Sam Worthington is Chief Data Officer at Crux Underwriting, a London Market (multi-line)specialty MGA. In this episode of The Practical Innovator, he speaks with Unitary's Nasri El-Sayegh about Crux's approach to AI, automation, data and scalable growth.
Episode summary
Featuring: Sam Worthington, Chief Data Officer at Crux Underwriting
AI in insurance is usually discussed in terms of underwriting models or claims automation — but for Crux Underwriting, the more important transformation is happening in the data foundations underneath all of it. In this episode, Naz talks to Sam Worthington, Chief Data Officer at Crux Underwriting, about how the London Market specialty MGA is building automation that lasts: starting from business strategy rather than technology, deciding what to build versus buy, and drawing a hard line between where AI is trusted to act on its own and where a human still has to sign off. Sam also makes the case that, done right, automation doesn't strip the human element out of a relationship-driven business like insurance — it buys back the time brokers need to actually spend with their clients.
“Brokers don't come to us asking for AI. They want their lives made easier. The best automation is invisible: we quote fast, we take the paperwork off their hands, and that frees them up to actually advise their clients instead of doing admin. It's not making our customers less human. It's giving them more time to be more human.”
Sam Worthington, Crux Underwriting
Key takeaways
- Start with the business problem, not the technology. Crux starts with what the business is trying to achieve, then works out where AI and automation can help. Sam compares it to the dot-com boom: technology can accelerate a strong strategy, but it won’t fix a weak one.
- Automate the bottleneck that’s actually slowing the business down. Crux initially assumed quoting speed was the problem for one major broker. In reality, the bottleneck was policy documentation. Removing that operational constraint is what made higher-volume business workable.
- Use AI where it adds value, not everywhere. Crux uses AI for specific tasks like extracting and structuring data, while deterministic rules and human oversight control what happens next. The level of automation depends on the risk and the tolerance for error.
- Buy the commodity technology. Build what gives you an advantage. Crux uses external foundational AI models rather than trying to build everything itself. Its internal investment goes into the things that differentiate the business: its data model, underwriting logic and accumulated business knowledge.
- The goal of automation is better work, not less human interaction. Brokers don’t care whether AI is involved behind the scenes; they care about faster service and less admin. By automating repetitive operational work, Crux gives brokers more time to advise clients and focus on the parts of the job where people add the most value.
About The Practical Innovator Series
Hosted by Nasri El-Sayegh (Naz), VP Commercial at Unitary, The Practical Innovator is a series featuring the people changing how insurance organisations operate. Each episode profiles a leader challenging established ways of working, introducing new technology, or taking a practical approach to AI and automation — with a focus on the real decisions, experiments and lessons behind meaningful operational change.
Full interview transcript
Naz El-Sayegh (00:03)
Hello and welcome to our Practical Innovator series from Unitary, where we're talking to people who are changing how insurance operations happen. We're talking to leaders who are challenging those established ways of working, introducing new technology, looking at automation and their approach to AI. For those that don't know me, I'm Naz — I look after everything customer at Unitary, supporting teams from ideation all the way through to automation. And it's a great honour to be joined today by Sam, Chief Data Officer at Crux Underwriting. Sam, why don't you do your own intro?
Sam Worthington (00:33)
Well, thank you very much, Naz, and I'm very I'm delighted to join your show.
Naz El-Sayegh (00:38)
Lovely, great to be here.
Sam Worthington (00:39)
I chose Crux for a few reasons. First of all, the business strategy appealed to me. Crux is a London Market specialty MGA, multi-line, and the business is very much about finding the profitable niches that will generate long-term, through-the-cycle profitability. Those niches can be anywhere — it's about finding the cracks in the floorboards that are underserved by other parts of the market, where it's therefore profitable to write. Crux is able to do that in a different way to other MGAs because it has a single connected view of risk that allows it to automate quoting and policy issuance, so it can do straight-through processing — with the benefit of making brokers' lives easier, which means Crux is able to access risks and distribution that other MGAs can't.
And as Chief Data Officer, and effectively head of the entire actuarial function, I had the chance to combine data and actuarial, which gave me a big remit to shape how we organise ourselves from an operating model perspective, and how we set ourselves up for success. My key role is designing the data foundations and the data pipelines that allow us to ingest and process data so that we can automate underwriting and the back-office processes we need to do — for example, bordereaux ingestion, or producing our outwards bordereaux — and designing the data structures that support that.
Naz El-Sayegh (02:24)
Makes a lot of sense, Sam — that's really interesting, and it's quite a unique take on a London MGA. With that in mind, and finding those niches of profitability across the market, where are you focused right now — both at an individual level and as a company, with the future growth plans you've got?
Sam Worthington (02:42)
Good question. It's grounded in our business strategy — finding these cracks in the floorboards and enabling ourselves to write them. There are two broad strands to our current work. The first is growth: we have good foundations at the minute, and we're building out the product set — for example, into property — and the distribution network we have, to increase the flow of business running through the Crux platform.
The second is strengthening the foundations for scalability. When most MGAs think of scalability, they think about hiring people. Our business is very much about using technology to minimise the need to hire people — our scalability is about putting the tech in place that allows us to grow without increasing headcount. As a guide to that, a third of our team works in product, data and AI — seven people in a company of 21. That's a big investment in terms of the cost of doing business, but it sets us up for long-term growth.
My focus is around that single connected view of risk — how we build the canonical data structures and the pipelines to be able to process that.
Naz El-Sayegh (04:00)
That's super interesting — what I really want to dive into is what you just said. If a third of the business is in data, AI and tech, and you mention it's quite a big investment, why have you opted for that?
Sam Worthington (04:12)
It's a good question, and it comes back to the principle of who we are. We're building a long-term business — our ownership is primarily individuals, with a lot owned by the founders, the employees and other private individuals. There are many MGAs out there that want to sell in five years. That's not us. We're building for the long term, and we know that if you get the foundations right, that's a prerequisite to being successful in the long term. Technical debt pays you back just as heavily and just as fast as interest compounds on a bank loan — you've got to deal with it, you've got to set yourselves up correctly. We're building reusable foundations, so the next product launch is easier than the last one, not increasingly difficult because we haven't thought about it.
Naz El-Sayegh (05:06)
Very true — I can relate to that, and I'm sure lots of our listeners will hear that and think, hang on, we could do something a bit different with our operation. Very inspirational. You mentioned you want to automate as much as possible — that's the basis of the whole organisational structure from a tech and risk perspective. Where do you see the automation opportunities? Where can AI support that within Crux today?
Sam Worthington (05:31)
Our approach is to start with the business strategy and then define our AI strategy in terms of how it delivers value for that business strategy — so it's grounded. There are two strands to this.
The first is that we use it directly in underwriting. I love a story from our co-founder and CEO, Mike O'Connor. Mike was in discussions with the biggest broker in the world, and went in thinking he knew what their problem was — that it was on the front end, about how they get business, how they could quote really quickly. He talked to them, and it turned out he was completely wrong. The problem was actually the back end. What the broker said was: great that you can quote automatically for us, but we need to process all those policy documents manually, and our team's not set up for it. So it's great if you can quote and write a thousand policies for us, but we can't issue the documentation quickly enough — so it doesn't work. What Crux said was, we'll build that capability so we can issue the policy documents on your behalf, and that resolves the broker's pinch point.
That's what happens — you keep finding these bottlenecks, and you run as slowly as the tightest bottleneck. So it's that automation of quoting and policy issuance that's key for us to make viable the writing of large-volume, small-premium policies, which the London Market has historically struggled with.
The second thing is how AI fits into the key pillars that drive our business. There are three engines for growth, if you like — three forms of capital. There's human capital, which is the expertise of our teams. There's financial capital, which we use to build product and pay our people. And the third pillar is intelligence capital, which is the accumulation of business knowledge. Every business — every insurer or MGA — has data and experience, but in many cases it's lost or just not usable. That's what we're doing: structuring it so we can capture it. So what we're using AI and automation for is to extract, store and structure every submission, every underwriting decision, referral and meeting note, so we can use agents to generate insights about what's actually happening in our business and feed that back into improving underwriting. As we store more data, our intelligence capital grows and compounds.
Naz El-Sayegh (08:23)
I really like that, and I don't think many businesses would have looked at an AI and automation strategy to have such a byproduct — or even put that at the forefront. Most businesses today have an AI board, an AI committee or an AI strategy. But what I'm curious to learn is — you said do business strategy first and then use AI to support it. What happens if a company does it the other way around?
Sam Worthington (08:45)
I see parallels and echoes. It's like Mark Twain said — history doesn't repeat itself, but it rhymes. There are things we can see that have happened in the past where we can see shadows and shapes of what might happen. I think if we look back to the dot-com boom, when the web was first introduced, every company built a website. Some of the prevailing views were that the only way you would sell in the future would be through a website — you had to get on the web or your business would die.
What actually happened is that companies with sound business strategies, who adopted a web strategy that supported it, amplified their success — Amazon is obviously a great example of that. Companies with weak business strategies just amplified and sped up their demise — pets.com is a great example, spending loads of money advertising, selling essentially heavy, bulky pet food at low margin, at a loss, and it managed to bankrupt itself very, very quickly. I suspect we'll see something similar, or at least the same sort of shape, with this wave of AI enthusiasm we're currently living in.
Naz El-Sayegh (10:07)
I agree, absolutely. You had a lot of ideas earlier about where you're seeing automation — submissions, bordereaux, policy document creation. How do you actually prioritise? How do you say, let's do this one first, let's sort out our bordereaux? What are some of the metrics behind the decisions that guide you and the team on what to automate?
Sam Worthington (10:30)
It's grounded in business value, and to assess that we ask three key questions. The first is: does it resolve some of the bottlenecks that allow us to do straight-through processing, which is key to our business? The second is: does it help our customers — does it give a better user experience, allow them to operate more effectively or quickly, or give them fewer touch points? And the third is: does it save time for our team — does it allow us to scale our business without having to hire more headcount? Those are the three areas we look at to determine whether something should be prioritised.
Naz El-Sayegh (11:11)
Now, with so many opportunities, how do you decline options? How do you say no as a team to an automation or an idea?
Sam Worthington (11:17)
Right now it's more like prioritising and saying not yet, rather than no. We're living in a time where the technology is evolving quickly, and use cases and supplier capabilities are evolving very quickly too. So it's more about what are the right things to do now, and building a backlog of the ideas our team has, that we can pick up at the right time in future.
Naz El-Sayegh (11:44)
Yeah, I like that — so it's not a straight decline, never to be seen again. It depends on the idea, right?
Sam Worthington (11:51)
Exactly. I mean, some things may never be seen again, but they're on the backlog.
Naz El-Sayegh (11:57)
We hear a lot about — and it's something we support teams to try and clarify themselves — do we build this or do we buy this? On that whole debate, which I think is getting ever more prevalent with the introduction of more sophisticated, perhaps easier-to-use models in the LLM world, how does Crux think about that kind of decision?
Sam Worthington (12:16)
We've invested a lot of the company in product, data and AI, but we're still a very small team — we don't have large resources, we're tiny compared to big MGAs or carriers. So our preference is to buy where we can, and we buy anything we deem commodity. Foundational models, for example — there are multiple different ones, and they're all very good; which ones are best changes very quickly, so we're not going to anchor ourselves to just one of those. In fact, we use all the major models for different purposes across our organisation — we stay open and receptive, and that's very much tactical, swap in, swap out as needed.
Where we think it's worth investing our scarce resource is where it's building a moat — something that's not a commodity, that actually differentiates us and protects us from the competition. That's the canonical data model — how we actually structure our data, which is proprietary and valuable — the underwriting logic we use to determine whether to accept, decline or refer a risk, and the intelligence capital, the business data we're accumulating. All of those are things we need to own and build ourselves.
Naz El-Sayegh (13:41)
Absolutely. As the AI models you've brought in get better and outpace each other, what we're seeing more is a view that, okay, I can give this to an agent — I can give this whole task to an agent and have it done for me. Do you have a debate internally, or with yourself, about at what point you let the human let go? At what point can you remove the human entirely from the process, and how do you decide that?
Sam Worthington (14:06)
That's a good question, and an interesting, philosophical one too. To some degree it comes back to — I like Unitary's approach of it not being just one AI to do a large task. Essentially, it's a lot of deterministic rules and software, plus a thin layer of AI for the certain elements where we need an LLM. That's the approach we've adopted with our operating model, and I call it controlled autonomy. Where we allow AI to play is very narrow and distinct, and it's controllable. For example, we use it for submission ingestion — AI is used to extract data into structured form.
We say AI proposes candidate data — it's not taken as the authoritative truth, but rather it's extracted, flagged as what's been extracted, and comes with a number of different signals. Then we have a set of deterministic rules that determine how we process that — do we decide it needs human referral, or can it be cleansed? That means we can review, update and control that set of rules to determine what happens and how it works.
The second key thing for us is that we see this question very much as a continuum, rather than can this be done by AI all the time, or does it need a human all the time. It's a combination of using AI in some contexts and circumstances, to different degrees depending on context, and switching between human referral and using AI for straight-through processing.
To give some examples: when we're ingesting bordereaux, which is used for contractual documents, this has to be as watertight as it possibly can be — nothing's bulletproof, but the tolerance for error is minimal, tiny, negligible, and we surface everything we can't process automatically for immediate human review. When we're ingesting a terrorism schedule of values — imagine we've got locations all over the US and we can't geocode some of the lampposts in Ohio — well, it just doesn't matter, because that's not a material location for the risk, there's no accumulation there. So we can say, just disregard that, it can be processed and ignored. Depending on the context, and our business appetite for getting it right, that determines our approach. The key thing is we use human judgment where expertise adds value.
Naz El-Sayegh (17:18)
I really like that — that's a wonderful way to frame it. And human value also translates to customer experience, right? That's a reason your customers deal with Crux and routinely come back — it's not just that you've got the best coverage and pricing, and you've managed to insure weird and wonderful things, there's a human element to it. So do you ever fear that automating processes is going to deter a customer from working with you because it's lost that human touch?
Sam Worthington (17:46)
No, I don't, actually. Our customers are our brokers — our direct customers are the brokers we interact with. It's interesting — a few years ago there was a question over whether there'd be broker disintermediation. Was there still a need for a broker, when you could process and buy your insurance automatically? And actually, in this country, we do have high degrees of people buying insurance online without a broker, from a personal lines perspective, and that seemed to be the direction of travel. But there's still a lot of people who buy insurance from a broker.
I remember speaking to some friends who live in the US — insurance professionals — and I was interested in their perspective. I asked why they buy from a broker, and my friend said: because if I have a claim, I'm going to be on the phone to the broker asking, what are you going to do for me to pay my claim, what are you doing right now? That ability to have a person you can interact with, at the time you need it, was extremely valuable to her. I think it's one of the things that's often underestimated about insurance — that kind of crisis management. As well as the financial recompense you get, there's someone there to help you — for example, in a travel insurance situation where you've lost your documentation. If you go through a broker, you can have someone on your side, fighting your corner. And interestingly, since then I've come across more people in this country, insurance professionals, who say they buy insurance through brokers too, because they know their claim will get paid.
Coming back to your question, in a very long-winded way —
Naz El-Sayegh (19:39)
And I've got a follow-up to it now, but yeah, go.
Sam Worthington (19:41)
The brokers don't come to us and say we want AI or automation. They want us to make their lives easier. And we do it — we use it, and it's invisible to them. So we help them look good because we quote very, very quickly, and we take away the work they need to do in terms of producing policy documents.
What that means is we give them more time to be able to talk to their customers and advise them on the things they actually need advice for, rather than doing admin work. So I would flip it around and say automation is actually allowing our customers to become more human, and not less, because it gives them more time with their clients.
Naz El-Sayegh (20:22)
Preaching to the converted on that one, Sam — I love how you frame that. I wanted to ask what you think brokers are sensing from this, but you've answered it for me. Let's talk a bit deeper into the future — you mentioned Crux is still a relatively small MGA, you're building the data moat, that speciality, the singular pipeline. Where's that competitive advantage going to grow? How do you see that improving over time, for both you and your brokers?
Sam Worthington (20:52)
I think our mature and established competitors have far more data than us — we're new, we've been writing terrorism business for eighteen months and had no claims, so that doesn't help us much in terms of claim prediction or loss modelling. There's just no in-house data that we've generated. But the thing is, our competitors' data is often not accessible.
I like to think about it as — imagine you had a library stuffed with every book in the world, but it wasn't organised, and the books had pages torn out. The data content is there, but it's not consumable, it's not accessible. Our philosophy is very much around building that. The data moat is about constructing our data in a way that's accessible, and building the feedback loops so we can improve our decisions about the next referral or renewal — we've got trading data, risk data, exposure and claim data, and we can put it all together to generate insights.
Naz El-Sayegh (21:58)
I like it, and I can almost predict the answer here, but let's say another MGA comes along, a competitor of yours, forensically looks at your tech stack, and decides to just copy it — and does it successfully. Would they still be missing things?
Sam Worthington (22:14)
Almost certainly. Our tech stack is not off the shelf — it's not just a product that does its thing, there are heavy degrees of configuration that we make of it, through our team of experts. They'd need to replicate the expertise we have in product, data, AI, marketing and underwriting to implement the software the way we have — and they'd need the high levels of collaboration we have between those different teams, the right people with the right mindset and willingness to collaborate.
It's interesting — Mike O'Connor, our CEO, observed to me that he's been working for 30 years, and at Crux he works differently to how he's ever worked before. He said it's like — we have to work in lockstep across marketing, data, product and underwriting, so that it's like the three-quarter line in rugby: you've all got to move up as one, and stay in that line, or it falls apart. You've got a leg, and the opposition runs through. It's the same here — you need to replicate that collaboration. And the way we've set ourselves up is to implement our business strategy specifically. So if you had a different business strategy, our tech stack and configuration might not be appropriate for it — if you were focusing on one particular segment, I wouldn't approach setting up the business, or the operating model, in the same way we've done at Crux.
Naz El-Sayegh (23:40)
Interesting. Sam, there's been some amazing insights over the last half an hour — really value your input and everything you've said. I guess a final question to end our episode on: if another business leader was in the room with you now, who wasn't too clued up on AI and hadn't taken it seriously yet, what advice would you give them?
Sam Worthington (24:01)
First, I'd say start with your business strategy, and ground your AI strategy in terms of your business. Second, build a data layer — work out how you store data and get feedback loops on it, and build that intelligence capital. Work on that, build the data foundations. And with some humility, I'd say our collective knowledge of AI is immature — things are moving rapidly, there's huge potential, and the limitations are unclear. So I'd say focus on proven outcomes, from vendors like Unitary who've done it and won — get some successful wins, get some success, before you roll out.
Naz El-Sayegh (24:39)
Brilliant, Sam — it's been amazing to have this conversation with you, and I think the team at Unitary are all very excited to see what you and the team at Crux do next. We'll be following the story very closely.
Sam Worthington (24:48)
I'll be happy to talk to you again, Naz.
Naz El-Sayegh (24:51)
Thank you so much, Sam.
