# Your Company Is Already Using AI. Just Not Safely. _Dave DuPont, Shane Emmons, and Future Point of View's Scott Klososky on why carriers stall at their third AI agent, and what separates a governance policy from day-to-day oversight._
On September 17, Swept executive chairman Dave DuPont spent an hour with Swept founder and CEO Shane Emmons and Scott Klososky, founding partner of [Future Point of View](https://fpov.com). The premise was that at most carriers the adoption decision is already behind them, whether or not anybody made it on purpose. Something is running right now: a model inside the vendor platform you renewed in March, an assistant an underwriter has been pasting submissions into since spring, an automation somebody in billing built for herself over a weekend. What follows is what stood out to us from the hour, and then the full transcript. ## The wall at agent number three A carrier buys Copilot licenses for everybody. Somebody in claims builds an agent, somebody in billing builds two more, and they work. Then the list of the next thirty gets written down and the whole thing stops. Klososky sees this often enough that he raised it unprompted: "Now that's a huge step, and now we can get ourselves in trouble, and so then they stop." Emmons had been watching the same stall from the vendor side. A company he talked to the week before had a name for the employees doing the building, citizen developers. What worried them was not the chat-style tools but the agentic workflows already out in production, because "the cat's a little bit out of the bag of some of them are already out there." The stall is strange because it happens after the hard cultural work is finished. These organizations already argued their way past risk appetite, already shipped, and already have people who want to build more. Then they find that nobody can describe what the first three agents are doing well enough to sign off on a fourth. Klososky reads it as governance and risk appetite collapsing into one unanswerable question, where pausing costs less than answering. Emmons was more concrete: "most of the solutions out there take a whole bunch of engineering, and they don't know how to do that." That version has a budget consequence. Oversight built as an engineering project means the number of agents a carrier can responsibly run is bounded by the number of engineers it can assign to oversight, which at a lot of carriers is zero or one. Meanwhile the [AI nobody provisioned](/post/how-to-detect-shadow-ai) keeps arriving. ## Bots, agents, and agentic workflows Ten minutes in, Klososky stopped things to deal with a word problem. He had been saying automations. Emmons had been saying agent. Neither had checked whether they meant the same thing, and in most carriers nobody checks. Emmons's working definitions: a **bot** is a chatbot, one question and one answer, maybe with some lookup happening behind the scenes, but the shape is one forward and one back. An **agent** gets kicked off by something that is not a person, an inbound email or a workflow or some other trigger, and then works with as much autonomy as you left it. An **agentic workflow** usually starts with a person and then runs on its own, and what separates it from the static workflows we have been building for twenty years is that it exercises judgment about what to do next. Keeping the three apart has a practical payoff, which is that they need different oversight. A chatbot that answers a coverage question needs different evidence before anyone trusts it than a triggered agent that reorders a claim file, and the escalation path is different, and so is the review cadence. A policy written about "AI" has no way to express that. Emmons raised a failure that only shows up once the count gets high. Automations written independently in different departments start running into each other. "You never predicted an automation done by somebody in billing is interacting with an automation in underwriting and claims." Every one of them cleared its own review, and nothing at all reviewed the seams between them. Seams are where this kind of thing tends to go wrong, which is an argument for filing them explicitly in an [operational risk taxonomy](/post/ai-risk-taxonomy-operational-governance). ## Governance and day-to-day oversight are two different jobs "Governance is policy and procedure, for the most part, and then active risk control is what are we doing to watch over, minute by minute, the automations and all the AI that we've done." That is Klososky's split, and most carriers have not made it. He was blunt about the failure mode on the policy side too: governance that lives as a pile of documents in a Teams channel is not governance. It has to be monitorable and enforceable, and he expects more of that enforcement to be done by AI. Emmons handled the watching side. Conventional observability is built to wait. Something breaks, an alert fires, a human goes and looks, and that sequence works for systems that fail in recognizable ways. AI systems tend to change how they behave without breaking anything, and they can do it between one Tuesday and the next. So his answer starts before deployment: "starting with day one before you're ever putting it out there, take a fingerprint, understand the behavior of it, such that you can then supervise it and watch it." The reason the ordering matters is unglamorous. [Drift](/post/ai-hallucinations-vs-ai-drift-understanding-and-managing-ai-drift-for-long-term-success) is a comparison, and with no recorded baseline there is nothing on the other side of it. You end up holding a stream of outputs that all look reasonable and no way to show that this quarter's behavior matches what you signed off on. That demonstration is most of what [supervision](/offering/reliability) buys, and it is what an examiner is asking for when they ask how you oversee your models. ## Three classes of risk, not one Klososky splits AI risk three ways: operational and technology, workforce, and strategy. Most carriers have staff, budget, and a standing committee for the first, and improvised coverage of the other two. The [frameworks they borrow from](/hub/ai-governance-frameworks) are mostly organized around the first one as well. Workforce risk is where he spends client time, and he is unsentimental about it. "If you don't handle things right, AI can be something that makes your workforce hate working for you." He described adjusters twenty-five and thirty years into a way of working, watching auto-adjudication remove the parts of the job they were good at, inside organizations so pleased with their deployment pace that nobody was tracking what he called the feng shui of the workforce dropping. It is a soft phrase for something that turns up eighteen months later in retention numbers. Strategy risk arrived by way of a client, who pointed out to Klososky that failing to keep up is itself a risk and probably belongs on the register. Moving slowly costs margin and share, and he adds a third cost: "Some of your A-player employees leave because they think you're the Flintstones." Emmons connected that to hiring, which bites harder in insurance than in most industries. Careers here run long and turnover is slow, so a carrier can go several years without registering that it has stopped attracting the people who would have been running it in 2035. "Without having felt like you made an active choice, you're just not going to be able to find these folks, let alone know you lost them." The applicant who never applied is not an event anybody logs. ## "Fast follower" of whom? Fast follower is the most popular AI strategy in insurance and, Klososky argues, frequently not a strategy at all. "Tell me specifically of who. A fast follower of one of the big people in the industry? Somebody you compete with every day?" When no name comes back, the phrase has been doing the work of a decision that nobody made. Emmons went at the arithmetic instead. The organizations in front are compounding, in tooling and in the pattern recognition their people accumulate by using the tools every day. "Everybody who's ahead of you, their speed is compounding much faster. You are going to have a lot harder time breaking atmosphere and catching up to them." Following fast enough to catch a leader who is accelerating means accelerating harder than they are, which is a materially different plan from the one most fast-follower carriers have written down. On where the money goes, Klososky splits a budget three ways: workforce skills, meaning licenses and training and bonuses and the hackathon you run to get people interested; the automation layer; and customer-facing tools. He floated 50/30/20 as a shape and said plainly it was a shape rather than a recommendation. Emmons's addition is what keeps the split from being fragile. Every line item in it is a bet. "If we're only going to make one investment next year, there's a high risk that one's not going to work out." The usual ending is a single failed pilot that sours an organization on the whole category for two years, which is an argument for funding more of them rather than fewer. ## Where the winners come from The labs end up as commodities. That was Emmons's prediction, by way of Scott Galloway's line about airlines: brutal competition on nearly identical capability, margins compressed to almost nothing, and enormous value delivered to the people buying the service rather than the people selling it. "I think we end up becoming the winners of AI much more than the labs." Klososky ran the same logic one level down, inside insurance. Small and mid-sized carriers that get genuinely good at this acquire a lever for competing upward that they have never had, and he expects some of them to take share with it. His comparison is early e-commerce, where a few unremarkable-sized retailers turned out to be very good at it while several much larger ones spent a decade being bad at it. Emmons took the short side. Mid to large carriers that move slowly, stay bureaucratic, underfund the work, and believe they are already competent at it will watch revenue go flat. Many of them have spent years watching insurtech for the threat. Emmons expects it to arrive instead from a carrier a third their size operating in three of their states. Where the two converged was on what is left once price and service homogenize, which both of them think is coming. Klososky's example was a council of experts: a customer-facing panel a policyholder could talk through their actual exposures with, legal angle and business strategy angle at the same time, instead of one agent asking the standard intake questions. Emmons moved it to commercial lines immediately, then to distribution. Hand that panel to your independent agents and you have given them something nobody else can offer the communities they already sell into. ## Transcript

Lightly edited for readability. Filler and false starts removed, meaning preserved. Speaker attribution was reconstructed from context: the recording was captured on a single room feed with no speaker labels.

Dave DuPont: Hello, and welcome to our webinar: "Your company is already using AI, just not safely." My name is Dave DuPont and I'm the moderator. I'm the chairman and chief executive of... actually, I'm not the chairman and chief executive of this company. I'm the executive chairman of Swept.

Shane Emmons: Hi, I'm Shane Emmons. I'm the CEO and founder of Swept.

Scott Klososky: I'm the only one not with Swept. I'm Scott Klososky with Future Point of View. But let's say friend of Swept.

Dave: The topic of our webinar is that AI is already in your organization. Insurance companies, through employees, vendors, existing software platforms, and formal initiatives, are using AI. Today's discussion is not about whether your company should be using AI. It's how to govern it responsibly while continuing to innovate and create value for policyholders.

Before we get to the meat of our topic, I want to raise something that's been in the news: the existential risk, the x-risk, that we've been hearing about with AI. I'd like to hear your thoughts on that risk briefly, because we could spend our whole time talking about it, and on how it relates to insurance companies and how they should be thinking about AI.

Shane: It's funny, because while we probably could talk about it a lot, we don't need to. It is, quite honestly, a lot of hype and scare, things we don't need to worry so much about. There is definitely risk involved with AI that we can talk about today, but the one dominating the news cycle, the potential of ending civilization before the decade's out, is so intensely small that it just doesn't matter right now. We need to weigh the incentives of why people are saying that. For the most part, I think we should be focusing on how we can use it and make sure we use it safely, because there are far more actual risks involved than these Terminator-style end-of-the-world scenarios.

Scott: You can't say x-risk without talking about doomers. Now we have a word for it, and we have these different classes of people, and it's interesting out in the industry, seeing people line up. Jensen Huang and Mark Zuckerberg, Zuckerberg just yesterday, both lined up on the "this is not as big a problem as people are making it out to be" side as far as x-risk. They both said, look, there are risks about making sure that AI has controls on it. Make sure it's not biased or discriminatory, make sure it's behaving in appropriate ways. And we have the ability to control those risks. Even Microsoft has come out and said the same thing. So it's interesting to see that you have people on the doomer side, and then you have people on the other side, and I'm talking about the top people in the industry, lining up.

I agree with you. I think there is a lot of need, when you have a powerful technology, to make sure it's governed well, to make sure you have active risk control. I don't worry about x-risk. I just don't. As deep as I've been in this technology for a long time, and I know you have as well, I believe there's an off button. I believe we're smart enough to control the off button. What we're doing is panicking a lot of people who don't understand this technology very well, for no good reason.

Dave: Well, let's talk about the opportunities and risks, then. We could look at it by function, we could look at it by type of AI, but we're starting to see AI in insurance companies being implemented in underwriting and pricing, claims management, fraud detection, and customer service. And of course there are different kinds of AI. Where do you see the greatest opportunities and the greatest risks?

Shane: The greatest opportunity everywhere we go, especially in insurance, is around claims. It's one of your biggest expenses, and it's the thing you have to manage all the time, so working in claims is always number one.

The biggest risk in AI is a simpler one. These systems are non-deterministic, which just means they're not consistent. They're going to occasionally bring back different results. We talk about that all the time. Not baking that into your understanding of the AI projects you're rolling out, either internally or with vendors, is probably the biggest risk, because it's the risk of the project failing. You expected it to be right 100% of the time and it's right 90% of the time, and you get bummed about it once you're out there. The biggest risk to these projects succeeding is not accounting for the variability, and just getting started with it.

Scott: All right, I'm going to go a completely different direction, just for fun. I'm going to say the biggest opportunities are improving the IQ and the capabilities of our teams. Just raising the general skills by amplifying what people can do, because they become high-end users of AI tools. So that's one of the biggest opportunities: let's raise the general skills of all of our people.

The next level up from that is automations. Instead of picking a department, I'll just say, I don't care whether it's accounting or legal or any department. Automations, and I mean hundreds of automations. A lot of small ones, some medium-sized ones. To me those are the two big areas of opportunity in the insurance space: raise all boats on the effective IQ that people have, and then work on the hundreds of automations across the organization.

As far as the risks, it's the exact same answer. Okay, raise all boats with your people, but if they're using AI indiscriminately, or the wrong way, or the wrong models, or they're not adding the human element to the work they're doing and they're just copying and pasting what the AI says, then it's inappropriate use of these tools in your workforce. That's a risk that worries me. And in the automations, not having good controls over all of these AI automations you put in place, not having good ongoing oversight to make sure they haven't gone off the rails and started making wrong decisions. Automating inappropriately, taking out a human element where a human element should have been left in. I know my answer's more general, but that's the way I would look at it.

Shane: I think yours is absolutely the right answer, specifically, in departments and claims and underwriting and things like that. And I like yours too, on automating all of these small tasks. One of the risks that comes through there that a lot of folks don't think about is how these small tasks and automations start interleaving with each other. You never predicted an automation done by somebody in billing is interacting with an automation in underwriting and claims, and all these new interactions happen. That's a risk many of us aren't thinking about. We're looking at that single agent instead of how they weave together with themselves and with humans as well.

Scott: Okay, we should cover this early. You said agent, and I said automations. This is an area of confusion inside insurance organizations. It's the vocabulary. We talk about agents, or bots, or agentic, and we talk about automations. Do these all mean the same thing? Do they not mean the same thing? What would you tell somebody, just from a vocabulary standpoint, so people aren't confused when we say, hey, we're building hundreds of agents, or hundreds of automations, or hundreds of bots?

Shane: I see them as different, though I will admit to sometimes using them interchangeably myself. First and foremost, what most of us see are bots, and I would call them chatbots. It's anything where we're asking a question and pretty much getting an answer. Maybe there's a little bit of activity that happens in the background to find that answer, but it's mostly one-to-one communication: one forward, one back.

Then we get into what I think of as agents and agentic workflows. These get a little bit interchangeable. It really comes down to this: an agent is often something that is going to get automatically kicked off, maybe not by a human. A workflow comes in, an email, something that triggers it, and it can go do work autonomously or semi-autonomously, depending on your human oversight.

Your workflows are more often slightly stepped down. That's where I get too nuanced, and it's hard to tell, but it's where you're triggering something, usually personally, and it can go off and do its work. It might never come back to you, or you might see the end result somewhere else. Sometimes I bring in agentic workflow just to mention that these workflows have judgment inside of them, versus our traditional, more static workflows. But those are the three levels, and then we start getting to human oversight levels on each of them that are much more complex.

Scott: So when we talk in these terms, and use any of these terms, we're often talking about hundreds of these that are going to get built. And this is where the governance comes in. What I see in the insurance space, and I'm just curious if you see the same thing I see, is that there are a number of mid to large, sometimes smaller, insurance organizations that are a bit stuck in their AI journey. They gave Copilot and gave AI capabilities to the team, but the next step is starting to build all of these tools, automations, bots, chatbots, agents. And they build two or three of them. And then they go, oh my gosh, if we go build the next 20 or 30, that's a huge step, and now we can get ourselves in trouble. And so then they stop. Are you seeing the same kind of thing?

Shane: We're seeing the same thing. We were just talking to a company last week, and I really liked the term they used. They were calling their employees who are building these "citizen developers." So we've got these citizen developers out there building agents, and they're not so much worried about the chat style, where they're building a chat agent that they can just repeatedly use. They're building these more... these automations, these agentic workflows, or these true agents. And they are stuck and worried. They don't know how to govern them, and the cat's a little bit out of the bag: some of them are already out there. So they're trying to slow it down and figure out how do we govern this, how do we manage it. Because most of the solutions out there take a whole bunch of engineering, and they don't know how to do that. That is one that we see consistently, all the time.

Dave: Are there other reasons why organizations are getting stuck, as you put it? I do see this when I talk to insurance leaders and insurance boards, and I'm just curious whether the governance or the risk that you're describing is the only issue. Is there more behind organizations just kind of not being able to proceed?

Shane: For me, what I see is there is pure governance, or there's this risk appetite and the unknownness of the risk, so it's easier to stop. But you're consistently stuck between a rock and a hard place. They know that they need to use this to compete in their very red ocean environments, where they don't just have policyholders that are out there without policies. They're going to have to get it from competitors. That's where their competitors are.

But they have these questions about how to invest. Making even the questions less about risk: how do we know what to invest in? There's a new AI tool every day. The models come and go. ChatGPT was the darling for about two years. Claude's been the darling for the last nine months, but last week, for the first time in about a year, ChatGPT overtook Claude in terms of usage in the enterprise for a week. So we're having this up and down all the time. What horse to bet on, how are we going to do that? Sometimes it's just easier to go choose Microsoft Copilot.

Scott: I agree with all of those. I'll give you two more. One is just the variability in the regulatory environment. If you're in multiple states, every state is coming up with its own regulatory environment around AI and insurance, and then you're wondering what's going to happen with examiners. How's AM Best going to look at this? So there's the unknown environment issue. I have heard some boards and executive teams say, we wish we understood this better before we moved faster. We don't want to get over our skis and get in trouble later from examiners or regulatory.

And then the other thing, I'll go a bit of a different direction. It's the concerns about how this impacts our workforce from a human and then a technology standpoint. We're not clear. We've heard a lot about, oh, it's going to take jobs away. If we are automating parts of people's jobs, what do we do with those people? Are they going to be happier? Are they going to be unhappier? There are a lot of insurance organizations that, however many team members they have, they've been doing pretty well. They've been making money. This year was better than last year. They're growing. And so I see a little bit of, hey, we're not sure, when you start at scale automating a chunk of the workforce, what this is going to do.

And they like to be fast followers, a lot of the insurance folks I talk to, and we don't have a set, who are you following yet. And so for them it's like, we don't know what the impacts are going to be on the workforce, we haven't seen leaders that we can just fall in behind just yet. And so maybe I want to get out of the mud that I'm stuck in, but I don't have any vision of who I can go chase to get out of the mud. So those are other things I see.

Dave: You've both talked about risks and concerns, and I wonder if there is a good understanding of the upside of implementing AI, and whether that might be an issue. What are your thoughts there?

Scott: Well, for sure that's a problem, because now you're talking about vision. People have to have a vision of what my insurance organization would look like on AI. If I had a clear picture of that, and it's a robust picture where we're amplifying profits, and we're providing much better customer service, and our team members love this environment. If I had a clear vision and a clear picture of how this comes out, I'd be moving way faster. A lot of them are not capital constrained. They could make the investments in the insurance space. They just don't have the clear vision.

And I think that happens in the early part of technology cycles. It was no different when the internet came out and we started building a lot of websites. People didn't have a vision of what that website was going to do until later, when we got examples. And so there is huge upside with AI. But people like us, we live in this world, we see it more clearly than a lot of leadership teams and boards do. So that's the number one thing I would say: there's huge upside with this wave. In my mind, there is no world in which insurance companies can't be more profitable, operate much more efficiently, lower their risks, make their team happier, make their customers happier. No world in which this doesn't happen.

But I think leaders and boards have to be able to create a clearer vision that they can touch and feel, that gives them the confidence that this is how the AI wave is going to go. Because right now people are spending most of their time and energy on, well, what are all the things that could go wrong? What are all the risks that I need to control? And we do need to control the risks. It's just that, generally, insurance people can see the risk management better, because it's the business that they're in, versus the vision of what this could do for me.

Shane: I think the other thing that they really have the opportunity for right now with that vision is that AI is going to, and already does, give them a lot more control for what they can do with their technology. It really is true that you could be back to the 80s and 90s, where you were building everything yourself on the mainframes and being able to, bespoke, make your company do exactly what you want, versus buying software off the shelf that you have to align your company to in order to use. So I think that's the other personally exciting thing for myself, looking at it and going, how can we transform a lot of these tools that for the last 20 years have forced us into their boxes? Now with AI we can have the vision of what our company can do and not be constrained by the tools we're selecting, because AI is going to give us a lot more power to do exactly what we want to, and make our company different such that people would want to become our policyholders, versus just picking between generics.

Scott: That comment about picking between generics is interesting. Elon Musk and others have talked about how AI and robotics might bring a lot of people's services and prices to the mean, that everyone will get the same power and everybody will homogenize, in some ways, their products and services. And I think that that is a very possible future. And then you have to ask yourself, well, how do we differentiate ourselves? And there's a danger if you don't keep up with what everybody else is doing from a price and capability standpoint. But if you do, then in the insurance space we really might get to a point where everybody's auto insurance is priced about the same, if we just looked at something simple like auto. It's all going to be the same, it's all going to be easy to put a policy in place. Then you do have to ask yourself, what is the difference between the organizations?

And the intriguing thing to me is, I believe in the future it's going to be your human component.

Shane: 100%.

Scott: What will separate you, once you are elite with AI, is going to be your human connection with your policyholders, your agents, anybody that has a relationship with your company. But I think that's an interesting thing in banking, insurance, legal, and service industries, accounting, to think about: your ability to do services and the pricing may homogenize. And then how will you separate yourself? It's going to be human connection, way more than it is today.

Dave: That makes a lot of sense to me. The nature of your relationship with your customers, and the information you have to provide the service or product, whatever it is.

I'd like to dive into the risks that we've been talking about. We've identified they're there, talked about them in a variety of ways, and I think our audience would be interested in understanding: so what do you do about them?

Shane: There is a lot of risk, and I think first you just tackle them head-on. We see so many people sitting back, actively choosing, well, instead we're just not going to engage, and therefore we've avoided the risk. But choosing to actually engage with the risk is first and foremost for me.

And then looking at what the risk actually is. So looking at your chatbots, your agents, and so on. Making sure, I guess, first and foremost, a governance plan. So what's your governance? What's your human governance plan? And then going to and looking at these risks.

When I look at the risks, I'm looking primarily for how this agent, once it's deployed, is potentially going to change its behavior. Because we know it's going to change, and how am I going to catch that? That's probably the number one risk to find, and that's not difficult. It just takes what we think of as active supervision. A lot of what we do in technology today is we monitor things. We put observability out there and we watch for activity, and then that flags it to somebody when there's something wrong, and you go deal with it. With AI, because of the speed and the dynamism of it, you have to take a different approach in terms of the activity and how active you are. And for that, we think about supervision. So what is the risk? How fast can we catch this risk when it's changing, drifting, shifting, whatever it is? That's number one for me. So starting with day one, before you're ever putting it out there, take a fingerprint, understand the behavior of it, such that you can then supervise it and watch it. That's our first step on the risk path.

Scott: I don't like going second after you, because I'm forced to always have to do something different. All right, so here are my different comments than yours.

First of all, it's really important for everyone listening to this to be able to discern governance from risk control. I think governance is policy and procedure, for the most part, and then active risk control is what are we doing to watch over, minute by minute, the automations and all the AI that we've done. So I think it's important to separate the two concepts. If we're going to manage risk, we've got to have both. We've got to have great policies and procedures, and I don't just mean a bunch of documents in a Teams channel. I mean that we have governance that can be monitored, that can be active on its own, enforced sometimes by AI. And then there's active risk control. So I think it's important to separate those.

The next thing I would say is there are a lot of risks outside of just "is the technology working well or not" that we don't always talk about. The risk of impact on the workforce. If you don't handle things right, AI can be something that makes your workforce hate working for you. If you handle it well, it can be a blessing. And I'm watching clients doing both. I'm watching some that are so excited about just putting AI in that they don't look at the impact on the workforce emotionally. So you're taking work away from people, that for you sounds great, because you're capturing hours, and you're creating efficiency, and you're turning work over to an AI tool. But to some of your people in claims that have spent 20, 25, 30 years doing claims in a certain way, maybe they changed technology, but doing claims a certain way, and all of a sudden now you're doing auto-adjudication in a whole new way, and you've taken things off their plate. There are a lot of people that don't love that. And I see a lot of organizations that seem to have turned a blind eye to it. We're getting really good at putting AI tools in place, but what's happening is the mood, the feng shui of the workforce, is dropping. So I think that's a risk that we don't talk about, that we're going to have over the next five years: are you amplifying and making your workforce a healthier, happier place, or not?

I'll do one more that, again, is maybe a different way of thinking. This came from a client of ours one time. We were talking about risks, and the client says to me, isn't one of the biggest risks not doing well with AI, or not keeping up with the market? And I thought, well, sure, I should always have that on my list of risks, but I just assume that you understand that. So for this discussion I'm not going to assume that. There is a strategy risk. If you do not move at the right velocity, if you say things to yourself like, oh, we're going to be a fast follower, yet you have no idea who you're following. You're just saying that because you think it sounds good.

When leaders say to me, when I say, do you want to be ahead of your competitors? Do you want to be with your competitors? Do you want to be a fast follower? And somebody says to me, I want to be a fast follower, I say, okay, of who? Tell me specifically of who. A fast follower of one of the big people in the industry? Somebody you compete with every day? Because when you say you're going to be a fast follower, that's one of the most important things. Who are you following? Or don't say it. Pick some different thing that you would tell me about speed.

I think that that's probably one of the more dangerous risks, a strategy risk, not just operational and technology risks. You move too slowly, you cost yourself profits, you cost yourself market share. Some of your A-player employees leave because they think you're the Flintstones. So in the world of talking to people about risks, probably we need to start thinking in terms of, hey, we've got operational and technology risks, we've got workforce risks, and we've got strategy risks. And let's monitor all those. Let's be good at all those, and not just over-focus on operational risks.

Shane: I really liked a couple of things there. One is that risk of losing your employees, or potentially even just never getting those next stars. Especially in the insurance industry, you have very long-lived careers, and so there's always that slow turnover. And your next stars, they're going to want to work in these new ways. And if you're a fast follower, which I start to question how fast you can follow on some of these now too, you are never going to attract that next one. And so you are going to, without having felt like you made an active choice, just not be able to find these folks, let alone know you lost them.

The other one, too, that we didn't get into a little bit earlier, that's less about the risk but the opportunity, is that I see often in these organizations there's opportunity for new stars to emerge, because there are people who innately are going to be able to work with the technology better. They're just going to get it faster. Right today they might not be the star for whatever reason in their department, but I am always amazed when we go and do a training or something, and you find this person that it just clicks for, and they just wildly bring AI.

The other part of that risk, too, I was thinking, was the fast follower. Right now, when you're thinking about being the fast follower, you have to really make sure you understand that "fast," and know that everybody who's ahead of you, their speed is compounding much faster. You are going to have a lot harder time breaking atmosphere and catching up to them with the momentum. So if you're going to be a fast follower, you have to potentially really start to think about, how am I accelerating faster than what they're doing right now, because you're just never going to catch up otherwise.

Scott: All right, I'm going to take the one thing you said that's off your topic, I apologize in advance. I was looking at my watch, and we've got a few minutes, so I can make this comment.

I really like what you said about the opportunity thing: that there are team members now who are gaining AI skills and taking big leaps forward from where they would have been. I often talk about how new people, young people especially, that are coming into the insurance space are going to judge you on, are you the Flintstones or the Jetsons? That's what I always say. And then I realize that you have to be a certain age to just know who they are. Fortunately, all of us are. Those of you who don't know who the Jetsons and the Flintstones are, I apologize, you'll have to look that up.

I think you're right, it's going to be hard to attract A players if you're not the Jetsons. But I've seen the same thing with, let's say, developers. I remember talking to a CIO that said, hey, I have developers that on a scale of 1 to 10 were 3s or 4s. And I have some 8s. And he said what was crazy was, you give a 3 or 4 an AI and a harness, and they become a 7 in 30 days, 60 days. He said it's incredible how fast they come up to speed.

I heard the same thing about people with Excel use. You can have people that all they could do is sum a column. They couldn't really build a good Excel spreadsheet. You give them an AI tool, and they go from a 2 to a 7 at being able to build a spreadsheet, just talking to Copilot. Just talking to it and saying, here's what I need, or here's a sample of an old one, I need to improve this one.

And so I love that ability for somebody whose skills are not elite skills, or even very good skills, but they want to work hard and they want to learn, being able to partner with an AI tool and take a giant step forward. And I think we're seeing that in the workforce now. We're seeing people self-learn how to build agents or automations. We're seeing people self-learn how to be citizen developers, even if they're not asked to. People watching YouTube videos and figuring it out, where maybe they were an average team member, and then as a citizen developer they have a specialty that nobody else has. And I think they are proud of themselves when they do that. They bring huge value to the organization when they do that. But leaders have to be intentional about finding these people and rewarding them, or inspiring people to do these kinds of things.

These are the stories that make me feel good about AI on days when the doomers are trying to make the narrative that we're all going to die. I'm like, okay, but did you see what Joanne taught herself to do? Who was 28 years old, not a distinguished employee, and this is a true story, right, at her place. She taught herself how to build automations, and now they call her the Agent Queen. And I asked her, how did you learn how? And she said, I taught myself with YouTube videos. I just think that's such a great story. She'll be paid more, she'll move up faster, she'll provide more value. It's a shame those aren't the stories that blow up on the internet.

Shane: It reminds me of another story that I've really liked, that I think is an excellent usage, because you're bringing up one of my favorite usages of AI, which is to train yourself. You don't feel stupid in front of the AI with whatever question you ask, and it has infinite patience for whatever you're going to ask.

One thing that I see in the industry a lot, and we were just alluding to it: if you have these long careers, you have planned retirements with training the new generation. And my experience back in the past with that is that it is always difficult. It is a year to two-year-long process that is planned out. This person's going to retire in two years, they're going to get trained for 18 months, and finally be able to take over this job. And the problem always in the past has been this person retiring, their job, they have done this job excellently, but they aren't necessarily a great teacher. I remember when I was transitioning, I had two weeks to teach people everything I'd done for a decade. I did a terrible job at it. And they asked me questions for about the next nine months, from elsewhere.

But now what we're seeing is this excellent usage of AI to bridge the gap, so that the person leaving can put their wisdom into the AI and have it transformed to also be given back to this new trainee in the way the trainee wants to learn, because the trainee doesn't necessarily want to learn in the way that this person who's a non-teacher and has never been trained wants to teach. I think that's a fantastic way to start doing it, and it also gives that great AI exposure to this new employee right away, to not be afraid to start automating and seeing these opportunities.

It's always one of my favorites. If you're not ready for full automations, and you have turnover that you're trying to do training for, it's a great way to then give that harness to your incoming employees, and they're going to do new stuff.

Scott: How do we do knowledge harvesting? How do we do knowledge harvesting from the older people?

Shane: Yes. The older people.

Scott: The older people, right, that have lots of experience and wisdom. The older people. How do we do knowledge harvesting?

Shane: There's a mirror over there, we can see ourselves.

Scott: Exactly. But older people. How do we do knowledge harvesting?

Shane: I love the concept of knowledge harvesting. I think today people think it's weird. They want it, but they think it's weird. I just don't think it'll be weird in the future.

Scott: Right, I don't think it'll be as weird. And I could understand a bit of when it's knowledge harvesting, and you feel like, oh, this knowledge is getting harvested out of me, and now it's going to go into the machine. But when the knowledge is getting harvested out of you and given to the rest of the team, to kind of tide lift everybody's boat here, that should feel powerful. Because for me, that's always been the power of being a leader, and that is being able to lift up by imparting as much as I can. And I think that is where, if you can share that more with your colleagues when you're doing knowledge harvesting, it's not to go learn all your workflows and then say goodbye to you, it's to raise everybody else up. Because having these knowledge taps is fantastic. And you're able to be a blessing to everyone else. And I think most people I know who are retiring, if you said, hey, we have a way that you can do an interview with an AI for an hour a day, it'll just talk to you, and it'll ask you a bunch of questions, you just have to talk to it, they would be fine doing that.

So, on knowledge, you made a comment earlier about when you work with an AI tool. For any of you who are listening, when you hear us talk about teaching yourself how to be a developer, or teaching yourself how to do automations, if that sounds daunting to you, there is such a great ability to just talk to an AI tool and ask it, teach me how to do this, teach me how to do this. And it doesn't judge you, and it'll slow down, and it'll do it as many times as you need it to do it.

And I think about, I was starting when I went back to developing with tools like Codex. And I didn't understand what looping was. And I kept hearing people talk about looping. I'm like, oh my gosh, people are talking about looping like everyone knows what it means, and I have no idea, really, what this means. And I remember asking Codex, hey, I feel like I should be telling you to loop something, or I should do looping, but I don't understand the concept of looping. Can you just explain it to me and how I might use it? And it was like, okay, here are different ways, you should probably ask me to do it this way. And so then I just did. I did what it asked, and then I saw the outcome. But I say that to, again, any of you who just heard this and went, looping, I have no idea what looping is. You are one AI request away from learning what looping is.

Shane: Yeah, totally.

Dave: I'm going to seize control, because I have a couple of other questions.

Scott: Fair enough.

Dave: What are the best investments? There's a wealth of opportunity here. There's a need to manage risk, address concerns. But what are the best investments that organizations can make right now in artificial intelligence?

Scott: First of all, my number one thing is just have an idea of what you're going to invest. I get concerned that when I ask people, as we're sitting here today, hey, what's your 2027 investment or budget, people have no idea. So I think first of all, it is what is the amount that you're going to invest, and are willing to put to work? And then down from there, it's okay, now what do we put that in?

And again, I'm going to go back to a model of saying a percentage of this investment should go into raising the skills of our workforce in general. That's both the licenses we need to have, the education we need to have, the bonuses we need to pay, the hackathons we need to do to get people excited. What's everything we should invest to get our team three steps forward next year?

And then the next layer, that automation layer we talked about. Okay, well, how much of the budget or the investment should we put into building use cases? They're automating things, and how do we build the right use cases? If there are hundreds that we could do, in 2027 how do we pick the right 75 we want to build that year? And so I think you've got to look at that automation layer, put some there.

And then I think the top one is, what do we want to invest in customer-facing AI tools? This is just the architecture I think in when it comes to budget and putting money in. Those are three areas that you should focus on putting money in. Raise all boats, knock out as many automations as we can, and then put some investment into customer-facing. And depending on how aggressive you are, depending on where you are in your journey, the percentage would be different. It might be 50% in my workforce, 30% in automations, 20% in my customer-facing AI tools, or a different recipe depending on where you are. So that's how I would do it.

Shane: The other thing, I think, in the investments, before we talk about any specific investment, is remembering these investments are bets right now. We're not in the fast follower stage, where we can look at somebody who's done everything right and we're just going to copycat them. Right now we have to take some bets on what is right. I actually really like your hackathon idea for citizen developers, but that's a bet that anything's going to come out of it. And so we have to also, when we're thinking through these investments, be comfortable with some of them not turning out, that some of these investments aren't going to work, but on the aggregate they are. So I think the other part of when we're looking to make investments is making enough investments that we can deal with some of them not paying off. If we're only going to make one investment next year, there's a high risk that one's not going to work out, and then I think one of the major failure modes is people have that one not work out, and then they just get turned off.

Scott: You've got to have a portfolio strategy.

Shane: Exactly. You've got to have a portfolio strategy here.

Dave: Makes sense. Another question: who are going to be the winners and the losers in the insurance industry itself with AI, and in the greater technology world? Predictions.

Shane: Predictions? If I'm going to get bold with a prediction of the winners and losers in the AI world, I don't know about losers, but I think the winners are more likely going to be us, companies that are actually utilizing the AI, more so than the labs. I'm of the opinion that the labs themselves are closer to commodities. They're just constantly competing and getting so much of the same capabilities that, to reference Scott Galloway, he always talks about them, they're just going to be like airlines. They work on razor-thin margins, but we all benefit from very cheap airfare and being able to conduct business and travel in new ways. I think we end up becoming the winners of AI much more than the labs there. That's my first prediction.

Scott: Interesting. I wouldn't have gone there, but I think that's very true. The people who verticalize, who support AI, who design it, build the tools. So first of all, I think you're probably right if you look on aggregate at how much money is generated and all that kind of thing. So I think that's an interesting take.

If I look at insurance companies in general, I think there are going to be some winners that are small to mid-sized insurance organizations that become elite with AI, and it is their lever to be able to compete up the chain. And it's a lever, it's an amount of power they haven't had before, and they'll move faster, they'll move better, and it will allow them to eat up market share, perform really well from a profit and a risk control standpoint. So I think we're going to see some smalls and mediums that become elite.

And the analogy I would give you would probably be e-commerce and retail. There were some retailers that were smaller retailers, not that big necessarily, that really did an outstanding job with e-commerce, and then kind of blew up on the internet. And then there were big retailers that just sucked at e-commerce. And I think we'll see the same dynamic. I think that happens a lot in technology waves. So I look for some winners to be small to mid-size.

Shane: And so I'm going to go to the loser side of that equation. I think there's some mid to large that are going to take steps back in the next three to five years. I think they're going to go slow with AI, I think they're going to be bureaucratic, they're not going to put the right investment in, they're going to be arrogant, thinking that they're good at it when they're not. And I think they're going to see their revenues kind of flatline, if not go backwards. And they're going to figure out that, hey, they were worried about insurtech. They thought insurtech could come take a chunk of their market. Maybe insurtech does in some ways, but I think what they're going to ignore is the small and mid-sized insurance organizations that become elite with AI and take some market share from them. So I think you have winners there, and I think you may have some losers up at the mid to large.

Scott: Yeah, that makes total sense. I've been implementing technology in the market for years. And when I look at any industry, and you look at the small, medium, large players, when there is a technology wave, one would think that the mid to large have lots of resources and they can win with technology. But how often do we see the mid to larger are slow to get things into the market, for many different reasons? Too bureaucratic with making their decisions, too risk-averse, and again, they think they're going to be fast followers. And I think we've just seen so many times the smaller side of the market leveraging technology faster, better. So that would be one of my answers with winners and losers.

I will say the other thing is creativity and the human connection are going to make winners out of some insurance organizations. You are building capabilities no one else has thought of yet. You are giving customers things nobody else is giving customers. You are doing things for your workforce nobody else is doing for the workforce. And again, in that world we're talking about of homogenized pricing, the two things that are going to separate you are your level of innovation, things that you're doing that escape the price homogenization or the service homogenization, and then the human connection.

And so I think the winners are going to be insurance organizations that kind of break the stodgy mold of an insurance organization, and they just become much better white space thinkers, and they're able to mix a highly automated AI-driven organization with a huge human connection out in the market. And some innovative aspects to what they do. That's my prescription, I guess, for what the winners will look like. Losers will be the opposite. They will just be the opposite.

Shane: And I think, too, right now with AI, your whole point of they're going to be doing things that nobody is doing right now, and probably nobody has thought about yet. And the only way you're going to get there is by embracing it early, so that you're giving your people on the team time to build the pattern recognition with these new tools so that they can get there. Because the reality is they're so new, we don't have the patterns to follow. And so this gives that true first mover advantage, which is a rare thing, where you are going to develop pattern matching that nobody else has had the opportunity to yet. But if you fall behind on it, you're just not going to catch it, and then somebody else is going to do that. So if you want to be that first person that really gets there, it's such a compounding benefit that you can't wait on it.

Scott: All right, so let's do something crazy. Instead of talking generalities, because you and I know this, we were talking about this before we started: let's talk about one thing that an insurance organization can do that's innovative and creative. Let's talk about how you would apply the council of experts concept.

Think about creating an AI council of experts that is customer-facing, that helps a customer decide what kind of risk control they need. I don't care what the insurance is, P&C, homeowners, health, anything. If you had a council of experts that the customers could talk to, so it's not just an agent, right, or a person, but you can talk to a council of experts that would be very good at asking you, what do you want from this protection? What is everything you really have to protect? And let us give you a lot of different opinions from different experts. That's the kind of tool that an insurance organization could build that I don't think anybody's got right now. So I know I'm just throwing that at you, but you understand council of experts. Isn't that an example of something an insurance organization could build?

Shane: It is. It's something an insurance organization can build, and I like it. I like the innovation of it, because when I think of being inside there, I immediately went to commercial lines. If I'm a business going to look for commercial lines, I'm probably going to an agent, and the agent's going to ask me some reasonable questions to try to get me across policies, but it's not going to do what a council of experts could do, which is take it from different angles. Let's look at our legal angle, let's look at our business strategy angle, all these other things that would be wildly amazing if my insurance company is giving me this ability to do it, because otherwise I've got to develop it myself, and most of our customers aren't going to think about that. But being able to do that is highly innovative.

And it lets us think, too. I know all of us have, usually, pretty deep relationships with our agents. And so since we have all these independent agents that we're trying to get to write business for us and with us, how do we give them the council to be able to give to our potential policyholders as well? That's something that is a leg up that right now no other carrier is giving to their agents. And as much as giving them golf outings and trips helps to write policies, giving them something that makes their job not just easier, but more dynamic, and gives more value, because agents live pretty much almost entirely off of their relationships to the people in the community. So being able to give them a way to make even deeper customer intimacy is going to drive that back to you as well.

Scott: Make them more valuable.

Shane: Exactly.

Dave: I have to make the comment that I know now how directors feel when they're dealing with very talented actors who go off-script.

Scott: I can't imagine why you would feel that way.

Dave: But this has been a fascinating discussion. I think we are at time. Sorry we can't just continue, because I have more questions, and I've found the discussion fascinating, and I'm sure our audience does as well. But anyway, I want to thank you both, Scott and Shane, for an engaging discussion. I'm not going to thank you for following the script, but that's a different issue.

And I'll say to the audience, if you want more information on this topic and more, please visit our respective websites: www.swept.ai and, correct me if I'm wrong, Scott, www.fpov.com.

Scott: That's perfect.

Dave: Okay, thanks for your time today, everyone, and we look forward, ideally, to doing something like this in the near future.

## What to do with this The adoption argument is over at most carriers, and many of them missed the moment it ended. Employees, vendors, and a platform that shipped a feature nobody asked for settled it between them. The question left over is smaller and more awkward, and worth answering before the 2027 budget locks: can you describe what the agents you already run are doing well enough to approve the next one? That is the problem Swept works on. We [govern the AI your team is already using](/offering/governance) and [establish how a system behaves](/offering/reliability) before and after it ships, so the answer you give an examiner, your board, or your own head of claims comes with a record behind it. If you are stuck somewhere around agent number three, [talk to us](/contact).