy systems like mainframe or IBM power or, or, or old, um, Oracle databases, et cetera. I can do that, uh, and loosely couple and tightly integrate those things with legacy systems. So I can have modern developers, younger developers using modern tools could do development on a legacy platform. So if I'm a CIO and I walk in and say, ah, okay, I got a legacy system.

[00:15:08] What do I do? Should I take three years or five years and a hundred million dollars and probably fail and, uh, migrate that whole thing to cloud? Or do I take six months and one 20th of that money and one 20th of the risk to kind of fix some of the things that I don't like about that platform? Uh, and I think just as an advert for, uh, for mainframes, I think what people don't understand, probably a little, including myself, if I'm totally honest, the, uh, non-functional requirements of applications.

[00:15:37] Meaning some of the legacy tech, uh, the mainframe power series, et cetera, it is one giant piece of infrastructure, the ultimate hyper-converged piece of infrastructure, the ultimate HCI. You have your CPU, RAM, storage, networking, security, all in one giant optimized made to work together device that is also deeply integrated with the operating system that runs it.

[00:16:02] Like an iPhone, incredible piece of hardware that is inexorably tied to an incredible piece of software in an iOS. It can do things that cloud can't. And clearly cloud can do things that mainframe can't by, by a stretch, but they both have a place in the ecosystem. So if you're trying to produce massive transaction processing at scale, at incredible reliability, incredible security, that's, that's probably a good, it's probably a mainframe thing.

[00:16:28] But if you want, uh, sort of the agility that, and that, um, maybe, uh, disbursement in the, um, you know, availability zones, et cetera, et cetera, uh, multinational, um, spread out closer to the end compute to the edge, et cetera. You know, cloud's going to be the answer. So having both of those work together, I think the industry is coming around to, ooh, AI can be a giant enabler of that. AI agents are only as strong as the data that they're given.

[00:16:55] 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 denodo.com to learn more.

[00:17:19] And I was also reading before you joined me today, and so I worked with the market study group to analyze, I think it was six million lines of RPG code, all using AI. So to dig a little bit deeper on that, what did that project demonstrate about AI's ability to modernize rather than replace legacy tech? Yeah, that, that is, uh, we talk about that, uh, example probably too much. A, we love those guys. Marker study is the name of that company, big, uh, six largest insurer in the UK.

[00:17:48] Uh, had done a number of acquisitions and they had a, uh, IBM I platform that was by far the most efficient, um, platform for, uh, transaction processing for, I think it was rating of insurance programs. But, um, so it was great at that. Like we talked about massive transaction processing at scale, security, reliability, but the ability for their developers to drop code on that and make new insurance products in a matter of weeks versus months was limited.

[00:18:15] So six million lines of RPG code analyzed, figured out that's, I'm glad you asked that question. Perfect example of, uh, migration is a subset of modernization. And there were some applications on that platform that should go. They didn't belong there. This platform was not suited for them. So we move those off and then create a DevOps environment that makes it easier for their developers to drop code on that. So they can create new insurance products, uh, products very quickly.

[00:18:42] And then still rely on the execution model that the, uh, that the IBM I provided. So perfect example of everything I just talked about. Yeah. And also something else I wanted to bring up was what two, two and a half years after everyone got excited or distracted by the shiny object syndrome and going all in on AI everything. CFOs are now predictably becoming increasingly involved in some of those AI investment decisions.

[00:19:08] So how is that changing the way that organizations evaluate AI projects and also measure success and balance innovation with the most important thing of all accountability, which was accountability, which was somewhat neglected for a couple of years. Yeah, for sure. And including my own CFO, by the way, I think we, we were ahead in this regard. Um, our board was wonderfully three years ago, all about us making the investments necessary. So we've done a number of things.

[00:19:37] We, we have a very robust and mature AI ops platform. Uh, we hired a guy from meta who's built, um, a system that allows us to do predict incidents before they happen, govern change management, uh, help our techs resolve issues quicker. Our mean time to repair, for example, is this is a, we're a mature company that was already good. And our MTTR has dropped 50% in one year, meaning 50,000 tickets a month processed.

[00:20:03] All those tickets now take half the time that they used to, uh, which is amazing for the, from a quality of service perspective. But we had also parallel to the ops stuff. That was an advert for our upside. Uh, that's not for sale. It's just how we deliver a better service and more efficient. Uh, we had also built a tool, a different platform internally. So every single associate, 4,000 people at the company had access to, uh, LLMs in a way that was governed and controlled. So you have to, you should not restrict that which should be platformed.

[00:20:32] Don't block Claude and everything else. And open AI and Gemini create an environment where our associates can use it and doesn't let our data leave the home and go out and train anything. Uh, but also makes it easy to use. So any kind of persona who uses this internal platform can make an agent. Our CEO can make an agent, a super technical guy in Poland can make an agent. And so a thousand agents got built. So now we're like, Oh, what are those costs?

[00:21:00] And if you actually choose an LLM in the dropdown, so now the LLMs have little dollar signs next to them to show you, Hey, you're, you're hooking up to an expensive LLM or you're hooking up to a cheap one that's fit for purpose. Next step is to not even allow people to do that and let what will have governed rules on the backend that decide which model is used, which one's suitable for purpose. But now with a thousand agents out there, my CFO, and I'm sure a lot of others who's asking, how much does each one cost? Which one's the most expensive? Which one's the most effective?

[00:21:29] So I think what's about to happen. And I was a finance guy by trade early in my career. One thing I learned early is chargeback showback fixes a lot of problems. So right now centralizing AI for control is a good thing. But if you don't disperse the cost and people don't even know what they're spending, they're not going to be responsible. So now we're going to see a lot of in the industry and in my own company, like, okay, it hits your budget now. Oh, well now I'm going to make a different decision than I was going to make a minute ago.

[00:21:57] And when corporate, you know, sugar daddy was carrying the expense, I'm going to spend what I want. And now chargeback showback, like, no, you're going to have an explicit budget for AI and you're going to have to work within those parameters. So I think in the next six months, you're going to see a ton of that. And for anybody listening in an organization that is building AI into their operations today, are there any architectural decisions you think that could better determine whether they remain flexible as models and providers and regulations?

[00:22:27] All these things will continue to change, continue to evolve it. Anything you'd advise here? Yeah, the old IT ad is loosely coupled, tightly integrated. The guy that we brought in from Meta was very specific about this and knowing this environment is going to change at a pace we've never seen before.

[00:22:47] So building a system that is, as usual, a collection of tools from different vendors and such and then stitched together internally, you have to have the ability to swap stuff out because the leaders of the race are going to change constantly. And when you see it happening, even at the highest level, you know, OpenAI just came out with, you know, they were kind of written off and now they're making a comeback. After Anthropik had been written off sort of last year and then all of a sudden they burst ahead in front of everybody else.

[00:23:14] So when that's happening with trillion dollars, maybe not quite trillion dollar companies, all the little upstarts at the lower edges doing interesting things like, for example, FinOps. Like there's an explosion now of small companies building FinOps tools and or services. And no doubt, you know, some of them will look good now. And then in a year, you're like, oh, but something's come along that's much better. I want to be able to disaggregate that tool and put in a new one as the case may be.

[00:23:42] So I think I hear a lot from my clients about lock in. And again, in this environment, I think it's probably as important as it's ever been with the tradeoff being we've had this conversation about cloud, the people not wanting to get locked into Amazon or Microsoft or Google. And I would always use, again, the iPhone as an analogy because people would resist adopting some of the microservices that the cloud providers made available. But those are really useful.

[00:24:12] Well, I don't want to get locked in. I want to be able to do cloud arbitrage. And I would tell them, do you really think that you will find some arbitrage opportunity between Microsoft and AWS and you'll be able to move your app from cloud to cloud and save some money? They're hyper competitive with each other. So it's like walking around with your iPhone going, I don't want to use any of the apps because I'll get hooked on it. Well, then why do you have the phone? There's this tradeoff between the two.

[00:24:39] Like you want to adopt maybe the special sauce that a given AI vendor has without locking yourself in fully. And that's a balancing. I don't have a good answer for that, but that's going to be the balancing act. And if we have a CIO at the beginning of their AI journey that is listening to our conversation today and they're more interested in that long-term vision rather than the instant gratification and shiny distractions, etc.,

[00:25:05] what advice would you offer to put them in place or what would you tell them to put in place first to ensure that that infrastructure remains resilient, adaptable and ready for whatever comes next? I know that is almost impossible. It's a great vision to follow rather than just the next big thing. Well, clearly you should hire a managed service provider called Ensono to run it. That's a given.

[00:25:26] But other than that, I would say when it comes to hiring talent, I've seen, even in my own company, maybe a tenancy or a move to like, ooh, I've got someone in my team that I want to reward that's been playing around with AI that I think has shown promise. I'll put them in charge of creating this loosely coupled tightly integrated platform.

[00:25:50] You should think hard about that and bring in sort of fresh talent that kind of knows AI from the ground up and can institute sort of a new way of thinking because you really have to be well-versed in this stuff rather than, you know, maybe it's been 20% of your job for the last year and you've been trying to figure out how to do AI on things instead of AI in things.

[00:26:12] So bringing someone from the outside that spent 100% of the last three years in AI, I would recommend CIOs do that versus you should try to reward existing associates. I couldn't be more about that, but you got to have a combination of those things happening. If you're not bringing fresh talent into the company that knows AI from the ground up, you're probably making a mistake. I absolutely love that. I think that's a powerful moment to end on. And for anyone listening wanting to dig a little bit deeper on anything we talked about today, we did cover a lot. Where would you like me to point everyone listening?

[00:26:43] On sonar.com. Awesome. Well, I will include a link to the website and indeed your LinkedIn there. And my LinkedIn. Yeah. You'll reach out to me on LinkedIn. I obviously love talking about this stuff. I don't know what happened in my career that moved me away from finance to tech. Like I'm wildly unqualified for the tech side, but I get lots of experiences that maybe make me a little qualified. Yeah. I love it. Well, I'll include links to everything.

[00:27:06] I think as we covered today, before we started recording as well, as AI investments grow, so do expectations around business outcomes, governance, and accountability. The question is no longer whether to invest in AI. It's how to do it responsibly while still moving fast. I think we hit the sweet spot today, but thank you for bringing it all to life. Thank you. Thank you. Always a pleasure, Neil. Good talking to you.

[00:27:28] Today's conversation, I think, was another timely reminder that successful AI strategies cannot be built on excitement alone. Yep, they depend on resilient infrastructure, thoughtful architecture, and flexibility to adapt as models, costs, and regulations all inevitably change. Now, Brian also challenged that assumption that legacy technology is always something to replace.

[00:27:55] And how showing the right foundations can become an advantage rather than a limitation when combined with AI. So I'd love to hear your thoughts. Is your organization building AI on foundations that can adapt over time? Or is it increasingly becoming dependent on technologies that are outside of its control? Let me know. Techtalks. So techtalksnetwork.com.

[00:28:22] You can find 4,000 interviews there, how to work with me, connect with me, or just send me an audio message. Whatever it is I want to hear from you. but thank you for listening today and i'll be back in your ears tomorrow morning bye for now