Yo Sub Kwon, Co-Founder & CEO of Voltaire
More Than Just Letters: Yo Sub Kwon on AI Claims Correspondence, Audit Trails and the ChatGPT Appeal
Every claim ends in letters: acknowledgments, payments, reservations of rights, denials. Someone has to write each one, cite the right policy language and get it out before a regulatory deadline. Yo Sub Kwon built a company around that one job.
Yo Sub is the Co-Founder and CEO of Voltaire, which sells software that drafts claims correspondence for carriers. He came to insurance from cybersecurity and fintech, and he says he started with letters and is expanding into more of the claims process.
In Episode 178 of InsurTechTalk, Yo Sub and I covered how an adjuster’s letter gets drafted and checked against the policy, what happens when an adjuster overrides the draft, why catastrophes made his first believers, and why he thinks the AI regulation conversation is missing a question.
About Yo Sub Kwon
Yo Sub says he was not originally in insurance. He was in cybersecurity and fintech, working on an AI project with his co-founder, and interviewing the users of that software to learn what they liked. One of those users was an adjuster using the software to write denial letters.
He had no idea what the person’s job was. The adjuster explained it, introduced him to their boss, who introduced him to their boss, and that person explained how claims and insurance worked. Yo Sub says he did not fully believe that writing this correspondence was a real industry problem, but he was willing to explore it. He built a prototype of what Voltaire is today and brought it to an insurance conference to show a group of executives. It was, he says, a Verisk conference, more than three years ago. He showed it on his phone, and he says the reaction was immediate: “we gotta get this right away.” That was when he thought there might be a product.
What Voltaire Does, and How Fast
I opened by saying that most of our listeners are in P&C, and that Voltaire takes one small niche of the big domain of claims: writing the letters for the adjusters. Yo Sub agreed that letters are where Voltaire started and the main focus of the product, and said the company has been expanding to help with more of the claims process.
88 Percent Go Out Unedited
I asked what sentence adjusters rewrite most often in the drafts. Yo Sub says 88 percent of the letters Voltaire produces go out to the insured with no edits at all. The edits he does see are custom to the claim: an adjuster who has talked to the insured on the phone and knows they would appreciate extra detail about one aspect of the claim, or a settlement letter that closes out a claim.
Where the Time Savings Are
Yo Sub’s point is that the more complex the letter, the more value the software provides. His two examples:
- Casualty: tender response letters, which can answer a demand letter, run 10 to 20 pages and are heavily legal. He says one can take an attorney six hours to two days to write. Voltaire finishes it in a couple of minutes, and the attorney spends about 30 minutes reviewing it.
- Home property: a denial or partial denial letter typically takes 30 to 60 minutes to write if the adjuster is not taking big shortcuts. Yo Sub says Voltaire’s average to finish one is 3.7 minutes.
Hurricanes Made the First Believers
Yo Sub says he learned quickly that an exaggerated reaction from an insurance executive does not mean an easy sale. A demo goes over well, especially when it runs on a carrier’s own letterhead and its own policies, and the sales cycle is still long.
Helene, Milton and the Independent Adjusters
After the prototype, he says, he went heads down, put a team together and built a production version. Where the company hit its stride was Hurricanes Helene and Milton in the Southeast. Carriers that had been talking to Voltaire suddenly became very interested, and Voltaire was able to onboard all of their independent adjusting firms and adjusters quickly.
What the carriers liked, he says, was consistent letters from adjusters who did not know the carrier’s policies, because they had been hired on short notice for a surge in claim volume. The industry tends to expect quality to drop in those periods, but regulatory deadlines still apply. Bringing the new adjusters’ letters up to the quality of the in-house team, Yo Sub says, made those carriers believers and gave him the confidence that there was a business.
I noted how much the industry has to dial up from a few adjusters to many overnight when a catastrophe hits. Asked about the most recent seasons, he said they have been very quiet this year and last, with small blips at customers and some IA firms onboarded, and business as usual for Voltaire.
How Many Letters Does an Adjuster Write?
I pressed for a number, and Yo Sub’s answer was that it varies widely by line of insurance and by how an organization triages. An adjuster who handles only easy payment letters writes about ten times as many as one handling complex cases, he says, maybe more.
He distinguished between letters an adjuster’s name is on and letters the adjuster touches. An acknowledgment letter at first notice of loss may carry the adjuster’s name but is generated automatically. Counting everything, he said an adjuster might be associated with something like 80 pieces of correspondence a week. Counting only the letters they actually work on, maybe 15 or 20.
Will AI Take the Adjusters’ Jobs?
I brought up Adam Price from Kinetic Comp, who told me on the show that a large carrier runs around 300 claims per adjuster, and asked Yo Sub the question everyone asks: is AI coming for our jobs?
Yo Sub says not with this software, because letter writing is one part of what an adjuster does, and many of the adjusters Voltaire works with are overloaded. More people are retiring than being hired, he says, and carriers he works with do not plan to replace everyone who leaves. They do worry about the expertise leaving with retirees. One selling point, he says, is that carriers no longer have to worry about how well younger adjusters write a letter, because that part comes off their plate.
I said that was a little scary for organizational memory, because their experience is gold. Yo Sub’s answer: many good adjusters are not good writers. They investigate carefully, review the policy and make a sound determination, and then lean on templates or old letters and copy and paste, which is where errors creep in.
The Error Rate in Carrier Letters
Yo Sub says Voltaire already outperforms human error rates statistically. His example is a large carrier that sends about 400,000 homeowners letters a year and estimates that about 100,000 of them contain some kind of error, which is 25 percent. Based on the data Voltaire has seen, he thinks that is a little optimistic and the real figure is closer to 30 percent.
Most of those errors are trivial, he says: a misspelling, the wrong insured’s name, a small mismatch. Most recipients throw the letter away.
The Appeal Arms Race
What Yo Sub is seeing now is that more policyholders are willing to appeal a claim letter even when nothing is wrong with it, because they paste it into ChatGPT, ask it to find problems, and send back the long legalese document it writes. I said I loved it: is my AI better than your AI? He agreed: “the arms race.”
When the Adjuster Overrides the Draft
I asked whether adjusters disagree with the drafts and change them. Yes, Yo Sub says, all the time at first, and it was one of the things Voltaire had to learn quickly in onboarding. Many carriers had let adjusters do their own thing, but once consistency was available they wanted to enforce it. His example is adjusters who always changed third-party language (“it was investigated and this was found”) to first-party (“we investigated”).
Those style decisions, he says, are now made at the director or VP level, and Voltaire can surface the pattern. If adjusters keep making off-script changes, the system can ask whether the AI should make the change permanently. Sometimes the answer is yes, and sometimes it is “we need to go talk to that person.”
The Audit Trail
I asked whether the system records the override. Yo Sub says the audit trail matters for two reasons. Voltaire was being blamed for mistakes in letters that turned out to be adjusters’ edits, and the trail lets the company show that the letter it generated was accurate. It also helps the legal department, which gets a record of who touched the letter, when, and what they did.
Policy Citations and the Adjuster’s Decision
Does the adjuster keep their own decision? Yes, according to Yo Sub. By the time an adjuster comes to Voltaire, the coverage review, the contact with the insured and the evaluation of the policy are done. The adjuster tells Voltaire the decision, and Voltaire writes the letter.
For a denial, the letter must cite the exact policy language that justifies it. That can be hard: a base policy modified by a special provision, further modified by endorsements, across hundreds of pages. Mis-citing, Yo Sub says, can be a very expensive problem for the carrier if the recipient acts on it.
Can the Platform Block a Denial?
Not exactly, Yo Sub says, but if someone tries to deny something for a reason that is not in the policy, the software reports that no language supports it and does not generate that letter. He calls it a safeguard rather than a decision-making system.
Off-Ramps for People
I asked about off-ramps for a human who wants to proceed anyway. Yo Sub says there is one: after a letter is generated, the adjuster can add any language they want. But the audit trail shows that the human did it, not the software.
I connected that to my conversation with Steve Johnson of insured.io in Episode 174. We talked about guardrails for AI inside a workflow, and the point is not only that the AI does not breach them but that the human does not either. Who owns the letter? Yo Sub: the adjuster. They sign it, they hold the license, and they are the ones summoned if there is a problem later.
Hiring, Onboarding and Job Satisfaction
Easier Hiring
Asked for a customer story, Yo Sub said one customer told Voltaire it made hiring much easier, because they could open up the threshold for applicants. Training new adjusters was simpler, he says, because so much of it had been about understanding how policies work and their hierarchical structure so that adjusters could cite them properly. With citations handled, new people onboard and become effective faster.
Insurance and the Technology Gap
Coming from outside the industry, Yo Sub says insurance is far behind other industries in technology adoption, and that as AI becomes more capable the gap is widening, because insurers are adopting more slowly than others. He expects the industry to adopt these tools like every other one, and processes that were tedious and manual to become obsolete. What new people learn and do, he says, will not be what their predecessors did.
Job Satisfaction
Asked whether Voltaire’s tool improves job satisfaction, Yo Sub says the company ran job satisfaction surveys and found that it greatly increased satisfaction, because many adjusters hate writing letters. Most of Voltaire’s users, he adds, are desk adjusters rather than cradle-to-grave adjusters who also go to the field, especially outside home and property lines.
The Name, and a Lightning Round
Is Voltaire named after the philosopher? Yes, Yo Sub says: Voltaire was a prolific writer, and the company writes a lot of letters.
In the lightning round:
- A letter he would never let AI send alone: by the way the software is built, he says, no letter goes out without the adjuster who signs it approving it.
- Where Voltaire operates: the US and Canada.
- The strictest states on letters: Yo Sub pointed to the Claims Compendium, a free resource that Voltaire makes available to anyone, covering what must be in each letter by line of insurance, with the case law behind the rules. The most complex jurisdictions, he says, are probably Florida, New York, California and Washington.
- Will anyone write letters manually in five years? Yo Sub says it probably happens sooner, depending on how long it takes carriers to feel pressure to upgrade. He sees that pressure rising as policyholders respond to more of the claims letters they receive and carriers are punished more for mistakes. He is confident it happens within five years.
- What he took from crypto and security: a lot of cybersecurity practices went into how Voltaire does things. He had expected insurance to care more about protections than it seems to, citing SOC 2 compliance and similar standards. I called those the basics.
What Is a Good Letter?
I asked for the best advice he received from a claims veteran. Yo Sub’s answer was a lesson rather than a quote: how many different opinions there are about what makes a good letter. The director of claims wants a letter that the recipient can read and understand. The legal team wants every clause in place. The QA team checks for something else. The adjusters want a letter that gets past their supervisor. Voltaire had to talk to all of them within each organization to build the picture, and Yo Sub calls it difficult and a challenging learning process, but one he says leaves the company well positioned.
What Insurance Should Talk About More
Yo Sub’s answer is a bigger conversation about AI regulation, and specifically what regulation is trying to accomplish. There is a lot of discussion about what AI should and should not be allowed to do, he says, and the skipped question is what the industry wants to achieve. He argues that the technology to handle claims end to end exists today, and can cut processing times, improve consistency and get policyholders paid faster. Instead of working out how to accelerate adoption safely, he sees a lot of conversation about what regulators might eventually say.
His warning is that regulation intended to protect consumers could end up hurting them by preserving a slow, expensive, outdated process. In his view, if an AI system can make claims handling faster, more accurate and more consistent while staying transparent and accountable, it should be encouraged.
I said that many people will want to talk about the accountability of AI, and that it is a topic for another episode.
My Take
Two things stood out to me.
First, the audit trail. When I asked whether the system records the override, Yo Sub’s answer was that the record protects both sides: Voltaire when it is blamed for an adjuster’s edit, and the legal department when it needs to know who touched a letter. It ties back to what Steve Johnson of insured.io and I discussed in Episode 174: you need guardrails for the AI inside a workflow, but also to make sure the human does not breach them. Recording the override puts that in writing.
Second, the scale. When Adam Price was on the show, he talked about a large carrier handling around 300 claims per adjuster. I said that if letters change, that changes the adjusters’ game and their effort, and then comes the question everyone asks: is AI coming for our jobs? Yo Sub’s answer was no, not with this software. I also found it a little scary for the organization’s memory when experienced adjusters leave, because their experience is gold.
As for the ChatGPT appeals, I called it an arms race: is my AI better than your AI?
Key Takeaways
- Voltaire started with claims letters and says it is expanding to more of the claims process
- Yo Sub says 88 percent of the letters Voltaire produces go out with no edits
- He says a home-property denial letter takes 30 to 60 minutes to write and Voltaire averages 3.7 minutes, and that a long casualty tender response can take an attorney six hours to two days against a couple of minutes plus about 30 minutes of review
- A large carrier he spoke to sends about 400,000 homeowners letters a year and estimates 25 percent contain an error, and Voltaire thinks it is closer to 30 percent
- Policyholders are starting to appeal claim letters by sending them to ChatGPT, which Yo Sub calls an arms race
- The adjuster makes the decision and signs the letter, the software cites the exact policy language, and it will not generate a denial for which the policy has no supporting language
- An audit trail records who changed a letter and when, which protects the vendor and gives the legal department a record
- Catastrophes drove early adoption: after Helene and Milton, carriers onboarded their independent adjusters quickly and brought letter quality up to the in-house level
- Yo Sub argues that AI regulation could hurt the consumers it is meant to protect if it preserves a slow, expensive process
Yo Sub’s figures, including the 88 percent of letters with no edits, the 3.7-minute and 30-to-60-minute letter times, the tender response times, the carrier’s estimated 25 percent error rate, Voltaire’s estimate of about 30 percent, and the job satisfaction results, are company-reported and were not independently verified.
Related on InsurTech.me
- Adam Price, Kinetic Comp: how a wearables company learned to listen to its customers (Episode 155)
- Steve Johnson, insured.io: the Pizza Tracker for claims and AI with guardrails (Episode 174)
Connect
- Voltaire: voltaire.claims
- Voltaire’s free Claims Compendium
- Yo Sub Kwon on LinkedIn
Strategy for carriers, MGAs and investors: let’s talk strategy.