The case for AI in insurance claims does not need another abstract argument. Insurance leaders need to know whether it can fit into an operating claims environment with legacy systems, sensitive information, uneven data, limited IT capacity, experienced skeptics, and real coverage decisions on the line.
That is what made this Insurtech Insights USA 2026 panel useful. Moderator Sabine VanderLinden brought together Yo Sub Kwon, CEO of Voltaire; Greg Cornett, AVP of Claims at Harbor Claims; and Tiffine Wang, CEO and founder of Onsen Global. Their discussion stayed close to the work: how to start without waiting for a core-system overhaul, where to draw data boundaries, how to train new adjusters, and what changes when employees see AI as support rather than replacement.
The full 41-minute panel is below, followed by the most practical lessons for claims, operations, IT, and innovation leaders. Closed captions are available in the player.
Watch the Full Panel
Key Takeaways for Claims Leaders
- Start with operational proof, not an AI thesis. Greg’s confidence changed when he challenged the tool with a partial-denial letter and inspected the policy language it returned.
- Do not make a core-system integration the price of learning. A standalone workflow can establish value while the carrier’s IT team plans a deeper integration on a realistic schedule.
- Set data boundaries around the job being done. More information is not automatically better. Claims teams still need deliberate rules for relevance, access, security, retention, and human review.
- Pair technology with a talent strategy. Harbor Claims’ Claims College gives newly licensed adjusters structured practice before live claims. AI can support that learning, but it does not replace claims judgment.
- Adoption becomes real when employees ask for more use cases. The useful signal is not executive enthusiasm. It is the adjuster who moves from avoiding a tool to asking whether it can help with another letter.
- Use QA to correct work while it is still in motion. The panel’s strongest quality argument was not faster drafting. It was earlier detection of trends, recurring mistakes, and unsupported positions.
Moving AI in Insurance Claims From Hype to Operational Proof
Greg described himself as the person who translates between insurance language and computer language. He has worked with claims technology since the Windows 95 era, so he did not enter Harbor Claims’ first AI-vendor meeting ready to be impressed.
He entered ready to break the tool.
“I purposely went into our first ever meeting with an AI vendor looking to break it.”
Greg Cornett, AVP of Claims, Harbor Claims | 04:00
Instead of accepting the proposed payment-letter demonstration, Greg asked for a partial-denial letter. The point was not to reward a polished demo. It was to test a harder workflow where the draft had to reflect a mixed coverage position and cite relevant policy language. He said the result exceeded his expectations and surfaced language he did not expect the system to find.
That is a better starting point for an AI evaluation than a general promise of efficiency. Pick a consequential, bounded workflow. Give the tool a case that represents the real difficulty of the work. Then have experienced claims professionals inspect the inputs, policy citations, reasoning boundaries, draft quality, and review path.
For leaders evaluating AI claims correspondence, the unit of proof should be the actual claim workflow, not the stage demo.
Working Around Legacy Systems and IT Backlogs
The panel did not pretend that carriers can simply replace core claims systems. Insurance technology environments are layered, regulated, and full of dependencies. A useful AI product has to work with that reality.
Yo explained that Voltaire initially expected the final implementation to sit inside the carrier’s claims management system. Carriers wanted that too. Then the implementation conversation reached the IT queue.
“I think you have to meet the carrier where they’re at.”
Yo Sub Kwon, CEO, Voltaire | 10:50
Some carriers told his team that IT was backlogged for six months, a year, or even two years. The practical response was to let carriers start with a standalone workflow that could generate a letter without consuming scarce integration resources. A deeper integration could come later, once the carrier had capacity and enough operating experience to define what it actually needed.
This is not an argument against integration. It is an argument for sequencing. Claims leaders can separate two questions that often get tangled together:
- Can this workflow create safe, measurable value now?
- What integration is justified after the team understands the workflow, controls, and adoption pattern?
Starting outside the core system also creates new responsibilities. Leaders still need approved data-transfer methods, identity and access controls, retention rules, security review, auditability, and a clear system of record. “Standalone” should mean faster learning, not weaker governance.
Data Boundaries, Sensitive Information, and Coverage Decisions
Greg’s example of the data-boundary problem was deliberately ordinary. When an AI tool helps draft a claim letter, should the workflow receive the full field-adjuster report, or only the desk adjuster’s observations relevant to the communication?
The answer is not “always provide everything” or “always restrict the record.” The team has to define the purpose of the workflow, the information necessary for that purpose, and what a human reviewer must still consider. A field report may contain relevant facts, opinions, or material outside the scope of a particular letter. The control question is whether the selected inputs are appropriate for the task without excluding information that could change the claim analysis.
That distinction matters most around coverage. Yo described a workflow that returns a draft for review when it cannot find policy language supporting a proposed denial. Greg then stated Harbor Claims’ operating principle more directly.
“If you’re trying to find the policy language to support a denial and you can’t find that language, there’s a pretty good chance you shouldn’t be denying that.”
Greg Cornett, AVP of Claims, Harbor Claims | 22:42
The careful reading is important. Failure to find supporting language is a warning to reconsider the position, not an automated coverage determination. The adjuster remains responsible for the claim decision, the complete record, applicable law, and the final communication. AI can help surface policy language and identify a gap. It should not turn a partial record into a denial.
For teams reviewing this workflow, Voltaire’s public chapter on denial-letter history and practice provides useful context on why these communications deserve close review.
The Claims Talent Gap and Harbor Claims’ Claims College
The panel’s talent discussion was more grounded than the usual shortage narrative. Greg focused on what happens after a newly licensed adjuster arrives.
A license does not provide practical experience. New adjusters face a large body of policy, process, system, and communication knowledge at once. Harbor Claims responded by building Claims College, a structured series of courses that introduces the work step by step before participants move onto live claims.
Greg acknowledged the tradeoff. The program takes longer before a new employee handles live files. Once they do, he said they are more efficient from the start and less dependent on constant over-the-cubicle questions.
AI fits beside that program, not in place of it. A drafting or review tool can reinforce consistent structure, surface policy language, and give a new adjuster a clearer starting point. The organization still needs experienced instructors, supervised practice, escalation paths, and accountability for decisions. The strongest talent model combines domain expertise with technical fluency, which was also Tiffine’s recommendation for internal AI teams.
Top-Down Support Is Necessary, but Resistance Is Managed in the Middle
Greg was unequivocal that top-down support mattered at Harbor Claims. Its ownership group backed AI adoption from the beginning, which gave teams permission to learn and kept the effort from disappearing during the first difficult implementation decision.
Leadership support, however, does not resolve the employee-level question: What happens to my role?
Tiffine described the “messy middle” of adoption. A manager may own a workflow that AI changes substantially. If the organization waits for that person to feel threatened, resistance is predictable. Her advice was to involve those managers early, define the work they will own next, and give credible internal champions enough authority to move an initiative through the company.
She also offered a sharp example of the cost of delay.
“I finally went and did some hackathons over the weekend and basically built what I asked them to build in five hours. I’d been asking for three months. This took me literally five hours.”
Tiffine Wang, CEO and founder, Onsen Global | 32:59, lightly edited for clarity
The point was not that every enterprise problem can be solved in five hours. It was that direct exposure changes the adoption conversation. Teams that regularly use the tools develop a more realistic sense of what is easy, what is immature, and where governance or integration genuinely needs more work.
AI as Support for Adjusters, Not a Replacement
When the audience asked what claims representatives were saying about AI-written letters, Greg’s answer was plain.
“They understand that this is not here to replace them. This is here to support them.”
Greg Cornett, AVP of Claims, Harbor Claims | 30:01
The employees who initially hesitated did not become advocates because of a transformation memo. They changed when the tool handled part of the work they found tedious, while leaving them responsible for the claim. Greg’s favorite adoption moment comes when the question changes from “How can I avoid using this?” to “Could you add this letter too?”
That is a stronger adoption measure than logins alone. It shows that claims professionals understand the boundary of the tool and see a practical benefit. It also aligns with a lesson from Voltaire’s earlier discussion of AI that adjusters actually adopt: workflow fit and confidence matter as much as technical capability.
Leaders should still be precise about the word “support.” Support means the adjuster can review the source material, challenge the draft, change the decision, document the rationale, and decline the output. If those controls are missing, the organization has automation without accountability.
Claims QA, Trend Detection, and Mid-Flow Correction
Greg called claims QA “low-hanging fruit” because much of it remains manual. His proposed starting question was simple: present the claim file and ask, “Could we have done this better?”
The more interesting opportunity appears when review moves closer to the work. Traditional sampling can reveal a problem months after it began. A well-controlled AI-assisted QA process can help organize findings sooner, identify recurring language or workflow patterns, and direct human reviewers to files that need attention.
“AI might find a trend that a human might not recognize. So it’s giving us the opportunity to put a stop to bad habits or to implement corrections mid-flow instead of having to wait for that three-month delayed review.”
Greg Cornett, AVP of Claims, Harbor Claims | 35:31
This does not eliminate QA judgment. It changes the review cadence. Claims leaders still need validated criteria, representative samples, false-positive review, documented remediation, and a way to distinguish a drafting issue from an underlying coverage, training, or process issue.
The operational payoff is not merely a faster letter. It is a shorter distance between a problem entering the workflow and the organization correcting it.
The Monday-Morning Test
The panel closed with one sentence from each speaker. Yo emphasized compounding gains. Tiffine encouraged leaders to test tools while experimentation is inexpensive. Greg offered the most balanced summary: AI is a short-term investment pain point and a long-term benefit.
For a claims executive, the Monday-morning move is modest: choose one bounded, language-heavy workflow; define the data and decision boundaries; put experienced adjusters in the review loop; and measure quality, rework, cycle time, and employee response. Do not wait for a perfect enterprise architecture, but do not confuse speed with permission to skip controls.
For teams working specifically on claim-letter quality, the Claims Correspondence Compendium is a useful place to review letter types and jurisdictional requirements. If you want to compare that workflow with Voltaire’s drafting and review approach, request a focused conversation. The goal is not an AI project for its own sake. It is better claims work with the adjuster still in control.
Full Cleaned Transcript
Transcript note: This transcript was edited for clarity and accuracy. Filler, repeated words, and false starts were removed; adjacent fragments were joined; and obvious misspellings in names and company references were corrected. Meaning and useful timestamps were preserved. Statistics and company descriptions remain as spoken unless noted. The moderator’s opening project-success figures and workforce figures are included for completeness but are not repeated as verified facts in the editorial article because their exact source editions were not confirmed. Timestamps may vary by a second or two in the player.
Expand the full cleaned panel transcript
00:00–09:49 | Opening, Introductions, and the First Reality Check
00:00 | Sabine VanderLinden: We felt it would give us the best access to the strongest possible group of senior insurance decision-makers. Insurtech Insights is a must-do conference, I think, for the network that is built here. So welcome to the Orange Stage.
00:18 | Sabine VanderLinden: Let’s do a quick, honest check before we start. We know our industry has been suffering from a lot of hype and promises. We’ve seen the statistics. MIT highlighted that, I believe, 5% of AI projects technically succeed. [The next organization name is unclear in the audio] puts that at 20% within the industry. A recent report [publisher unclear] highlighted that 20% of projects actually go to completion within insurance. So there are a lot of transformation projects, a lot of AI, and the results are mixed.
01:00 | Sabine VanderLinden: I’m sure you’re a bit fatigued with vendor hype. We want to create a little reality check here. I’m sure I have the right audience and panelists to make sure we talk about AI transformation the way we should.
01:20 | Sabine VanderLinden: We’re here to talk about the AI reality check, overcoming legacy tech, the talent gap, and organizational resistance. To get started, I will ask the panel to introduce themselves in the context of our discussion. There is a Slido, so please send us questions a little later. We have a nice 40 minutes together. With no further ado, let’s go through the introductions. Let’s start with Yo.
01:52 | Yo Sub Kwon: Thank you, Sabine. My name is Yo Sub Kwon. I’m the CEO of Voltaire. Voltaire is an insurtech startup that helps insurance carriers write their claims correspondence and, in effect, helps them close claims faster and more accurately.
02:13 | Sabine VanderLinden: We have some numbers here already. Thank you. Greg?
02:18 | Greg Cornett: Hello, I’m Greg Cornett. I’m AVP of Claims for Harbor Claims, based in Florida but with operations across the United States. I’m in charge of claims and also vendor relations.
02:33 | Sabine VanderLinden: Thank you. And Tiffine?
02:34 | Tiffine Wang: Hi, I’m Tiffine Wang. I’m CEO and founder of Onsen Global. I typically work with large conglomerates on capital deployment and AI implementation. I was formerly with MS&AD, working on their investment team, so I have spent quite a bit of time in insurance as well.
02:53 | Sabine VanderLinden: I’ll throw my first question to Greg. You’ve been in claims automation since the Windows 95 era, right? You have now deployed AI inside Harbor’s claims operation, and you and your team are proud to go on the record publicly about having exceeded expectations. You work as a translator between insurance language and computer language. Take us back to before you flipped the switch to AI. What did you expect to happen, and what actually happened when you looked at the work you were doing at Harbor Claims?
03:37 | Greg Cornett: To expand on what Sabine said, I joke that I’ve been in technology and insurance so long that, when I started, all our machines were operating on Windows 95. That’s how long I’ve been doing this.
03:51 | Greg Cornett: What did I expect when we began to implement AI? I expected less than what I actually got. I purposely went into our first ever meeting with an AI vendor looking to break it. What happened very much exceeded my expectations. I was a little blown away.
04:13 | Greg Cornett: I think I was given, “Hey, let’s write a payment letter.” Instead, I threw out, “Hey, let’s write a partial-denial letter. Can you put that in there?” And it did. It was fantastic. It pulled policy language that I wasn’t expecting it to pull. The results were, frankly, a little astounding. It was very eye-opening that we were going to need to adopt this soon. That was my beginning experience with AI.
04:43 | Sabine VanderLinden: Let’s go into legacy tech. I want to start with the structural problem. Insurance is highly regulated, with many systems, a lot of complexity, and financial infrastructure. The stakes are often too high to rip and replace when things are broken. A lot of vendors come to you expecting pristine data and a multi-year IT overhaul. Let’s talk about what is broken and what works now.
05:19 | Sabine VanderLinden: Greg, when you started, there were a lot of manual processes before bringing AI into the organization. You worked on APIs, data management, and communication. When you look at what you’re doing at Harbor Claims, how much information do carriers and TPAs share with their vendor partners? Where do you draw the line around how and what you share?
05:50 | Greg Cornett: That is definitely a large part of the challenge because we’re dealing with sensitive information every day. How do we draw the line between giving these systems more of our data so they can do a better job of interpreting, reading, and finding things we might have missed, versus restricting and locking them down to specific sets of information and feeding them what we want to temper the results? That’s always part of the conversation, and it’s an ongoing conversation. Finding the balance between the two has been a constant challenge.
06:35 | Sabine VanderLinden: Can you share an example of that AI reality, what actually works?
06:42 | Greg Cornett: Take something as simple as writing a claim letter. Do we upload the entire field-adjuster report, or do we limit it to the desk adjuster’s observations of damage? A field-adjuster report might contain an opinion, or something extraneous to what we’re actually addressing when it comes to coverage.
07:10 | Sabine VanderLinden: I’ll go to Tiffine now. During our preparation call, you shared an interesting number. Last week I was in Barcelona chairing another conference, and that number was repeated. You said an AI-native company can deploy in weeks, while incumbents can take six months or more. That is a massive gap. Based on your experience as an investor, what should we expect in this new world?
07:45 | Tiffine Wang: To share some context, I invest out of a couple of different buckets. I have an earlier-stage fund that’s mainly AI-native. It’s tied to Stanford and has very engineering-heavy teams. Then I work with organizations writing larger checks, all the way to $40 million. Some are larger conglomerates, and some of the industries are more regulated, so it’s much harder to implement an AI solution.
08:15 | Tiffine Wang: A key difference is having the right DNA on the team. With large organizations, it sometimes takes much longer because you have many different people making decisions. I’ve worked with corporates and in the corporate-venture world for a long time. Many of the more innovative corporates, whether inside or outside insurance, have what I would consider innovative legal counsel.
08:46 | Tiffine Wang: If you have a legal team that’s always saying no and blocking everything because that is some version of the safest way to mitigate risk, you’re not implementing anything. You want someone on the legal team who works with you to support adoption and mitigate the risk of adopting AI so you can move faster.
09:12 | Tiffine Wang: The other thing is having the right technical talent in-house. The insurance industry tends to outsource a lot of it, though many organizations are trying to bring more engineering in-house. I work in Silicon Valley, where the majority of the teams are engineers. There is no discussion about whether to adopt AI. People are using AI all the time. It is natural. The question is, “What are we building?” They are doing optimization here and there, but it is less about how to implement AI than about what they are building, with AI used as second nature.
09:49–17:25 | Legacy Systems, IT Queues, and the Talent Gap
09:49 | Sabine VanderLinden: Yo, coming to you, Voltaire sits on top of carriers’ systems. One question I hear in the industry is whether, because of the AI-native revolution, we modernize and put things on top of legacy systems or rip and replace. In your case, you sit on top of the carrier’s existing system. Once you adequately explain what Voltaire can do, the carrier wants to implement the product. With the bureaucracy that exists in large enterprises, echoing what Tiffine said, talk us through how Voltaire is designed to be easy to implement in a legacy environment and how you manage the bureaucracy.
10:50 | Yo Sub Kwon: I think you have to meet the carrier where they’re at. A lot of times, when you’re trying to integrate something into a carrier, the IT team becomes a blocker.
11:12 | Yo Sub Kwon: When we first got started, we would show our product, and carriers would say, “Okay, that sounds great. Let’s move forward.” We always envisioned, and carriers told us, that it had to be inside their claims management system. We believed that would be the final, full implementation when a carrier was using our software.
11:37 | Yo Sub Kwon: When it came time to move forward, they would say, “Our IT team is backlogged for six months, a year, or two years.” That’s not ideal for a startup, to wait that long for a carrier to begin using the product. The carriers also wanted to start using it.
11:55 | Yo Sub Kwon: We had built a full working demo that we brought to shows like this, where carriers could see us immediately generate a letter. It created that magic moment when they saw the letter being generated. That demo ended up becoming the primary way we onboarded carriers.
12:19 | Yo Sub Kwon: It became a standalone tool outside the claims management system that carriers could implement immediately without using IT resources. Over time, we could schedule the full integration with their IT team when capacity opened, if they even wanted it at that point.
12:40 | Sabine VanderLinden: That’s perfect. We’ll move now to the talent gap.
12:43 | Sabine VanderLinden: I want to set out some statistics that I’ve been reading from the U.S. labor website. Demand for claims adjusters is projected to go down by 5%, underwriters by 3%, and actuaries to increase by 22%. When you look at the current gap, we will need approximately 18,000 claims adjusters per year for the next few years, up to 2030, but they are not there. We will need approximately 2,300 actuaries and around 8,000 underwriters, and they are not there. [Figures are presented as spoken; the exact source edition was not confirmed.] Five people are leaving and one person is joining the industry, so we have a major talent gap.
13:37 | Sabine VanderLinden: Greg, Harbor’s answer was to build Claims College. Tell us more about how you train people, get them to want to come into the industry, and build a program that works and inspires talent to join.
14:07 | Greg Cornett: We realized quickly that part of the problem with onboarding new adjusters is the overwhelming amount of information thrown at them. Take somebody who has newly acquired a license. They’re licensed, but they have no practical experience with what they’re supposed to do.
14:25 | Greg Cornett: We started Claims College, where we run them through a series of courses, teach them step by step, and slowly integrate them into the process. It has been very successful. Does it take longer to get them onto live claims? Yes. But once we do, they’re more efficient from the start. There is much less standing up over the cubicle and asking, “Hey, how do I do this?” It has been a very successful investment.
15:02 | Sabine VanderLinden: Wonderful. I’m checking whether we want to use the Slido. We had questions for you as well to test some of the conversation we’re having. Please answer the question if it is working, and then we can debate it.
15:27 | Sabine VanderLinden: Coming back to Tiffine, you’ve watched AI-native teams operate. It was fascinating to learn how differently they operate. You shared that the gap isn’t usually budget; it is that the tools don’t fit the people. What kind of people should an insurance company or large enterprise hire today to deploy AI well?
16:02 | Tiffine Wang: I’ve worked with enough large organizations to see that they tend to have subset teams. Sometimes what works is carving out an entity. They need to make sure enough is invested, that the team is empowered, and that it can tweak and build something new.
16:24 | Tiffine Wang: Sometimes initiatives change a lot and might get abandoned. Whatever you’re building should be integrated closely enough that people care about its success, while giving the team enough room to do what it needs without getting bogged down by the corporate world. Hiring the right type of talent and empowering it inside the organization is very important.
16:58 | Tiffine Wang: From a startup perspective, I usually see success when there is very good AI-native talent and very good industry expertise. Bringing in someone who understands how to navigate the regulatory aspects of insurance and pairing that person with very good AI-native talent can be a win-win.
17:25 | Sabine VanderLinden: If you have some time, go into Slido and answer the questions. The first is: What is the biggest blocker to deploying AI on your legacy system today? The second is: What is the most urgent talent investment for your firm in 2026? The last is: What is one thing you will do differently after this panel? We’ll look at the results toward the end and compare what you think with what the panel thinks.
18:20–25:13 | Human-Centered Tools, Coverage Support, and Organizational Resistance
18:20 | Sabine VanderLinden: Yo, Voltaire is AI software, but on our preparation call you said you take a human approach. That is important because an insurance claim is a highly emotional experience. It’s fine to bolt on or build great AI systems, but the technology works only if the human understands how to use it. Your customer-success team gets on a short coaching call with customer adjusters. Does the right AI tool make the job more interesting for claims adjusters?
19:13 | Yo Sub Kwon: Absolutely. We hear time and again that adjusters using Voltaire are significantly happier with their jobs. We’ve talked with thousands of adjusters, and we’ve never had someone say, “I’m so disappointed Voltaire will write these letters for us,” including letters that cite complex policy language and need regulatory compliance.
19:46 | Yo Sub Kwon: Besides the time savings, another thing we’ve heard is that adjusters are learning from the software. We have a feature where, if someone is trying to deny some aspect of a claim and there is no relevant policy language supporting that denial, the software won’t let them. It gives them that message and kicks it back.
20:16 | Yo Sub Kwon: Conversely, if they are denying something and relevant language exists, we pull the supporting language. We’ve been told, “That is really good supporting language. I never would have thought to pull that from the policy and add it.” It makes the adjusters better and produces higher-quality results.
20:40 | Yo Sub Kwon: One thing I’ve heard from Greg is that there are fantastic adjusters who care about the insured and the claims they’re working on, but might not be great at writing letters. If the software makes that part of the job easier, it is less stressful and gives them more confidence as adjusters.
21:06 | Sabine VanderLinden: We’re going to organizational resistance. We have to admit that claims adjusters and teams within carriers and TPAs have a right to be frustrated. Middle managers have a right to be frustrated. They often show skepticism toward new technology after 15 years of having technology thrown at them that didn’t work.
21:33 | Sabine VanderLinden: Greg, that resistance exists, but it is less of a struggle at Harbor because AI integration is a top-down priority embedded in your strategy. What did Harbor get right or do differently that we could borrow? What warning signs should leaders pay attention to so they manage that resistance?
22:10 | Greg Cornett: Top-down support is critical. It is absolutely the most important thing. You have to have ownership and leadership that believe in this process. I think everyone here agrees that, if we’re not on board with this, we’re going to get left behind. It is as simple as that.
22:30 | Greg Cornett: I want to go back to something you started to say. It is a small change of topic, and I apologize. You were talking about looking for the right language to support a denial.
22:42 | Greg Cornett: We have a saying at our company that, if you’re trying to find the policy language to support a denial and you can’t find that language, there’s a pretty good chance you shouldn’t be denying that. The support we get from the letter-writing side can help somebody reconsider coverage. If someone thinks they’re looking for something and can’t find it, they need to reconsider their position.
23:16 | Greg Cornett: I’m sorry, I strayed off topic.
23:18 | Sabine VanderLinden: No, that is superb. That gets me to a conversation I had with Tiffine. Looking at this from an investor viewpoint and observing patterns across carriers, the hardest resistance pattern you see is management, the “messy middle.” These professionals may have done the job brilliantly for 20 years and watched AI tools being thrown at them. We hear that Claude can do everything today. What is the playbook to address resistance at the human level, not only the organizational level?
24:07 | Tiffine Wang: Typically, I see very strong support for AI adoption at the top, and usually younger people enjoy playing with it. Sometimes the middle is different because those people have different responsibilities and have been in the organization long enough. That’s a general statement, because I also see strong adoption in the middle.
24:28 | Tiffine Wang: Getting a couple of champions to help push solutions through is very helpful. Sometimes, when you reach the middle layer, a person is in charge of a workflow that is their whole job. When AI comes and completely replaces the workflow, they think, “What am I going to do?”
24:48 | Tiffine Wang: Instead of letting them reach that “oh no” moment, help them figure out the next steps earlier. Say, “You’re now going to be in charge of all of this. You’re going to be in charge of a group of agents instead of a group of humans doing this.” You’re reframing their role. There aren’t enough people coming into insurance anyway, so help them understand that it can make the job easier rather than completely replacing them.
25:21–29:24 | Decision-Making, Procurement, and Signals of Real Commitment
25:21 | Sabine VanderLinden: We have self-appointed chief AI officers and decision-making committees that can stop AI projects from going into production, plus procurement processes designed for a slower world. What warning signs do you see when you enter an organization and determine that the decision-making process is broken? What should founders selling solutions be aware of, and what should they cut out?
26:15 | Yo Sub Kwon: Every carrier is different, and every company selling a product is different. What I’ve experienced is a shifting landscape. When we started a few years ago, we were essentially creating a new product category. No one was looking for our product, so we had to do a lot of education and explain a lot of things.
26:46 | Yo Sub Kwon: That process has gradually changed. At an event like this, more conversations start with an understanding of AI. Carriers are actively looking for AI tools and adding them to their organizations if they haven’t already. They may already know who we are and what we do, which makes things easier.
27:15 | Yo Sub Kwon: Early on, we didn’t have the luxury of being too choosy in how we approached carriers. Now, if someone doesn’t understand the problem we’re solving, it isn’t worth trying to convince them they have a problem inside their organization. It is up to them to find out, and we’ll be available when they’re ready to discuss it.
27:44 | Yo Sub Kwon: Many people are already enthusiastic. Every demo we give, people are excited about the product and what it could bring to their organization. Ultimately, the biggest signals are the actions they take.
28:04 | Yo Sub Kwon: When you’re working with an organization where people respond to messages, and they do what they say they will do, that is the enthusiasm you want to see. If they don’t respond unless you follow up four times, or don’t respond at all, that isn’t where you should spend your energy and time. Spend it with people who sound enthusiastic and show it with their actions, who are trying to move it forward inside the organization.
28:44 | Yo Sub Kwon: That said, there are huge insurance entities with a lot of bureaucracy. You go through vendor procurement, an AI-team review, a legal-team review, and an IT-team review. Sometimes you still have to go through all of that. We see organizations move at drastically different speeds. The biggest differentiator in whether something will close quickly is, basically, whether someone responds to messages quickly.
29:24 | Sabine VanderLinden: It looks like everybody is shy. We don’t have any Slido answers, so I’ll move to the audience questions. Mitchell, tell me if anything comes through.
29:40–33:49 | Adjuster Adoption and the Employee Light-Bulb Moment
29:40 | Sabine VanderLinden: The first audience question is: What are claims representatives saying when AI can write letters, and do you think they felt they previously owned that work? That is for you, Greg. How does it affect them?
29:56 | Greg Cornett: I think they love it at this point. The idea is already out there. They understand this is not here to replace them. This is here to support them. If it makes their job easier, they’re all for it. It is as simple as that. It is becoming a more useful and accepted tool every day.
30:16 | Sabine VanderLinden: Yo, you’re working with Greg and many other organizations, alongside claims adjusters. What are they telling you when you’re augmenting their jobs?
30:35 | Yo Sub Kwon: Initially, there are always some holdouts. Some are really hesitant. They say, “I’ve been doing this for 20 years. I know exactly what I’m doing.” But once you put the tool in their hands, walk them through it, and show them, there is genuine excitement.
30:58 | Yo Sub Kwon: Humans aren’t optimized to read 100-page documents and then produce highly specific documents that need to be consistent every time. People have moods and bad days. There is inconsistency in humans, and AI can outperform a human in this type of language task.
31:27 | Yo Sub Kwon: That produces a good outcome for adjusters and for the people who review the letters and everything that comes from them. There is less stress and less back-and-forth. I think that is a huge pain point that has been alleviated.
31:52 | Sabine VanderLinden: Greg, what you’re saying is that an expert with AI will outperform someone who is not an expert. An experienced claims professional can apply 20 years of expertise alongside the system, which is better than a non-expert trying to work with claims.
32:22 | Greg Cornett: One of my favorite moments is when the staff begins to adopt these processes. Instead of, “How can I avoid using this?” it becomes, “Could you add this letter too? Can you do this for me as well?” They expand that thought process and ask, “How else could we use this?” It is always fun to see that light-bulb moment happen.
32:45 | Sabine VanderLinden: Tiffine, what about your environment at enterprise scale?
32:52 | Tiffine Wang: I agree. I remember trying to coach one team to use a solution. I finally went and did hackathons over the weekend and built what I had asked them to build in five hours. Suddenly they were all trying to adopt it the next week. I said, “I’ve been asking you for three months. This took me literally five hours.”
33:16 | Tiffine Wang: When people spend more time around others who are actually using and enabling the tools, it makes them want to adopt them. If you’re around people who always say it is too complicated or the solution needs to mature, you sit on the sideline, wait, and complain. If you’re using it day to day, you realize how much more it can do. These AI tools are evolving very quickly, so they’re getting easier to use.
33:49 | Sabine VanderLinden: I received a question asking whether the poll was meant to be active. It was supposed to be active, but I’m not the technical person, so maybe it wasn’t activated. I’m sorry. We are answering your questions now.
34:04–37:27 | Compounding Gains, Faster Feedback, and New Problems to Solve
34:04 | Sabine VanderLinden: What is the biggest benefit you can see if AI is fully adopted by carriers, at repeatable scale? I’ll start with Yo, then go to Greg.
34:16 | Yo Sub Kwon: Sorry, what is the biggest benefit?
34:18 | Sabine VanderLinden: What is the biggest benefit you can see if AI is fully adopted by carriers, hopefully in a repeatable, scalable way?
34:31 | Yo Sub Kwon: It is an incredibly powerful technology, and it is advancing so quickly. If you’re sitting on the sidelines and not implementing some kind of AI technology, especially in the places where there is so much low-hanging fruit and AI can bring immediate benefit, you are falling behind.
35:05 | Yo Sub Kwon: If you implement it now, you aren’t necessarily trying to reach where the earlier adopters are today because the gains are compounding. You’re chasing a moving target. If you implement now, those who implement after you will have a much harder time catching up.
35:24 | Sabine VanderLinden: Greg, what is the biggest benefit in your environment? Do you have any numbers to share?
35:31 | Greg Cornett: For me, it is about identifying trends, the speed with which we can write a letter and get it reviewed. In addition to letters, AI is giving us the ability to analyze this information much faster. AI might find a trend a human might not recognize. It gives us the opportunity to stop bad habits or implement corrections mid-flow instead of waiting for a three-month delayed review. That has been one of the things that jumped out for me.
36:05 | Sabine VanderLinden: So it is about identifying daily signals. I do that too. I have scheduled daily signals, insurtech radars, and top topics for carriers. I rate and score them, then focus on the most important. That is an example of how it can be used.
36:26 | Sabine VanderLinden: Tiffine, as an investor, what top benefit do you expect? How do you make an investment choice when you think about AI-native teams?
36:35 | Tiffine Wang: I invest in many AI-native teams, so that is basic bread and butter. The fun part of this job is that there will always be more interesting problems to solve. Once you automate a group of boring tasks, I guarantee new risks will emerge that insurance has to jump on and figure out how to solve. That is part of the innovation and investment cycle.
37:03 | Sabine VanderLinden: Greg, how can you effectively manage change if most of the board members oppose AI?
37:15 | Greg Cornett: I can’t answer that. I’ve never had to deal with it. I’m very fortunate to work for a company that pushed from the beginning. I have a very forward-thinking ownership group, so I can’t say I’ve experienced that.
37:30–41:33 | Holdouts, Claims QA, and Monday-Morning Advice
37:30 | Sabine VanderLinden: That is a blessing. Yo, you probably meet the adopters and transformers, those who want to become frontier organizations. But you may encounter people who say no.
37:51 | Yo Sub Kwon: I don’t know if I encounter them directly. There are bureaucratic blockers we run into, but we often don’t interface with them. They happen internally. We’ll introduce what we have, everyone is excited, and it sounds like they want to move forward. Then we get an email saying they couldn’t get the budget or something happened. I don’t have full insight into those situations.
38:25 | Yo Sub Kwon: Often we move on. Sometimes they come back later. There are holdouts who resist change or adopting AI technology. Sometimes we talk to people who don’t fully understand what AI could do for them. They write it off as a fad or say they’ll get it later.
38:57 | Sabine VanderLinden: Greg, what is one example of low-hanging fruit a carrier can use to upskill personnel quickly today with AI?
39:15 | Greg Cornett: I think QA review is low-hanging fruit. In the past, it has been a very manual process. I think it can be much more fully ingested and automated. Verification is so simple. You can present the claim file and ask, “Could we have done this better?” It is amazing what some of the results will be.
39:44 | Sabine VanderLinden: There is a question about what is hardest and where we should start, legacy tech or culture. Remember, AI is a business transformation and a cultural transformation. Most insurers have legacy systems and a lot of systems from M&A. To do it right, the data foundation will be critical. Most of all, there must be alignment between the organization and the culture. It is a business transformation.
40:15 | Sabine VanderLinden: I will ask my last question. Thank you so much for joining us. One sentence to the room, with no vendor language or corporate speak. What is the one thing you want this audience to take away from the Orange Stage? What should they do on Monday morning to get started? Yo?
40:40 | Yo Sub Kwon: AI provides compounding gains. That is the message I’d like to share. If you aren’t already using it, you absolutely should.
40:52 | Sabine VanderLinden: Greg?
40:53 | Greg Cornett: Everybody needs to remember that AI is a short-term investment pain point, but a long-term benefit.
41:02 | Tiffine Wang: I encourage you to test as many new AI tools as you can right now. They are venture-backed and subsidized, so go get your free credits.
41:13 | Sabine VanderLinden: Get the free credits before you have to pay a lot for them. Let’s give a big round of applause to the panel. I wish you a great rest of the afternoon. Enjoy the drinks and networking, and hopefully we’ll see you soon. Thank you.
41:33 | Greg Cornett: Thank you.
