Reinventing Operational Scale through Agentic Underwriting

Summary

CFC, a Lloyd's of London cyber specialty insurer, built an agentic underwriting system in production that takes an email submission and generates a fully fledged insurance quote ready for underwriter review in under 15 minutes.
Built on Databricks' Delta Lake architecture and LangGraph, the agent reduced quote turnaround from 24 hours to 5 minutes — a 99% time decrease — and cut cost per quote from £5 to £0.50, a 90% cost reduction.
This video covers the engineering journey, the human-in-the-loop confirmation process, change management at go-live, and lessons learned from deploying an agentic system at specialty insurance scale.

Reinventing Operational Scale through Agentic Underwriting

CFC as a Lloyd's of London leading Cyber speciality insurer, is globally scaling our business by amplifying every Underwriter. For this we've created the "Agentic Underwriter" in production to augment critical steps in the underwriting journey. By combining Databricks' delta lake architecture with modern LLMOps processes and open source frameworks like LangGraph we've rapidly created agents capable of creating fully fledged quotes. The agent is able to intelligently navigate our systems & the complexities of the speciality lines insurance process to deliver ready made quotes for our underwriters to confirm or adjust via a HITL process. The result is our underwriters have trusted quotes ready in 5 minutes vs 24 hours - a 99% time decrease. Equated to cost, we have quotes ready for £0.50 vs £5 - a 90% cost decrease. Our customers are feeling the impact with quotes arriving faster and prices more competitive. In this way, we have created the industry's first true agentic underwriter.
Talk By: Christopher Mullan, Head of Data & AI, CFC ;

Chapters

FAQs

What is CFC's Agentic Underwriter?

CFC's Agentic Underwriter is an AI system built on Databricks' Delta Lake architecture and LangGraph that takes an email submission for specialty insurance risks and generates a fully fledged quote in under 15 minutes. Underwriters can then confirm or adjust the quote through a human-in-the-loop process before it is issued.

What cost and time savings does agentic underwriting deliver at CFC?

The Agentic Underwriter reduced quote turnaround from 24 hours to 5 minutes — a 99% time decrease — and cut cost per quote from £5 to £0.50, a 90% cost reduction. This frees underwriters to focus on high-value, complex risks rather than routine quote generation.

What are specialty insurance risks and why are they hard to automate?

Specialty risks are unique, large, or highly specific insurance needs — such as satellite launches or cyber policies — where risks are not homogeneous and require navigating complex rule books and underwriting guidelines. CFC trained their agent on the same rule books used by human underwriters to handle the full complexity of specialty lines.

What lessons did CFC learn from deploying an AI agent in specialty insurance?

This video highlights the importance of assembling a cross-functional team and managing organizational change carefully at go-live, particularly in a regulated specialty insurance environment. CFC also found that human underwriters remain essential for the most complex or novel risks where the agent defers judgment.

Full transcript

[00:08] Uh so good afternoon. Uh I'm Chris Mullen. I'm head of data and AI at CFC and I'm here to talk and share about how we've reinvented the operational scale through a aentic underwriting and what it took to get that live it for real in
[00:21] production in a specialty insurance business. So, I'll try and cover off the who, the what, the why, the how, some of the so whats, and end with some lessons learned about how we did so much so quickly and
[00:33] the kind of impact it's already having at CFC. Uh, so who am I? Like I said, I'm Chris Mullen, head of data and AI. All the boring stuff's on LinkedIn, so if you want that stuff, go look for it there.
[00:45] What you won't find on LinkedIn is I'm a father of five. I'm a recovering actuary, and this is only my second ever time in San Francisco, so I'm really enjoying it. And then who are CFC?
[00:57] Football fans, soccer fans, sorry, uh will be disappointed to know that I'm not here to talk about Chelsea Football Club. I'm in fact uh here looking after CFC, a private equitybacked global MGA.
[01:09] Uh we specialize in uh cyberme and special cyberme and specialty risks. We were founded in 1999 as click for cover as a com hero dist when insurance first
[01:24] started to get disrupted by internet and digital based distribution. Um one of the leading global cyber insurers and we've been growing for like 20% kagger for 15 years and that's an important fact that comes into play later and that
[01:38] that that growth that improvement over time has always been powered by innovation in our products in our technology and in our data. So what did we do? Why did I get the
[01:50] opportunity to come and speak here today? So this is what we did in one sentence. Pretty simple. It's an agentic system that allows us to go from email to quote in under 15 minutes for specialty risks. But let me take a
[02:04] second. We'll go through that sentence backwards and hope we're all end up on the same page. So what are specialty risks? Right? These are unique and difficult risks. I'm sure we're all familiar with like personal line stuff, your car insurance, your home insurance,
[02:16] your health insurance. Businesses have their version of that, commercial lines. So, it's not home insurance, it's warehouse insurance. It's not car insurance, it's fleet insurance. It's not health insurance, it's business interruption cover. Specialty starts to
[02:30] kick in when those risks are no longer homogeneous, when the risks are really large, when they're very specific or they're very valuable. So think uh a satellite launch, a supermax uh
[02:43] container ship, this convention center, the the giant stadium. Those are the types of risks that end up in the world of specialty. Um, and increasingly businesses are much less defined by
[02:56] their physical property, their plant, their their trucks, much more defined by intellectual property, their cyber estate, their reputation, their M&A dealings. And that's the part of the
[03:09] market that we play in, the sharp end of the market where underwriting still truly matters. It's not homogenized risk. You can't just chuck it at a simple pricing algorithm and leave it be. So that's the specialty part. Next
[03:22] bit, email to quote in under 50 minutes. So insurance is like at least a 300 year old business. I know some people argue it's two 2,000 years or more. Because of that history, because of that longevity,
[03:34] it's got a whole bunch of built-in ways of working, a whole bunch of locked in path dependence that drive how everything works today. It's also necessarily a very intermediated humanto human expert to
[03:49] expert distribution chain because it takes a lot for a business to understand its risks for those to be worked with and interpreted by a broker and turned into the kind of product needs that can be taken to the market. And as a result
[04:02] of all of that human to human and expert to expert, the market principally transacts via email. So, we get loads and loads of emails with like PDFs and word documents and
[04:14] spreadsheets and JPEGs and TIFFs, all kinds of stuff. And clearly, this is a distribution chain. This is a way of working that was not designed to be digitized. Uh, and it makes d starting directly from that
[04:28] inbox an absolute necessity to meet the market at its natural entry point to meet our customers where they are. So, we have to start from that entry point and we have to handle all of those
[04:40] unstructured documents, all of that complexity and all of that data and insight uh from the get-go. And then finally, like I said, it's an agentic system. Um, it's not an agent.
[04:52] It's not one thing. We didn't just chuck our underwriting guide into chat GPT and go have a go. Um, we built multiple specialized agents. I'll go into a little bit of that.
[05:04] produce evidencebacked recommendations, product configurations, coverages, subjectivities, all of that stuff. So, we had to build a thing that represented micro decision by micro decision what an underwriter did. So, that's an agentic
[05:17] system that goes from email to quote for specialty risks in under 15 minutes. And if you put all of that together, you're transforming certainly you're doing something that transforms how we do business at CFC. But we know as much as we want to build
[05:32] all of these agents and change the world, build a future that whatever we do with the agent, they will always we will always have humans working alongside them, working with them. So we knew setting off when we did this, we can't introduce
[05:45] um complete separation. We can't introduce whole new weird dimensions of operational complexity. uh we had to build and design a system that had a lot of sympathy and empathy for the business, for the underwriters
[05:57] and for the ways of working. And the simplest way to think about this is really yeah, it's one underwriting system with two different types of users. Historically, what did we do? We we hired an underwriter, we gave them training, we gave them the guide, we
[06:10] gave them an email, and we hoped in the minutes, hours, days that followed, we'd get a submission back. What do we have today? kind of exact same thing. We give the agent the underwriter guide. We give
[06:23] them the email and within minutes we get a risk decision and a quote recommendation. But it's important to note that everything that the agents doing is doing it through our existing systems. Uh and that was a really important
[06:35] design approach to build a lot of confidence in the rest of the business to let us do what we wanted to do because what it meant was we were using the same battle tested enterprisegrade scalable disaster recoverable
[06:48] infrastructure controls governance observability alerts that sit on top of our existing systems and our existing human underwriters. the agents had to go through the exact same control gates, the exact same steps with the exact same
[07:01] amount of reasoning. The result is we get to use the same rule books for the agent as the human. We use the same controls and it just keeps everyone happy. It was a lot easier to do it this way than try and build something completely on its own.
[07:19] So, how did this kind of come about? Why did we do this? Um, I think especially given the news and what's going on, a lot of companies tend to start with start their answer to the AI question going, "How do we use it to cut costs? How can we have fewer engineers? How can
[07:31] we, you know, make everything cheaper?" And my CEO came to me with a very, very different challenge, which really ties back to that history of growth at CFC. How do we scale? How do we triple our
[07:45] revenue globally without really adding any headcount? And that's a much better framing for for someone like me and our team to to grapple with rather than how do we use cursor or something to fire people which is really boring.
[07:58] But in that question there's actually a whole bunch of embedded sub questions that I started to figure out how I was going to come up with a good answer to it. So yes, it's great that the first bit is it's a challenge about using it for growth and not not not reduction.
[08:11] But also underneath that question is how do we remove a lot of those boring, really time consuming, manual, repetitive tasks from an underwriter's day without replacing them with some completely unyielding computer says no
[08:23] automaton. Really important.
[08:35] limit to how many you can operationally support. If we make that drastically cheaper, we can serve way more of the market for the same kind of cost that we're prepared to spend. Yeah, LLMs are magic. What does it mean
[08:47] if we can now suddenly take in emails, application forms from basically any language anywhere in the world? Does that enable us to scale our operations globally a lot faster, a lot easier than we're able to before?
[09:00] And also crucially, how do we build an agent and a system, an assistant almost that works alongside our underwriters and elevates them, elevates their expertise and also elevates their experience of coming to work at CFC.
[09:18] So, how so if that was the why, what's the how? And this is a very non-useful diagram deliberately. I think I'd get into a lot of trouble if I gave you something that would help you too much. Um, but if we had to do a dry architecture diagram and you ask a complete non techy to do it, this is
[09:30] what you get. So, what do we do? Well, like I said, we built an an email system to take in those emails, pull them apart, take all the structured data out, use that, represent that in a bunch of different ways. You have to start at the front door.
[09:44] And like I said on the one user, two systems point, we had to build a system that used the same rulebook as our underwriters. So the next thing we had to do was build a little machine, a little agent, a little thing um that could absorb the same underwriting rule
[09:57] book and then understand it, decompose it and recompile it into the kind of structure, language, model, computer system that an agent could read, understand and
[10:10] effectuate. Um, and on top of that, we also recognize that in the real world, what's written down in the document might not always track what's actually happening in the market. and how things change. So, we also had to build a system, a
[10:23] little agent that watches and is constantly looking at real life practice, real life quotes going out the door and understanding how do I close the gap, how do I keep it fresh, how do I make sure that the agent instructions match what what seems to be the emergent
[10:37] instructions for our underwriters. On top of that, we had to build a thing that really replicates what an underwriter did. We had to really understand what an underwriter does step
[10:49] by step. And that might sound a bit trivial, but almost replicating underwriter chain of thought into a system micro decision by micro decision. So what are they doing when they open up
[11:02] that email and they're clicking around the application and they're clicking around some PDFs? What's going through their head? How are they going about understanding this risk, understanding this client? A good example of this would be one of
[11:14] the questions we have to answer is what is the industry classification the activity code of this client if it's three accountants in an office in de mo it's pretty easy these guys are
[11:26] accountants but what if it's deote right as a as a business grows in scale grows in complexity grows in operations it become increasingly hard to determine a single dimension a single answer to the
[11:40] question what does the company do I mean if you ask people about deote they might say they do tax or they do they do regulation or they do implementation um there's not a single simple answer for that and that's one of those areas of judgment an underwriter has and
[11:54] that's what we tried to teach the agent but it also turns out when we did that there isn't actually an activity code for a deote because there isn't a code that says they rob your watch to tell you the time so there's still some things that you have to go beyond the
[12:07] agent and make it really understand how our systems work um and also So in doing that in in in building that system that answers one question at a time what you find is that the agent can perform much much better than an underwriter. But we'll get on to
[12:19] a little bit of that later uh as we go through the talk. So what's the result of all of this? A hundred little micro decisions. Every little step every little chain of thought of the underwriters then captured in a whole bunch of micro decision agents. That's
[12:32] the system that we built. We had to reflect on how do humans think? How does the process work? Then of course finally you know some bits to talk to the legacy legacy systems and make sure we could actually effectuate and execute everything we needed to. Of course then
[12:45] we had to orchestrate and make sure we had something that went left to right sitting on top of of course a modern nice clean excellent data platform. Thank you data bricks. And then wrapped up in the kind of normal MLOps LLM ops
[12:58] governance observability layer that you need to do things safely and securely. And once again thank you data bricks. So if that's a pretty non-useful how, I'm going to give you some more really
[13:10] non-useful hows because I don't want you to catch up too quickly. So timeline. So we started the business, we started the business unit, we created the team and I was hired in like June 24. The amazing data engineering team basically had the
[13:22] data platform ready to go with our data modernized by September. Um, and then in May 2025, I I took a bit of a punt. I kind of put
[13:34] probably the cheekiest uh business case of my life to directly to my board and it was genuinely one slide with one line and it just said uh if you give me 2.5 million pounds I will
[13:48] create 250 million of value in 12 months. Um you got to go big or go home right sometimes. Um and they they said yes and I was as shocked as you are. Um
[14:01] they had some follow-up questions. Obviously, it wasn't just that simple, but that was the only thing I put to paper and that was my central argument. And really, at that point, it was a test of them. It was a test of the business.
[14:13] It was a test of the business business's desire and want and culture to do this stuff. Were they willing to have a go? Were they willing to try? Did they have a desire to be at the front? Did they
[14:25] have it a desire to lead? And by doing it that way, I got the answer. Uh, yes. Um, but obviously I was slightly taken a back. I was probably expecting to be told, you know, settle down, come back with some more thinking, uh, and sound
[14:38] sensible, but they said yes straight away. So, it took a bit of time to get my together and figure out who I needed and how I was going to do it. And we kicked off in September. Um, by December, we were in live in production with our MVP. We spent
[14:51] January to March really refining and building the product on top with our underwriters. User testing complete end of March, live in the UK uh, in April. And as of now, it's working for new business and renewals globally in the US
[15:04] and the UK. And we're binding policies that have been issued by an agent. Um, on the kind of tool chain, another unhelpful answer. It's all the usual suspects. I don't think there's anything
[15:16] on there that you're going to be like, "Oh my god, if only I had that, it would work." Lang chain, langraph, of course, paid for conformance, of course, AI gateway or whatever it's called today on observability and governance and control. ML flow, Unity catalog, the
[15:29] usual suspects. But these are just all of this stuff. This is just the visual artifacts of everything. This isn't really the answer as to why the whole thing worked, I think. So, here's a slightly better answer.
[15:43] Um, it just comes down to the ruthless application of a of a product mindset. The old DF classics, can we do it? Do people want it? Does it make any money? Um and having previously been like
[15:55] product in a SAS and and and product in corporate they are very different and in corporate land um desiraability dominant and feasibility dominates almost the entire time especially early on but dominates
[16:07] way more than the other do um because to be honest most people in the corporate can tell you what they want or think they want or what the CEO says they should want and things like viability in the corporate context when you're buying
[16:20] building and supporting proprietary software is a bit fuzzy. You're not the commercial business. You're supporting a commercial business. You're not selling ARR license fees. It's a bit fuzzy. What what really dominates is feasibility.
[16:33] Um can you do it? Can you do it safely and securely in a regulated environment in the constraints of your IT estate with all of the controls with all of the cyber security you need? Um it's a it's
[16:46] a very different challenge. And so really really focusing on the desiraability. keep doing it wrong. Sorry. The feasibility um is what made it work. So making this a bit more real. How do we actually get it live? So
[16:59] basically over the past three quarters starting in December last year, we kind of cycled through the DF and the V, but really focusing on different points as we went. Starting in September, we built those core fundamentals. We solved the hardest
[17:11] technical uncertainties first. Can we run an agent? Can we run 10 agents? All of that stuff. Can we orchestrate it? Can we bring it right to left? Obviously, this bit took us about three months to get right and get working. And if you're in a green field, I'm sure
[17:24] it's a lot quicker. I'm sure it's a lot easier. I'm sure it's a lot faster. But green field companies typically aren't writing billions of dollars of premium. And I think the other important point here was we kept it all in stealth mode. We didn't really talk to anyone. We
[17:37] needed to prove internally within ourselves that we had the certainty and the confidence that this would work, it could work and we could build it before we started inviting uh opinions from the guarding functions. Um and we were very
[17:52] deliberate to keep it quiet, keep it secret, even only let in one or two people from underwriting before we opened the floodgates, we opened the opinions and we uh unleashed all of the uh corporate slowdown machine on ourselves. But look, by December, we had
[18:04] those raw unfinished tech capabilities and it was good to go. We knew at that point there was something that was one day could be a product that worked. Um, but we had solved all the major technical hurdles. Then we moved on to
[18:16] co-designing the future with our uh underwriters. So we sat down with them, show them what was possible, listened to them, understood their needs, their workflows, and built something really really quickly, iterate iterating all the time um to build a thing that would
[18:29] work for them. This is when it was all about a bit more of that internal selling. Uh, and we couldn't have done anything like this without agentic engineering. I know it's all the rage, but honestly, it's absolutely transformed how we do things. Without
[18:41] agentic engineering, we wouldn't have been able to deploy something in the afternoon based on feedback we got in user testing in the morning. And we wouldn't have been able to like figure out how any of these undocumented legacy
[18:53] services worked without being able to just chuck an agent at it. uh and navigating that complex estate and that complex code without documentation would have been possible. And also during this time we spent an awful lot of time on
[19:05] the human factor stuff, human
[19:18] spent a lot of time thinking about how it would play out. And at that point, yes, that's when we start engaging your legal, your risk and compliance because at that point there's some clay on the table. You're not just having academic arguments over PowerPoint about
[19:31] dogmatically or academically how things are meant to work. There is a thing in front of them that they can check and look at and they can understand. And genuinely doing that saves so much time and energy for for everyone. And
[19:43] obviously by this point we're starting to get live. to really pushing that business ownership, making them decide what the risk appetite was, making them decide what the uh risk criteria and success was. And you know, a lot of technology projects fail with the
[19:56] business change when it feels like technology is being done to the business. So that was really about building a kind of collaborative mindset where it felt that uh technology was supporting underwriting in building their own future and being partners.
[20:10] I didn't want to own risk appetite. I didn't want to own what we needed to do. Uh, I just wanted to get it live. Um, then we shipped and iterated. We got it live as soon as we possibly could at the genuinely it felt a bit shaky on the day
[20:22] that we were really there, but getting it on, getting it turned on because everything's doesn't matter until you're live. Like nothing kicks your ass like prod. You've got to get there. You've got to see what works. You've got to find all the things you could never
[20:34] possibly plan for and understand without that real world pressure. With that being there, iterate, iterate, instrument the value. Becomes very quickly, very easy to demonstrate the impact you're having. We'll go on to
[20:47] that and how it all works. And as soon as we went live, it created all kinds of demand across the business. We basically had every line of business and every geography battering down our door and asking how do they get their bit sooner? How can they engage to go faster? How
[21:00] can they get their little piece? And we're already figuring out given this amazing harness thing we've built, how do we bring it over to the claims domain given so many similarities? That's the point where cash rules everything around you.
[21:13] So I said I'd talk a little bit about so what what's been some of the impact of this? How's it translating into changing our business? Yeah, 10x reduction in cost to quote. If I look at what it costs our
[21:25] underwriters, how much time all of that stuff, we're already 10x if not better and getting better. We haven't even bothered to do any phopsy type optimization. We haven't even bothered
[21:37] to figure out the right model for the task. We are just chucking everything at the frontier models at this point because we don't need to we don't need to absolutely optimize. Even at that sloppy um maybe not the most um
[21:49] disciplined data science, we're already seeing 10x. By the time we finish that cycle, I expect that number to go up even more. Already in live production, we're seeing quotes going out in under 15 minutes.
[22:01] Our average at the moment is about 14 minutes. That's a dramatic dramatic change in the broker experience and the broker service from being able to send an email and get an answer.
[22:13] And what does that translate into? Yeah, okay, maybe two to five times more quotes per underwriter. Still figuring that bit out, but what it means is very safely that 3x is within my error bars. I feel comfortable we'll deliver the
[22:25] core mission. Those are the kind of really hard easy to understand benefits. There's obviously some softer, some second order ones. So, kind of seeing a bit of a better hit rate. It's pretty obvious if you respond to your broker faster, if
[22:38] you get back first, if you make their job easy, you're going to get more follow through. They sent the email on Monday and you don't get responding till Tuesday afternoon. You've given everyone else in the market 48 hours to get a better answer in first. We can get back
[22:51] fast. If we can get back first, we get a much better conversion, much better quote to bind ratio. I hinted at it earlier. um but a lot more um consistency in in underwriting outcome.
[23:03] So using that business activity example that we talked about earlier, underwriters are they're risk experts, but they're also salespeople. Their job is to get quotes to bind.
[23:18] And with a little bit of experience, they can of course anticipate how the pricing system works. they can they can learn what happens if they press certain buttons in the pricing. So they know
[23:30] that if they call a client a professional services company, they might get a different price, a better price than if they call the the client a professional services company that deals with nuclear waste. Pretty
[23:44] obvious. They they know that that might have a big swing in the price. And you know, they're great people with great judgment, but they might be tempted to choose one classification over the other if it means getting a deal over the
[23:56] line. So by having an agent take those types of decisions, the result is a way more objective outcome and may way more repeatable approach to answering some of
[24:08] these risk questions compared to this subjective broker who kind of anticipates or appreciates how our pricing system works. And doing that consistent outcomes, consistent
[24:20] judgments, all auditable and traceable makes our heads of line really happy. They know they're not they're less likely to have like the rogue slightly tipsy after lunch uh underwriter making bad decisions. They know the agent's
[24:32] going to behave how they want it to over time. That makes our CEO happy and all of this makes our carriers very very happy. More more more uh precision, more repeatable uh uh
[24:45] better outcomes. Uh and most importantly, we are actually returning in a whole bunch of time to underwriters. And I know that might sound quite counterintuitive because a lot of the the the the spiel has been
[24:58] about speed and time and all of this stuff, but once you've taken all of that manual data entry, going through different screens, reading different PDFs, opening opening different emails, once you've
[25:10] taken that out of their job, even if you're asking them to do more of these quotes uh in a day, you're giving them huge amounts of time back. They're not having to go through all of these six screens to get stuff done and clicking a
[25:23] lot of buttons. And it means that as underwriters, they can focus a lot more on what underwriters are really good at, what humans are really good at. They can spend more time uh with the harder risks, the ones that do take a little bit of extra time because all of the
[25:35] easy ones have kind of been dealt with. they can spend more time talking to brokers, being out in the market, understanding broker client needs, coming up with product innovations, um, and just building those absolutely core
[25:48] human to human, expert to expert relationships that a chatbot can't replace. There's nothing that's going to replace shoe leather, reputation, and people. So, look, lots of really, really
[26:00] good stuff. I hope you'll agree. It's really awesome, but there's a whole bunch of stuff that we don't know yet. We're still kind of figuring out um we're still actively learning what does it mean to an underwriting
[26:13] organization, a structure, a team structure, a career structure when an entire department has been moved across to using these agents when all of them are supported by these the this this
[26:26] type of augmentation and automate automation. We have to design the job, the role, the process to make to ensure that our underwriters remain sharp, motivated, focused on doing their job
[26:39] and having fun doing it and don't simply become like I said earlier, auto approve, auto decline automatons. And the reality is you can't figure this out on PowerPoint. You you you don't know
[26:52] how it works. You just can't get a Tom consultant in to draw pretty stuff. You have to do it. You have to feel it. You have to see how your team dynamics change, your organizational structures might have to change. And we're still living that and figuring out and that's
[27:04] why we did the pilot as soon as we could so that we can understand with evidence and experience what the kind of transformation is going to be required. So we don't have a clue generally. Uh the next thing is we've very explicitly
[27:16] at this point chosen not to remove that human touch. The agent does absolutely everything, eats the email, configures the quote, applies all the subjectivities, calculates a tax, does all of that stuff, but it still comes to a human for the final decision and the
[27:29] push to client. And um one that just makes everyone happier, safer, feels like we're not quite changing the world just yet. Um, but the classic IBM apherism about computers can't be held accountable, so
[27:41] they can't make decisions is really true here, especially in this risk-taking commercial kind of decision- making. And also, no matter how good um all of the AI ever becomes at those types of tasks,
[27:53] reading rules, applying them, reading emails, taking uh data out, they can't read between the lines of what the broker has said or what they said in the pub or what they sent on WhatsApp on the
[28:05] unmonitored channels. they can't do that bit can't really do that kind of subtle understanding and price negotiation at the end there's still a huge role for that and it's again being clear about what we expect a computer can and do and
[28:18] be good at and what a a person really does and so keeping that bright line between what we can use the AI for and uh what we want to use people for is keeping us going
[28:30] I also promised I'd share some lessons learned I need to get a bit faster um so this you know all sounds quite fast and easy for us and you know relatively yes. Um so why did it work quite well for us? I think we didn't start too soon. I
[28:42] think if if you tried this say two years ago if you got over excited about chatbt 3.5 and tried then you probably just wasted a load of money and time. Uh we only went going when I was pretty bloody convinced that the tech was was good
[28:54] enough and we could go. I've kind of droned on about what makes our culture special. The risk takingaking, the the hunger, the desire to lead and change. You don't grow 20% a year if you're happy with the status quo. So, I'll
[29:06] leave that one. Um, I think we solved the thing, the problems in the right order. You know, step one, data modernization, data platform, make that stuff really easy. Step two, focus on the email front door, get that fixed, and only then jump to the operational
[29:19] process. If you try and do it in the wrong order, you're going to be let down by all of the other stuff. Um, we already had the APIs. might like when I when I started two years ago, this endowment I had was basically all of the legacy tech had the API in and
[29:32] out. That just makes everything easier. Otherwise, you've got to build that. As a business, we already comfortable with computers making decisions. CFC's been trading digitally via API for like 10
[29:44] years now. So, I wasn't the first person in the organization trying to convince people to let a computer make some decisions. Someone else crossed that Rubicon for me and I benefited from that. I talked about we embraced agentic
[29:57] engineering. I've only got three data scientists, one um MLE and like two software delivery people. So without aentic engineering as a force multiplier
[30:10] to the domain experts of my business and my technology stack, there's no way I could have got any of this done in this uh time period. And finally, yeah, we we picked the right partners. You know,
[30:22] I've got five kids. I know it takes a village to raise a kid and it takes something more than a village to raise a baby agent. And starting this off in that period after they said yes, it was having the humility to know what we do know, what we are good at, what we don't
[30:36] have, what we don't know, and what we're not good at. And then doing a bit of an Avengers assemble to find the right people to come and join in and help get it done. Um, come on. One more.
[30:50] There we go. Uh, oh no, wait. I've gone too far. That's my bad. So, if we were Where's it gone? If I was starting again, what would I starting again today? What would I focus on? Pretty obvious. You've got to have the right
[31:02] tools to start with. If you've got data bricks, you've got all of that stuff working and running. You've probably got 90% of what you need, if not 100 already. Um, you've got to have the right culture. Uh, it doesn't if you
[31:15] can't take risks, it doesn't matter what you're going to do. If you don't have people supporting you, doesn't matter what you're going to do. You need the right mindset both from the CEO all the way down to the um junior underwriter testing. If any of them don't understand
[31:29] what we're trying to do and the path we might be on, it doesn't matter. You need them to get that buy in so that when that underwriter is testing, he understands what we're trying to achieve and he doesn't go around saying everything's rubbish. You need the right
[31:42] data. You need the right team. Pretty boring. So, what were those lessons learned at the end of all of this? What have I figured out? Um, no one knows what they're doing. I mean, obviously, generally, none of us really know what we're doing. But in particular in this
[31:55] space, no one can actually say that they really know what they're doing. And I'm sure if you wander into the vendor booth, you'll find thousands of people on stands telling you that they they've got it all figured out. And I would just say like, don't believe them. Everyone's
[32:08] figuring this out. This is too new. You've got to figure a lot of this out for yourself. So have that mindset. Don't try and find a partner who tells you they've got it figured it out. Find a partner that's there to help understand it with you. Um, no one else knows your business like yours. No one
[32:21] knows how they how it works, why it works this crazy way. So, don't forget that when building your Avengers Assemble team. Um, you can definitely buy in some expertise. Don't get me wrong. You can definitely buy in some expertise, you know, raw capabilities,
[32:34] resource augmentation, bandwidth, and pointers on architecture and stuff like that. Uh, but don't forget that no one knows your business better than you do. And given given all of the lessons and the amount of times I've referenced things like culture, if you do want to
[32:48] get partners in, don't buy for logos, even ours. Don't buy for um deep expertise on the thing that's a couple of months old. Buy culture, buy mindset, buy people
[33:02] that are going to be additive to your team and understand what you're trying to do and why.
[33:15] to fail, you're going to have to reset every now and again. Classic, yeah, fail fast, learn early type stuff. But really, if you're afraid of how your business will respond to that, probably
[33:27] don't start. If you do not have that right culture to start off with, knowing that you're going to fail, knowing that there'll be some dead ends. If you're at all worried about that, chill. Don't do this now. Wait for it to be a really solved problem because without that backdrop, it doesn't really matter what
[33:40] you guys try and do. If you have the right culture, it's not a problem. Um, and then finally, um, yeah, look, there's a lot of froth. I think I heard the word agentic or agent about a bazillion times at the keynote today.
[33:52] It's a cool thing. Everyone wants to do it. Everyone wants to say they're doing it, but really don't fall in love with the tech. I mean, it's a bit of an apherism, but don't be someone who walks around with a solution in search of a problem in your business. You'll get
[34:04] laughed away. Um, I I didn't do this because I bloody love AI. I mean, I do bloody love AI. It's like my favorite thing, but I didn't set out to do it this way just because I wanted to solve the problem. I said to the team at the
[34:17] beginning, if it turns out we do not need agents to do this, I don't want them. just want to make it work and uh create the value. So fall in love with the outcome, fall in love with the value. Um assemble the
[34:29] best team people, assemble the best team of partners, get the technology and try and have a go. Thank you very much. Hopefully there's time for a couple of questions.

Learn more about the Databricks Data and AI platform.

The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.