Your AI Strategy Is a Document — The Operating Model Is the Difference

Paul Tyler (00:01) Hi, this is Paul Tyler, and welcome to another episode of the L&A Hub Podcast. We've got a great episode today about technology and the insurance industry, AI, and what it actually means to drive business results. With us we have a terrific guest, Sean O'Donoghue, who is the Chief Digital, AI, and Technology Officer for Security Benefit and Eldridge Wealth Solutions. Sean, welcome.

Sean O'Donoghue (00:34) Thank you. Great to be here. Appreciate the opportunity to chat.

Paul Tyler (00:37) You've got an enormous mandate. You're working for some very good clients here who want to push the envelope on what technology makes possible. But maybe first we could talk about your backstory. You've got an interesting trajectory — into insurance, out of insurance, back into insurance, is how I'd characterize it. Tell us about the path you've followed.

Sean O'Donoghue (01:05) Broadly, I've spent about thirteen years in professional sports, media, and entertainment, and that's sandwiched by thirteen years I originally spent in the London reinsurance market, working at Lloyd's of London and so on. Then the last eleven years or so back in insurance. So it's about twenty-four years in insurance all told.

It's been an interesting journey coming back. When you've spent time in entertainment and sports, one thing that becomes ingrained in you is the idea that every minute of attention you have to earn. Nobody has to buy a season ticket. Nobody has to go and see a movie. You earn that through the experiences you create — and the technology always needs to become invisible while you deliver those experiences.

Coming into insurance, and in particular the retirement space, one of the things that struck me is how much of the industry still deals with people performing high-stakes transactions on paper forms, on hold music on a phone. Nobody likes hearing "please allow seven to ten business days for me to process this for you." To some extent that's the way it's always been — competing on product and distribution, not so much on experience. The experience has been wrapped around the products for the carrier's convenience, not the customer's. I think that's where it's going to change. Straight-through processing becomes table stakes and the norm. And in a business where relationships matter, building that relationship through experiences with advisors and agents is what's going to make the difference in the future.

Paul Tyler (03:50) It's interesting. I've heard a few people say every company is going to be a media company at some point. I like your framing — it's not a media company, it's an experience company. Is that right?

Sean O'Donoghue (04:04) Yeah, everything is experiences.

Paul Tyler (04:07) Let's talk a little about the experiences. You mentioned a few of them. How do you think about the experience? Clearly there's the consumer, and especially in life insurance and annuities these tend to be infrequent but very important transactions. I just gave birth to somebody, somebody died and I'm filing a claim. How do you think about whose experiences are the most important for carriers to focus on, which are least, and — if we're going to be creating new experiences in the future — what do those look like?

Sean O'Donoghue (04:57) I think where this goes is, while we're not all there yet, table stakes is going to be transaction processing — end to end, straight through. I speak about that in terms of the button you click to make a change to your contract, the button that lets you perform a transaction. Being able to do those at speed and seamlessly is clearly needed, and a lot of people are getting there. I know at Zinnia this is a major focus.

But it levels up beyond that. What I call it is the conversation around the button. It's being able to provide context just in time for the advisor or the customer, ahead of performing the transaction. How do they know the consequences of what they're thinking of doing before they actually do it — before the first time they realize the consequence, which maybe is in their statement at the end of a quarter? How do you surface to them, just in time, the consequences of what they're thinking of doing?

For the advisor you can take it one level further. When they come in on a Monday morning, how do I give them the list of clients where something has changed or shifted over the weekend or over the last week — so they know somebody's free withdrawal period is now open, or there's some other shift because of dates and times where things are now available within their contract? It's surfacing to people the options they have, surfacing it to the agent so they're fully informed, just in time, ready to speak with their client without having to go do a lot of research and follow-up. It's almost a just-in-time service of real context, information, and intelligence.

Paul Tyler (07:47) One area of experience I'd like to get your sense on — how important is the employee experience at carriers? We've got a lot of people who've been in the business a long time, the baby boomers aging out. How much does the experience of the employee start to matter?

Sean O'Donoghue (08:09) I think the experience of the employee can actually become elevated. Even as we go into this world of AI — today we have people follow a process through step by step. As we're introducing AI, we have what everybody calls the human in the loop. An AI agent, or set of agents, moves something along, and along the way an employee stops, looks, and decides: okay, keep going, this is good, this is not good.

Where I think we get to ultimately is what I'd call human above the loop. The whole process can run and surface all the things I mentioned — the full context, the full consequence — and the employee is in a position of making an assessment of whether this is right. It's judgment. It's a higher level of thinking required in the role, which I think everybody would appreciate — being able to add more value that way, rather than moving things along in a process.

Paul Tyler (09:34) By implication that means new training, new hiring, new career paths. Maybe this bridges nicely to your ITL piece — which I really liked. You framed a question I've heard a lot of times: "Do you have an AI strategy? What's your AI strategy?" As if I had a business strategy. You're describing something more, and you talk about it as an operating model. It feels like what you're describing is that with this new technology comes a new way of actually operating internally. Talk a little more about that.

Sean O'Donoghue (10:09) The way I'd think about it: every carrier now has an AI strategy, and frankly for many people it's a document. An operating model is something that says how we know who decides, who builds, how it's governed, and how the data flows every day without needing a meeting. That's an operating model.

The test is pretty simple. When a use case shows up, does everybody already know how it gets evaluated, funded, built, and monitored? If the answer's a steering committee, then you've got a strategy. If the answer is "I have a system," then you have an operating model. The companies that make it through aren't necessarily the ones with the best tech — they're the ones with the operating model.

Paul Tyler (11:13) It's interesting. Building an operating model is a lot harder than a strategy. Much, much harder. I love that notion — if you need a meeting, there's a problem. If we looked at the results — I've been on both sides of these, where you have to create the big governance committee to run your AI and your pilots — why do some of these really good ideas die? Is it because you had a strategy, not an operating model?

Sean O'Donoghue (11:57) They're tightly related. The real reason things die is what I think of as the handoff moment — when it goes from a pilot stage to production. I need production data, I need a production owner, I need a place in someone's actual workflow for this to live. Often these are built as point solutions. There's no orchestration platform underneath, no governance path in front of it. So it succeeds technically but fails organizationally.

With a lot of what I've been hearing and reading lately, I'd move that a little further and say in 2026 the stall itself has moved downstream. Getting to production — people are figuring that out. Most people are in production somewhere. The stall point now is the value conversion. The deployment is live, but the value is still not being realized. Getting to production without an operating model is really just relocating the failure point.

Paul Tyler (13:36) You're also potentially just adding cost. If I roll out ChatGPT and Claude to my entire company, I think I just raised my expenses.

Sean O'Donoghue (13:43) That's true. Putting in place the operating model on which all of this can live — to convert a deployment into business value — requires what I'd really call unglamorous work. You've actually got to redesign the workflow around the AI. Back in the late '80s there was the whole business process reengineering wave, and in many ways we need to pull that back off the shelf. We don't just need to make processes go faster with AI, because we don't need to do stupid things faster. What we want to do is rethink the whole process from first principles, then use the AI to optimize how we're delivering an outcome.

That involves moving people into different work, as well as retiring the old processes — hopefully allowing people to do more value-adding work. And we've got to remember to measure the result. All of those things are what I'd call the operating model work. That's the part a lot of people skip on the way to production.

Paul Tyler (15:26) I love your analogy back to the '80s. Was that Michael Hammer? Michael Hammer, yeah. Those projects had tremendous impact on American businesses that I saw. Back then it was put the butcher paper up, cover the wall, and take the marker out. I'll date myself as well. It was effective. What's different this time? I ask because a lot of the best AI ideas at my prior company came way down at the bottom of the organization, whereas that '80s work was very top-down driven. Is this the same process, different technology — or a similar process with a slightly different approach that really releases the value in a company?

Sean O'Donoghue (16:27) In part it's many of the same processes. Things have evolved, hopefully, but broadly you're still performing the same process, still trying to drive to the same outcomes — but what's around us is now different. The technology's different. How people think about adding value is maybe even different. So we have different ways of solving the problem, but the fundamentals of thinking about reengineering — what I call process debt — are fundamental.

It's one of the things we spend a lot of time on at Security Benefit. Everybody's familiar with the term tech debt, but tech debt and process debt go together. A lot of technologies you struggle to get rid of because the process around them has been laid out and ingrained over decades. Getting an organization to shift away from the process debt it carries is frankly the hardest work.

Paul Tyler (18:01) Going back to technology — I think the technology is radically different, and the other dimension is the speed at which it's improving. Back in 1990, if you ran one of these big process reengineering projects and it came back five years later, how much had changed? You thought there was change, but nothing like now.

Sean O'Donoghue (18:17) It enables velocity, for sure. We've all heard people come in and say, "I created an app over the weekend." Frankly, who can't? It's not that difficult now, if you have the tools and a basic understanding of what you're trying to do and the outcome you want. Anybody can create an app.

Creating an app for an enterprise — especially a regulated one — is a whole different story. They're in no way related. That's where the conversations around governance come in. And by governance I mean a framework, not a committee — a framework through which every AI use case gets evaluated, built, and improved. Without it you get many disconnected pilots, inconsistent data practices, and no path to scale.

With governance you build institutional muscle. Every time you lay down a new capability inside that framework, every subsequent project gets to go faster than the previous one. Data is a core foundation. Governance without a unified data foundation governs nothing worth scaling — and without unified data you're not personalizing, you're guessing. So governance is critical, doing it up front is important, and your data foundation is one of the first things you want to have governed.

Paul Tyler (21:01) Let me push a little on the tech debt. To give you a sense of where I'm coming from: two or three years ago, call centers were probably the easiest place to put AI, because you already had infrastructure, structure, the data you're describing — a very process-oriented environment. Back then, if I said I want AI analytics in my call center, I want to listen, screen-scrape, you had to build a lot of it yourself. There were no pieces in place. Fast-forward to today and a lot of that is almost out of the box from some of the call center platforms. So all the stuff you might have amortized over a six- or eight-year period now suddenly needs payback in a shorter period. How do you think about that?

Sean O'Donoghue (22:01) Again, if you have a governance framework, it helps you make these decisions. Take the call center. If I wanted to create a new one, I could look at a lot of third-party vendors, and most will say, "Here's all the functionality, use all of our AI tools." To me the gotcha is: and you need to move your data into our environment, and you need to use our AI tools.

How we think about it: if the future is about creating personalized experiences, that comes through the data I have that really understands our customers. The AI tooling and orchestration platform I have would let us build exactly what our call center operatives want, and we can adapt fairly quickly to their needs. What we've done through our governance structure is create what I've mentioned — our AI orchestration platform. It's a platform on which we can drop any of the solutions we build. They all inherit the same data governance. They all inherit full observability and traceability of everything that runs through it. That's how we learn quickly from the people calling into our call center, as well as understanding our own operatives and what they need.

Everything's going faster, and that velocity is enabled by having a governed framework. If the rules are laid down, people can build faster because they know what the guardrails are. You know your lane, you can build faster, and you can focus on delivering outcomes within it.

Paul Tyler (24:54) Let me shift the conversation to where the industry is headed — you've painted a pretty interesting picture. I see some of the rules changing. Around 2011 it started to pick up steam: let's not keep our servers and data on-prem, let's shift to the cloud. That made sense for a lot of applications — websites, things with spiky usage. Now we go to this AI world. Maybe we keep our data ourselves. And it's expensive — maybe expensive to keep a dedicated server running full time on Amazon or Azure versus on-prem. How does the hardware strategy change?

Sean O'Donoghue (25:46) For us it wouldn't be coming on-prem. We'd own all our data. We use Amazon and get a lot of value out of everything they offer. The governed data is all the data sources we have — whether it's in a Snowflake environment or in some of our AWS environments — and we govern it as one. That's critical.

So for a company like us, certainly at this point, on-prem, your own hardware, wouldn't be practical. It may in the future, but there's so much shifting in the industry and in the technology and AI space in general. I think people are at risk of hopping to whatever seems like the trend at the moment, and then two months later it's something else. So where you're already efficient, stay the course, see where those things play out, and pick the strategic choices based on the framework you've put in place for the company to make its decisions.

Paul Tyler (27:28) Shifting to the software side — buy versus build. When I first got into insurance, I worked for Sean at a big company, and I'd sit down with the IT department. The first thing they'd tell me was how many years it had been since they touched code — it was a badge of honor. Now we have senior people vibe-coding apps. Do you think carriers are going to be more apt to take the approach you did — take all the pieces, stitch them together, and own that code base in the future?

Sean O'Donoghue (28:05) What I should distinguish is what I consider the future experience layer — the surface where our customers meet us and learn about us — versus how we actually process the business. For a company like us, Zinnia is the best-in-class organization for processing our kind of business at scale. No holds barred. There is no way we would ever think about trying to replicate what Zinnia does.

Where it differentiates in the future is that different customers have different needs and expectations. And by customer, it can be a firm of financial professionals, or a consumer. Say we're talking about a distribution partner. They may simply want a dashboard where they can process business — and that's great, Zinnia has a solution that solves that problem. You may also have carriers who are very invested in extending the relationship they have with their distribution partners. It's a relationship business.

In this business, what I see happening is the financial products, to a large degree, merge. They don't become one, but there's not a huge amount of differentiation between them. If you invest in trying to be the most innovative financial product manufacturer, you're spending a lot of money to differentiate for not a long period of time, because these products are fairly easy to replicate — it's all in the public filing. So because of what's now possible with AI, I think the expectations are going to shift dramatically, and AI enables that. You get the ability to create, for a customer, that just-in-time set of consequences about what they've not done yet. They may be thinking about doing a withdrawal — how do you tell them all about that before they go ahead?

To actually implement that, you have to be very thoughtful about deterministic versus probabilistic. You certainly don't want an AI model determining how much somebody should withdraw — that would be a very bad idea. But your deterministic models can be very specific about the data they're pulling, and then your probabilistic AI solution can take that data and wrap it into a conversation and context, just in time, for your agent or the end consumer.

Paul Tyler (31:46) I love it. I know we're right up at time. Related question: if you were an advisor or a client three years from now — and a lot can happen in three years these days — how would you describe what the best carrier feels like, the experience it provides?

Sean O'Donoghue (32:15) It's an extension of what I was just trying to say: every transaction arrives wrapped in its consequences, so customers and advisors act with context instead of discovering later. Our operations teams would run above the loop — which means agents take work from intake all the way through to outcome with a full audit trail, and people spend their day on judgment and exceptions.

What it gets down to later on is whose intelligence is answering the question. In my view, the carrier's intelligence answers wherever the customer shows up — in the carrier's own experience, or inside the advisor's workflow. The one caveat: right now you've got to be very careful about how you use the public models. The last thing you want is the information you've made available being converted by a model into a recommendation. If you've made certain information accessible to models like ChatGPT or Claude, and somebody asks, "How much should I withdraw?" — you do not want it saying, "Well, with your company's product, you should withdraw this amount." That's just not allowed right now. There's very clear regulatory guidance that AI agents cannot give advice to buy or sell to end consumers.

So what's important, until the day that's resolved, is that you build those guardrails into your platform — part of the governance. What I call the advice boundary: if anything looks like advice, it must always go over to the advisor. The advisor is the licensed person to give the advice. It never crosses over to a customer. A customer, I can give them information, and it stops there. But you do not want an AI model deciding in a prompt what kind of advice to give the customer. That guidance is built into infrastructure. In the world of AWS, that would be a tool like AgentCore, where you're very clear about what is said where.

Paul Tyler (35:36) Interesting. This is great. What you're doing at Security Benefit is fascinating. If people want to learn more or contact you, Sean, what's the best way?

Sean O'Donoghue (35:54) I can be reached through email — Sean.ODonohue@securitybenefit.com. Also on LinkedIn, I think it's Sean W. O'Donoghue. Maybe you can post those in the notes for anybody who wants them.

Paul Tyler (36:11) Absolutely. Sean, thanks so much. I want to thank our listeners — be sure to recommend us, forward this to friends who might find it interesting, and join us again next week for another good episode of the L&A Hub Podcast. Thanks, Sean.

Sean O'Donoghue (36:27) Thank you, Paul. Appreciate the time and the opportunity to chat. Take care.

Paul Tyler (36:31) Thank you.