webinar
Iryna Kravchenko Iryna KravchenkoChief Editor
Business·Insurance·

Webinar recap: How AI is changing underwriting

DICEUS hosted an online webinar “How AI Is Changing Underwriting: Opportunities, Limits, and Real Use Cases.”

Our experts Kateryna Monastyrska, Alina Demchenko, and Diana Mircheva spoke about how insurers could apply artificial intelligence in underwriting operations without losing human control.

Missed the session or want to revisit the discussion?

Watch the full webinar recording here:

How AI Is Changing Insurance Underwriting | DICEUS Webinar

The discussion focused on three key topics: where AI can support underwriting teams, which risks and limitations should be considered, and how organizations can shift from AI experimentation to practical implementation.

A working example of AI in underwriting is realized in DICEUS AI-Powered Underwriting Workbench. It includes an AI Advisor – a scoring engine that helps underwriters assess submissions in three steps:

AI doesn’t replace the underwriter. It speeds up triage and highlights risks so the human expert can focus on decisions rather than manual data tasks. To tailor the model to a specific company’s risk profile, we typically train it on 3,000–5,000 historical data entries from that company’s own portfolio.

If you want to see it in action, feel free to book a demo.

Below is a transcript of the webinar.

Kateryna: Welcome, everyone. Thank you for joining us today. We are going to have an honest conversation about AI in underwriting: what’s actually working, where the limits are, and what teams should be paying attention to right now. My name is Kateryna Monastyrska. I lead insurance technology solutions at DICEUS, and I spend most of my time working with insurers and MGAs on exactly these questions. I’m joined by two colleagues who approach this from different but very complementary angles. Let me hand it over for quick introductions. Diana, over to you, please.

Diana: Hi, thank you, Kate. I’m Diana, Product Owner of the Underwriting Workbench. I’m responsible for how the product actually evolves: what gets built, in what order, and why. That means I’m constantly working through the gap between what AI can do technically and what makes sense inside real underwriting operations and business processes.

Kateryna: Thank you very much. Alina, please.

Alina: Hi, everyone. I’m Alina, Business Consultant at DICEUS. My role focuses on bridging the gap between what underwriting teams need day to day and how the platform is configured to deliver that, whether it’s submission intake, workflows, decision rules, or broker collaboration. I spend a lot of time talking to underwriters to understand where the current process breaks down, before we even start introducing any technology.

Kateryna: A perfect combination of perspectives. So let’s start with the big picture. Before we talk about what AI can do, I want to ground it. When we talk to underwriters today, a lot of them are skeptical. They have heard that AI will transform underwriting for five-plus years, and they still start their day reviewing emails and building spreadsheets. Alina, is that a fair picture?

Alina: Completely fair. There is a real gap between what gets announced at conferences dedicated to AI and digital transformation and what underwriters actually experience day to day. The reality is that most teams still spend about 40% of their time on tasks completely unrelated to risk assessment: chasing documents, entering and structuring incoming data, clarifying incomplete submissions, manually pulling data from public registries, and so on. That’s the baseline we are working against.

Kateryna: So hasn’t AI fixed that yet?

Alina: Not broadly. We definitely see some progress, and certain carriers are getting real value from optimizing specific steps in their workflow. But when it comes to wholesale transformation, it remains a challenge.

Diana: Yes, and the real bottleneck isn’t the AI model itself. It’s data quality, change management, and fragmented underwriting systems that prevent underwriters from feeding clean data into the model in the first place.

In general, that’s the part that often gets glossed over in demos. Models in demos are trained on clean, structured data. The real world sends you PDFs that someone scanned sideways, data that isn’t clearly visible, emails with lots of attachments in different formats, and spreadsheets that have been reformatted sixty times. Real life is a little bit different. So the first challenge is getting structured, usable inputs before the model does anything interesting.

Kateryna: So let’s talk about where AI is actually adding value today, given those constraints. Diana, what do you think?

Diana: I would say the first area where we see consistent, measurable results is submission triage. Not the decision itself, of course, but the sorting.

In real life, most people nowadays hand their boring, routine tasks to AI. No one goes to Google and checks every website to find an answer to their question anymore. Even Google itself just gives you an AI response. It’s a similar approach here. An underwriter gets dozens of submissions a day, which is a huge amount of routine. Of course, it depends on the company and the business.

But what AI can do here is read the incoming submissions, extract the key data, score them against your appetite, and give you a ranked list. That job is done by AI, and it reduces the time underwriters spend on routine. Which ones are worth your time? Which ones can you decline quickly? Which ones need more information before you can even assess them? This part works well, and teams that have implemented it are seeing triage time drop significantly, sometimes by 60 to 70% on this step alone.

Kateryna: Alina, do you agree on this one?

Alina: Yes, I completely agree. The core value here is focus. Skilled underwriters can handle complex risks, but if they are spending two hours a day just sorting through an inbox, that’s two hours lost for critical judgment calls. Triage automation gives that time back. Plus, from a practical standpoint, it prevents underwriters from missing out on profitable business simply because a good risk sat in the queue for too long.

Diana: Yes, that’s true. And the next layer is risk scoring. Not just sorting, but evaluating. You have a submission, and AI scores it against your underwriting guidelines, historical conversion data, and so on. Appetite match: does this fit what you write? Bindability, based on premium size, broker relationship, and response speed: how often do you bind this kind of risk? These signals exist in your data, and AI can surface them.

Kateryna: And this is different from the traditional rules engines that underwriters have worked with for years, right?

Diana: Well, yes and no. Rules engines are explicit. Of course, we use them, and I believe we will keep using them for a long time wherever we don’t need AI. So what is actually the difference? AI can find patterns in historical outcomes that no one explicitly programmed.

Rules are, let’s say, hard-coded. They are defined, they’re there, they work as usual, and that’s fine. But with AI, the insight might be: this type of account, in this region, submitted by this broker, at this time of year, has a 70% bind rate. That combination of factors would never appear in a hand-coded rule book, because it’s a different approach.

The flip side is that the model can also learn bad patterns if your historical data is biased.

Kateryna: Alina, what’s your opinion on that?

Alina: That’s a very valid concern. If your historical book of business leaned heavily toward specific geographies or industries for reasons completely unrelated to risk quality, like legacy broker relationships or historical appetite, then the AI model will carry those patterns forward. That’s exactly why we need underwriters to be actively involved in validating what the model is learning, rather than blindly accepting its output.

Diana: That’s totally true. And another area where AI is not just able to help, but is actually helping today, is quote generation. For standard, low-complexity risks, AI can take the structured submission data, apply your pricing models, and generate a quote automatically. The underwriter reviews and approves it rather than building it from scratch, which is what actually takes time. As we said, routine can be simplified with AI, and that’s where you get the 50% improvement in quote turnaround that you start to see in better implementations.

Alina: And I’d like to add that the keyword here is standard. When a risk calls for deep underwriting judgment, like a complex account, unusual coverage combinations, or an entirely new risk category, straight-through pricing automation is not the goal. Instead, technology should give underwriters a head start: it pulls the data, calculates the initial technical rate, and flags potential anomalies and areas that need attention. From there, the human expert should definitely take over and make the call.

Kateryna: Now let’s be direct about where AI doesn’t work well yet, where the limits are real and underwriters should know them going in. Diana, what can you tell us about the limits?

Diana: Yes, there are some limits. The first and most crucial one is that the model is only as good as the data you train it on. In underwriting, historical data is often incomplete, inconsistently structured, or covers a period that doesn’t reflect current market conditions. Say, two years ago. The world changes, sometimes even within a month, depending on the case. If your data is thin for certain segments, the model will be unreliable for those segments, and you might not even know it. Still, the model keeps learning, and that’s why we also need underwriter approval in some cases, so it learns from fresh data.

Alina: I would also like to add that AI doesn’t handle completely new and specific risk categories well. When a risk falls outside the historical training data, human judgment is irreplaceable. A good example that comes to mind is cyber insurance. The exposures and loss patterns change so constantly that historical data from just a few years ago won’t help you price this risk today. That’s why underwriters’ experience and market knowledge are essential.

Diana: Yes, definitely true. Another limit we have is explainability, and it’s still a real issue in production. You can build a model that is accurate, but the question “why did the model score this risk low?” needs an answer that an underwriter can engage with, not just a probability score. Regulators care about this, and so do underwriters, because if they can’t understand the reasoning, they can’t override it intelligently when they need to. So without explainability, AI kind of blinds underwriters. That’s one of the limits.

Alina: And there is another massive limit that is often underestimated: change management. AI tools launched without underwriter buy-in are bound to fail. I’ve seen it. The system flags something, the underwriters ignore it, and those alerts quickly become background noise. Sooner or later, trust is lost. To make AI work, you must involve underwriters from day one. The teams that make it work are the ones that involve underwriters in defining what the system should flag, and why, before the platform ever goes live.

Kateryna: So it’s not just a technology question, right?

Alina: Exactly. In reality, it’s maybe 30% technology and about 70% people, process, and data quality.

Diana: And the truth is that technology is usually the easiest part to handle.

Kateryna: Given all of that, the real wins and the real limits, what does a well-implemented, AI-assisted underwriting operation look like? What’s the pattern you see in teams that are making it work? Diana?

Diana: There are a few things they do consistently. First, they don’t try to automate everything at once. They pick one high-friction step, usually intake or triage, get it working and trusted, and then expand. The teams that try to do everything simultaneously rarely get anything into production, because it takes time and effort. And as Alina mentioned, there are people who also have to adopt AI and help it work.

Alina: Second, I believe underwriters are kept in the loop during the design phase. And it’s not just about keeping them informed. They need to be actively involved in defining what the scoring model should care about, setting thresholds, and deciding what triggers a human review. When underwriters help build the rules, they naturally trust the output of the AI model.

Diana: True. And to clarify: AI makes a recommendation, and the underwriter makes the decision. That’s the model that holds up over time. When you try to remove the underwriter from the decision entirely, you take on either regulatory risk or portfolio risk you didn’t intend to.

Alina: And from a practical standpoint, robust audit trails matter much more than people may initially think. Every single decision must be traceable: what data the model processed, the exact score it generated, how the underwriter acted on it, and, most importantly, why they deviated, if they did. This is not just to satisfy regulators. It’s for your own progress over time. If you can’t analyze why your decisions went the way they did, you can’t improve and optimize your process and get better.

Kateryna: While preparing for the webinar, we gathered three of the questions we hear most often from underwriters, and we’ll try to answer them now. If you also have questions, please feel free to send them to our emails, which you can see in the chat. We will get back to you with more details. So, Diana, could you please answer the first question: how much historical data does a carrier need before AI scoring is actually reliable?

Diana: Well, it depends on what you’re scoring and how granular you need to be. For broad appetite matching, a few thousand historical submissions can get you started. For segment-specific bindability scoring, you generally want at least a few years of outcome data: bind rates, losses, and renewal patterns. The honest answer is that many carriers don’t have it in a clean enough form to use directly. That’s why the first project is often data consolidation and structuring, before the model even starts.

Kateryna: Alina, the next question is for you. Should underwriters be worried about AI replacing their roles?

Alina: I’ll answer this directly: no. Not anytime soon, and probably not in the way people might think. The tasks getting squeezed are the purely administrative ones, like manual data entry, sorting documents, or basic intake. But the strategic side, like complex risk assessment, broker relationship management, and pricing for non-standard accounts, definitely remains irreplaceable. Ultimately, AI just frees up underwriters to spend more time on high-value work that truly demands their expertise.

Kateryna: And the last question is for you, Diana. What’s the single biggest mistake you see in AI underwriting implementations?

Diana: Oh, skipping the data audit, of course. Teams get excited about the model and start building before they even understand the state of their own data. Then they discover that their historical submissions are in, I don’t know, three different formats, half the outcome data is missing, and the policy records don’t link cleanly to the submissions. Those discoveries generally cost six months. So definitely do the data audit first.

Kateryna: Thank you very much, ladies. We are at time. Let me ask each of you for one thing you want the audience to walk away with. Alina, over to you, please.

Alina: Thank you. To close, I would like to emphasize that AI doesn’t replace human judgment. AI’s real power is to remove the day-to-day friction that holds underwriters back from applying their expertise where it matters most. That is the true value of optimization, and I should say it’s incredibly significant.

Diana: Yes, 100%. And I would say: start with your data. Not the model, not the vendor, but your data. If you know what you have and where the gaps are, every other conversation gets much easier.

Kateryna: Diana, Alina, thank you very much for sharing your expertise. A really practical conversation, exactly what we were going for. For anyone who wants to dig deeper and learn more about the DICEUS AI-powered Underwriting Workbench, we will share resources in the follow-up email. We would be happy to hear your feedback, so feel free to ask questions directly. Thank you very much, everyone, for joining us today. See you soon.

Diana: Thank you. Bye.

Alina: Bye.

Software solutions bringing business values

gartner
5/5
6 reviews
clutch
4.9/5
49 reviews

    Contact us

    100% data privacy guarantee

    Thank you!
    Your request has been sent
    We will get back to you as soon as possible

    USA (Headquarters)

    +19293091005 2810 N Church St, Ste 94987, Wilmington, Delaware 19802-4447

    Denmark

    +4566339213 Copenhagen, 2900 Hellerup,
    Tuborg Havnepark 7

    Poland

    +48573568229 ul. Księcia Witolda, nr 49, lok. 15,
    50-202 Wrocław

    Lithuania

    +37069198546 Vilnius, LT-09308, Konstitucijos ave.7
    6th floor

    Faroe Islands

    +298201515 Smærugøta 9A, FO-100 Tórshavn,
    Faroe Islands

    UAE

    +19293091005 Emarat Atrium, 423 Al Wasl Area, Dubai,
    P.O. Box 112344

    Jordan

    +962787500100 Swafeya, Waklat St. 2, Office 302,
    Amman