ITC Vegas 2026 Field Notes

AI and distribution technology earn their place when they help a customer find the right coverage. Not when they generate more leads, and not when they automate for the sake of automating.

The panel was clearest about where specialty distribution matters most. As more admitted and excess and surplus carriers return to the market, standard risks are getting easier to place. The harder problem is complex risk, such as property in high-wildfire areas, where capacity is thin and brokers struggle to find any alternative to the FAIR Plan.

• Good triage includes knowing when to refer. No agency is expert in every product or appointed with every carrier. A strong process recognizes when a risk is not a fit and routes the customer to a carrier, agency, or specialist who can help, instead of leaving them with one poor option or none.
• Insurance-specific AI can improve the match. AI can help agents surface needs the customer did not mention and point the risk toward a better-fitting market. It supports placement and underwriting workflows. It does not replace the carrier’s underwriting.
• Automation pays off in retention, not just acquisition. Streamlining renewals and servicing frees agents for the ongoing conversations that keep customers.
• The opportunity is connecting a fragmented market. With so many specialized MGAs and carriers, the value is in helping an agent see which solution fits which risk.

My take: This is the version of AI in insurance I find most credible. It does not promise to replace judgment. It makes a complicated marketplace easier to work through and gives agents time back for the relationships that drive renewals. The test for any distribution tool is simple: does it get the customer to the right coverage faster, including when the right answer is somewhere else?

The real difference at a tech-native insurer is not a better screen. It is data that arrives structured and current, and that changes how authority is granted, monitored and pulled back.

The panel paired an operator with a builder. Charles Valinotti, who oversees delegated programs at Spinnaker, said data ingestion is the one system that is genuinely designed differently. Authority, by contrast, is still a human decision with a faster dashboard.

• Current data changes the program relationship. Bordereaux typically arrive 45 to 60 days late, and authority gets revisited at renewal. With monthly visibility, a carrier can tighten authority early or expand it in month five, six or seven for an MGA that is performing well.
• A program is a promise, so show data, not a PDF. Valinotti sees more than 250 program submissions a year. A strong underwriter with a weak data room rarely leads to a lasting partnership, because the carrier needs numbers it can query and trust for years.
• Ingestion works best in stages. Andy Moss of Duck Creek said teams triage on five to 10 data points first, then extract more as a submission advances. Extraction costs money in tokens and licenses, and the next step is surfacing the outlier in a submission, not just filling fields.
• The work shifts, it does not vanish. As premium grows, bordereaux reconciliation and initial triage shrink, while underwriting judgment and actuarial work on edge cases grow. Both speakers pointed to new roles: agentic architects who design the workflow first and place people around it, and teams that manage the whole portfolio rather than one program at a time.

My take: I am new to insurance, but this part felt familiar. In my work with auto manufacturers, public agencies like INDOT and nonprofits like the YMCA, AI projects succeed or stall on the quality of the data coming in, long before any model makes a decision. Insurance appears to be no different. Moss offered the best one-question test I have heard: as premium grows, what happens to your costs? If the honest answer is more people doing the same work, the operating model has not changed yet.

Giving people a general chatbot is not an AI strategy. Results came when AI was built around how the best people work and grounded in the company’s own data.

This was a practitioner panel: a carrier CIO and a software CTO describing what actually happened when they put AI into production. Both were candid about what went wrong first.

• Open prompting hurt consistency. Rich Fusinski of H.W. Kaufman said rolling out general chat tools to underwriters produced results that were “somewhat horrifying,” because everyone prompted differently. They wrapped it in an underwriting agent with tools designed with their top underwriters and brokers. On a good day, measured conservatively, they see a 50x return on token spend.
• Hallucination and drift are symptoms, not the cause. Glen Ceniza of Sapiens defined hallucination as a lack of accuracy and drift as a lack of consistency. The fix was not a better model but better context: Sapiens built a knowledge base from its code, documentation and tickets, and accuracy rose sharply.
• Code is now cheap, so prototype first. Both speakers said an idea can become three working prototypes within an hour. Since people do not know what they want until they see it, meetings now start from options instead of specifications.
• Lead from the front, and measure outcomes. Fusinski said he wasted a lot of time asking his team nicely to adopt AI. He now leads by doing the work himself and tracks throughput, time in backlog and quality, which at a minimum must not go down.

My take: I asked the panel whether moving from human in the loop to human at the end means fewer developers. Both said no: companies do not choose to do less when they can do more, and the backlog never runs out. That matches what I have seen outside insurance, in manufacturing and grant writing, where the gains came from building AI around how the best people already work, not from handing everyone a chat window. The practical test: whose expertise is built into your AI tools, and how will you know if quality slips?

Connecting carriers and brokers is no longer the hard part. When everyone has the same technology, the advantage goes back to data, expertise and a well documented process.

The panel paired a carrier and a broker. Reid Stanway of Ki Insurance described Ki as a fully algorithmic syndicate, about five years old, with 30 live classes and just over $1 billion in gross written premium. Belen Tokarski of Mylo runs a digitally enabled broker with more than 100 carriers and 40 integrations.

• Integration work has collapsed from months to a day. Mapping one party’s data model to another used to take a business analyst months. Now the schema and specs go into an AI model and the first pass is ready in a day. AI can also close the “data delta gap”: a broker sends 20 fields, the carrier needs 40, and neither side has to rebuild its API.
• Technology is no longer the differentiator. Integrating, quoting and recommending markets is now accessible to almost anyone. Tokarski said the edge is your DNA: solving the right problem for the client, inside the guardrails a regulated industry requires.
• Document the process before you automate it. Tokarski was direct: if a process is not documented, technology cannot improve it. With good integrations in place, Mylo now uses AI to compare carriers for a book of business, and its carrier panels are shrinking as a result.
• Agent-to-agent trading is real, but early. The first agent-to-agent trades have happened. Stanway expects 12 to 24 months of small pilots, such as one carrier with one broker on one class, and two to five years before it is common. The biggest hurdle is behavior, not technology.

My take: Tokarski’s point about documenting the process first is the lesson I have seen hold in every industry I have worked in, from manufacturing to nonprofits. AI speeds up a process that is already understood. It does not create one. Before any agentic project, ask a simple question: can someone show me, on one page, how this work is done today?

Parametric insurance is moving from global catastrophe cover down to small businesses and homeowners. It works when it is built into the sale, not pitched as one more product.

Parametric cover pays a set amount when an independently measured event happens, such as a power outage of a certain length, instead of paying on an adjusted loss. The panel was candid that it is still a nice-to-have that does not replace a core insurance program. But both speakers called this the first inning of a real shift.

• It fills gaps traditional policies leave open. Mike Gulla of Adaptive said typical power outage cover has a 72-hour minimum trigger, while 95% of U.S. outages last less than 24 hours. A parametric policy can put $10,000 to $15,000 back into a small business within about a day.
• Embedding beats educating. Peter Settel of Bold Penguin noted that optional add-ons sell poorly: about 1% of California customers carry a separate earthquake policy. So Bold Penguin embedded parametric cover as a cross-sell alongside business owner and commercial property policies, now live in 25 states. It appears where the data says it fits, with no extra questions, and the agent can remove it.
• An independent trigger simplifies everything. Because the payout uses third-party data, the customer cannot influence it. That makes pricing cleaner and changes the sales conversation from “what if you burn down” to “what does a 10-hour outage on your busiest day cost you?”
• Transparency is part of resilience. Some areas cannot be priced at all. California’s planned utility shutoffs are one example. Gulla argued that showing customers the risks no one will cover helps them plan, and builds the trust a new brand needs.

My take: As someone new to insurance, what stood out was that the product design and the distribution design were the same conversation. The data decides where the offer appears, the agent does no extra work, and the payout lets business owners decide how to recover. That is a useful pattern for any AI tool: put the right option in front of people at the right moment, and do not make them work harder to use it.

The underwriting bottleneck starts before underwriting does: in the inbox.

This was a 10-minute proof of concept. The speaker said 80 to 90 percent of an organization’s data is unstructured, and submissions often arrive as one email with many attachments. The demo took a broker email through four steps: capture the email and attachments, use AI to identify each document type and its fields, index and file it in one repository, and deliver the structured data into Salesforce.

• Keep the context with the document. Sender, recipient, subject and timing are captured with each file, so documents do not lose their link to the submission.
• The business decides what to extract. Teams define templates and fields for each document type, so everyone works from the same consistent data.

My take: This is the data ingestion problem the earlier panels kept raising, shown in practice. From my Salesforce work, I know a CRM is only as useful as the data that reaches it. The question to ask of any intake tool is what happens to the exceptions: the document the AI misreads or cannot classify.

Most AI in insurance is still stuck in the efficiency trap. The bigger payoff is better decisions, and proving that starts with a baseline.

The panel argued that AI has mostly been aimed at intake, the same place the industry has chased efficiency for 20 years. Expecting a different result from the same target misses the point. A small improvement in loss ratio is worth far more to the bottom line than a large cut in expenses.

• Look past the time savings. Jay Bell of Zurich described an AI agent that reviews contractor agreements. It saves underwriters time, but the real value is a consistent, saved review of every contract. In three to seven years, that record will show which contract terms actually predict losses.
• Set the baseline before you start. Lauren Dieterich of Arch said every initiative should begin with the underwriting behavior it aims to change, and today’s numbers for it: hit rate by risk tier, retention, growth in target segments, rate adequacy or loss ratio.
• From scores to context. Analytics used to hand underwriters a risk score and leave them to work out why. AI can now explain the key risk drivers and compare similar risks, which moves underwriters from gathering data to making judgments.

My take: The baseline point applies far beyond insurance. Without one, every AI project ends up defended with hours saved, because that is the only number anyone can measure. Before the next project starts, write down the one business result it should move and where that number stands today.

Insurance could fund resilience before a hurricane lands, not only recovery after it.

Tom Philp of Maximum Information opened with Hurricane Maria in 2017. The forecast left little doubt about the track or the strength, and underwriters wanted to help clients prepare, but indemnity contracts need a confirmed loss, so nothing could move until after the storm. In the speaker’s words, it was not a failure of science or of the catastrophe models. It was a failure of the structure around the information.

• Combine forecasts with catastrophe models. Together they can estimate payouts at four, three and two days before landfall. On a hypothetical $10 million limit, the Maria forecast gave 99% confidence four days out that at least $500,000 would be paid.
• Tie early money to no-regrets actions. Working with disaster response groups, early funds can go to shutters, flood defenses, generators and evacuation, so the money builds resilience even if the storm misses.
• It is a capital question. Paying for prevention helps most in the major events that hit capital hardest. Philp was clear this is early thinking, and the numbers are not yet worked out.

My take: This was the boldest idea of the morning, and it echoes a theme from every session today: the data already exists, but the structure to act on it does not. The question for any organization is where it has good information arriving too early for its processes to use.

Before deploying AI, leadership has to decide what it is for: sharper decisions or more volume per person. That choice shapes everything, including governance.

The panel described two paths. Some carriers use AI to help underwriters and claims handlers make better, faster decisions. Others use it to scale how many policies or claims each person can handle at today’s level of expertise. Both are valid, but the choice should be deliberate.

• Autonomy follows complexity. Simple work, such as checking a veterinary bill on a pet insurance claim, is already moving to true automation. On complex risks, such as a commercial D&O policy worth $1 million or more, agents will act as colleagues, and the support team around a senior underwriter will become a set of specialized agents.
• Governance should match the goal. When AI supports decisions, governance needs to be stronger, because people will come to treat the agent’s answer as the truth. For high-volume, low-complexity tasks, a crawl, walk, run approach works: test accuracy first, then allow more autonomy.
• Build and buy are converging. AI-native insurtechs treat their own models as their advantage, while established carriers lean toward partners. The panel described the two groups as moving from cousins to siblings.

My take: The point about people treating the agent as the truth is the one every organization should plan for, in any industry. Governance should be set by the kind of decision, not by the tool. A good test for each AI tool: if it were wrong, who would notice, and how quickly?

Insurers are spending heavily on AI, but most are stuck in what Cognizant calls the “blur zone”: high investment, little realized value.

This was a short sponsor presentation that launched nine use cases shown at demo stations afterward. It opened with a comparison of AI investment against value realized. Tech-native companies land in the “value zone,” while by Cognizant’s analysis only about a tenth to a third of insurers realize business value, and much of that is siloed and project based.

• Fix the experience and the P&L together. The framework pairs trust, empathy and moments that matter with protection, growth and retention. The argument: improving the customer experience alone does not move the business.
• Start small and fast. Each use case was designed to be implemented within 12 weeks from a reference architecture.
• Governance is lagging. The presentation cited that more than three in four organizations are deploying AI agents, yet only 13% feel ready to govern them.

My take: The governance number is the one I would take home. Agents are arriving faster than the rules for overseeing them. A simple starting test for any organization: can you list every AI agent running today, what it is allowed to do, and who owns it?

Your data does not need to be perfect to start. But proving value takes a plan: does the AI work, do people use it, and does it move the business?

The panel agreed that the most common mistake is treating data as a one-time IT project instead of an ongoing business discipline. The second is trying to boil the ocean before starting any use case.

• AI makes old data usable. David Hurst of CRC Group gave an example: 20 years of business, but only five of the 25 data elements he wanted were ever captured in structured form. Pointing AI at the stored documents turns what would have been a year of cleanup into a short project.
• Measure in three steps. Hurst’s team first checks whether the AI performs as expected, for example whether submissions it ranks highest actually bind at a higher rate. Next comes adoption, because a great model nobody uses earns nothing. Only then do they look at business value, through leading indicators such as quote rates, bind rates and cost to service.
• Use the smallest model that works. Kenny Holms of Accelerant said simpler and local models often handle simple tasks as well as the largest ones, faster and at lower cost. That makes it affordable to test and improve across a whole portfolio.
• Winners say no. With so much hype, the companies that pull ahead will focus on the few metrics that move their bottom line and turn down the rest.

My take: The three-step measurement was the most practical framework I heard today, and it applies in any industry. Many AI projects jump straight to “did it make money” and cannot answer, when the honest first questions are whether it works and whether anyone uses it. Try it on your own AI project: which of the three can you measure today?

Do not add a chatbot to your portal. The value comes from redesigning the workflow from the ground up, not from putting AI on top of an old one.

Agents’ expectations used to be set by digital experiences outside insurance. Now AI sets the bar: an agent who can build a submission in five minutes does not want to wait on a portal or an underwriter. But when automation breaks, Jake Gilbert said, agents want a person, not a chatbot or a ticket.

• Most friction comes from digitized paper. James Thom argued the industry has mostly turned paper forms into web forms and PDFs. One large carrier with 18 different intake points is building a single entry point and routing the work internally, instead of making agents figure out where to go.
• The opportunity is large and specific. Vertafore estimates $2 to $3 billion of economic gain inside U.S. agencies from AI-driven automation, across producer workflows, submissions, servicing and financial operations. Submissions and servicing hold the most value.
• Go hard at one process. Thom’s warning: experimenting around the edges will not deliver enough value to prove the direction. Pick a workflow that matters and redesign it with conviction.

My take: Gilbert’s practical advice is worth borrowing in any industry: take off your company hat and go through your own process as your customer would. Onboard through your own portal, call your own help desk, and notice how many questions come before anyone knows who you are. That walk-through usually shows which workflow to redesign first.

The morning’s most important numbers came from EY: 78% of insurers plan to have agentic AI integrated within two years, but only 7% have reached meaningful scale, and 46% of initiatives stall between proof of concept and deployment.

The Main Stage ran four very different conversations, from a Shark Tank investor to a free solo climber. Each one came back to the same ideas: listen to the customer, own the outcome and respect risk.

• Institutional arrogance is what kills companies. Daymond John was candid that FUBU stopped listening to its customers as it grew. It made retailers take full boxes of jeans, including sizes they could not sell. Competitors offered only the sizes stores wanted, retailers dropped FUBU, and it was left with about $100 million in unsold inventory. Every company should keep asking: why me, why now, why this? Bill Powers of Cambridge Mobile Telematics, which serves 21 of the 25 largest North American auto insurers, said AI is letting six people do work that once took 12, with the open roles shifting to AI talent.
• Insuring the robot driver. Skyward Group and Apollo ibott described a single autonomous vehicle liability policy built for Uber, combining what used to be several separate policies and charged by the mile. Autonomous vehicles appear far less likely to crash, but the speakers warned that jury verdicts could make each claim far more expensive.
• Pilots stall because nobody owns production. The AI readiness panel (Tokio Marine, EY and Microsoft) said to start with one document-heavy, high-volume, low-judgment workflow, write down the rules that govern it, test against a few hundred cases with known answers and name a business owner. That is six to eight weeks of work, not an 18-month data project. And with regulators, “AI told me so” is never an answer.
• Feeling safe can be the danger. Alex Honnold said many of the scariest climbs of a long career happened with a rope, because the rope made it easy to keep pushing further into the unknown. Without one, the decision to turn back is simple. Honnold also sets no-go conditions in advance so pressure on the day cannot change them.

My take: Honnold’s rope story is the best description of AI pilots I have heard. A pilot feels safe, so teams keep adding scope without ever deciding who owns it in production. The EY numbers show the result. One practical test: before a pilot starts, can you name the business owner who will answer for it when it is live?

Trusted, connected data comes before agents. AI will scale a good process, and it will just as quickly amplify a bad one.

The session paired a Salesforce product roadmap with a customer story. Regina Felts of Comparion Insurance Agency, which has more than 1,500 agents, said the first priority after joining 18 months ago was bringing the agency’s data into one trusted, connected place. Only then did AI tools make sense.

• Trust drives adoption. Agents will not use AI they cannot trust in front of a customer. Clean, connected data is what earns that trust.
• The goal is time in front of customers. Felts wants AI to handle the preparation and follow-up so agents spend more time in the field having better conversations.
• Ingestion is the hardest underwriting problem. Salesforce described an underwriting product coming in the first half of next year that pulls submission data arriving over days and channels into one view, and flags conflicting information for the underwriter. A voice AI agent for first notice of loss is due next month.

My take: From my own Salesforce work, I know the platform is only as strong as the data model underneath it. Comparion spent 18 months on data before agents, and that order is right. If you are planning agentic work on any CRM, ask first: would your best people trust today’s data in front of a customer?

Build the small things, buy the complex ones, and never give away control of your data and your decisions.

A carrier CEO and a startup founder took opposite sides of the build-or-buy question and ended up close together. AI makes small builds easy, but the hard parts were never the code.

• Small builds are now quick. Mike Zukerman of CSAA said CSAA’s IT team built a chatbot that answers questions about a member’s policy in two weeks. Large projects that span several systems are a different story, and Zukerman does not want to build those in house.
• Talent and data were always the hard part. Sean Merat of Owl.co argued that a vendor’s real advantage is data pooled across many carriers and the feedback from how users act on its results, which takes years to build.
• Heavy customization is building in disguise. Both agreed carriers overestimate how unique they are: roughly 80% of requirements are shared and 20% are truly different. Zukerman described replacing a home-grown payments system whose builders have since left.

My take: Merat drew the clearest line of the day: carriers should keep control of their data and their decisions, and let tools do the rest. Zukerman added that moving too slowly is now the bigger risk. A simple test before any build: is this one of our 20% differences, or one of the 80% everyone shares?

The most convincing AI demo of the day came from an operations leader, not an engineer, and the biggest lesson was about procedures, not models.

Palomar walked through how it uses AI agents from Pace across servicing and claims. The examples were small, concrete and already in production.

• Agents built by the people who own the work. Josh Woertz built an agent in under an hour that finds claims emailed to any of Palomar’s many inboxes and enters them into the online claims form. Pace said a year ago about 90% of its agents were built by engineers; now about 90% are built by operations managers.
• Surge capacity without overtime. When a hurricane threatened Hawaii’s Big Island on a weekend, Woertz updated the voice agent’s greeting and routing on Saturday night. Claims calls jumped from about a dozen on a typical weekend to as many as 100, and every one was answered and routed without calling in staff.
• Write down the edge cases first. Woertz’s advice, looking back a year: tighter procedures up front, because too much knowledge was still in people’s heads. Monitoring production matters too. Some callers learned that asking twice for a person skipped the agent.

My take: This echoes a point from Tuesday: you cannot automate a process nobody has written down. What impressed me most was the feedback loop, with call center staff flagging problems in real time alongside nightly quality reports. Before launching an agent, ask: who will review its mistakes every week, and how fast can they fix them?

Accountability for AI is shared, but it stops with the carrier. Vendors must explain their models, and carriers must own the final decision.

The panel covered who answers for an AI-driven decision, how governance should scale with risk, and why data quality sits underneath all of it.

• Governance follows the risk. Sandy Ulrich of Pharmacists Mutual said controls depend on what data a tool uses, what decisions it makes and whether customers see it. Internal tools get more room to experiment; customer-facing ones get more scrutiny.
• Adverse decisions need a human and an explanation. Stephanie Kuczynski of ZestyAI said regulators vary state by state, but when AI contributes to a negative outcome, the carrier must be able to explain and correct it. In Kuczynski’s words, it is someone’s grandmother on the other end.
• Expect more vendor vetting. Kuczynski predicted that within six months, carriers of every size will send AI vendors formal questionnaires on traceability and accountability.

My take: This is the conversation I expect most organizations to have next, in every industry. Ulrich’s company also invested in AI literacy and internal AI champions, because a human in the loop only helps if that person understands what they are approving. A practical starting point: for each AI tool, name the person accountable for its outcomes.

AI is making fraud cheaper and faster to commit, and insurers are in an arms race to keep up.

The panel described a sharp rise in AI-assisted fraud: fake photos, fake narratives, fake identities and even fake video requests for wire transfers.

• Connect claims and underwriting. A Liberty Mutual claims leader said the winners will link data that used to sit in separate departments, and move from reacting to fraud to spotting it earlier.
• Everyone has the same tools. If all insurers can buy the same AI, the difference comes down to data, operating model and people. In the panel’s words, everybody has an oven; the recipes and the cooks win.
• Use the right model for each task. Shift Technology now benchmarks models weekly and routes simple work to cheaper models and complex work to the most capable ones, which keeps costs under control.

My take: The same tools that help insurers also help fraudsters, so standing still means falling behind. The regulator point stuck with me too: an insurer stays accountable for any vendor’s AI it uses. Worth asking: if a document or photo in your process were fake, where would you catch it?

Reinsurance is one of the biggest levers for growth and capital, yet much of it still runs on spreadsheets.

The panel argued that better reinsurance data lets carriers grow with confidence, manage risk and run closer to their capital limits.

• Systems handle the easy 80%. Most systems cope with simple treaties but break on complex structures with multiple participants, different terms and reporting lags. Excel remains common, and it is hard to audit.
• AI reads contracts at scale. Finding specific clauses or inconsistencies across a treaty inventory used to take days of reading. Now it is nearly immediate. AI can also flag where a third-party administrator’s data conflicts with internal records.
• Precision is a capital advantage. The more precisely a carrier knows its position, the closer it can manage to regulatory thresholds without raising or tying up extra capital.

My take: As someone new to insurance, this was the session where the money behind the money became visible. The pattern is familiar from every industry: critical numbers living in spreadsheets nobody can audit. Ask where your most important figures still depend on one person’s workbook.

Agents already keep the data carriers need. The opportunity is to stop asking them to type it again.

Applied Systems, whose agency management software is used by about 70% of retail agencies, acquired Cytora, which works with seven of the ten largest global commercial insurers. The two are now connected, and the first carrier links went live the week before the conference.

• Rekeying is the real problem. The same submission data is digitized several times between the customer, the broker and the carrier, often ending up rekeyed from a PDF.
• Quotes without a submission. The first use case is renewal remarketing: a broker picks part of a book, and carriers return quotes from data already in the agency system, without a new submission.
• Email to quote. Liberty Mutual described letting agents email any document, even a voice memo, and get back a price indication and appetite check. The pilot was built in roughly eight to ten weeks on existing connections.

My take: This is the “meet people where they work” idea from Tuesday, now live. The agency system is both the record and the place work happens, so putting carriers inside it removes a whole layer of friction. The question for any business: where are you asking partners to re-enter data you could already have?

Day 3 opened with three Main Stage conversations and one message: technology is easy to buy and hard to put to work. Whether the subject was AI or a burst pipe, the value comes from how it is used.

The first panel tackled AI liability. The moderator debuted a heat map that crosses the type of AI system with where it fails, illustrated by a fictional AI-assisted law firm. In one example, a hidden instruction in a client email pulled another client’s confidential material into a draft. In another, the AI gave a contract notice date of October 31 when the documents said September 30. Both drafts passed a lawyer’s review. The second panel, “Stop buying AI, start shipping it,” featured Guru Vasudeva, CIO of Nationwide, and the chief client officer of a large insurance distributor. The morning closed with Jonathan Scott of Property Brothers on building homes that hold up.

• Underwrite the use, do not exclude the label. A lawyer on the liability panel asked carriers who want to exclude AI what they mean by it, since spell check and legal research tools qualify. AI mostly amplifies risks a business already has. And “silent” AI exposure is only silent until a court tests it.
• Three kinds of AI at Nationwide. Everyday AI tools now reach all 30,000 associates and contractors. In software engineering, Nationwide trained three people on each of roughly 400 development teams to coach the rest, and sees about 15% efficiency on average and up to 40% in some teams. Transforming the business itself is the hard part.
• Obsess over the outcome, not the feature. For agribusiness submissions, the goal is not faster triage but doubling the territory each underwriter can manage. Vasudeva also named a paradox: the more AI automates, the more domain knowledge people need to judge its answers.
• Shipping means solving the physics. One fix came from a carrier relations leader meeting a carrier over a meal and asking for attachments instead of links, because no AI can read a “click here” email. The distributor abandoned automated logins after a one-time password on a manager’s phone caused 30 escalations, and now seats business and tech teams side by side. The goal is to move service agents from about 80% tasks and 20% advice toward 50/50.
• Build back better, and let insurers lead. Jonathan Scott is rebuilding a family home lost in the Palisades fire with a system that has no timber and resists fire, earthquakes and termites, for about 10% more than a wood frame. Of roughly 800 homes being rebuilt nearby, almost all are wood. Scott said insurers often push resilient technology before building codes do, and pointed to cheap fixes too: fine ember mesh over attic and crawl space vents, no mulch against the house, and a whole-home water shutoff that learns normal use to avoid false alarms.

My take: The two panels fit together. The wrong-date example was a review that checked the draft and not the source. The shipping panel showed that the hard part is the unglamorous detail of how work really flows. Both come back to people who know the business well enough to know what to check. A useful test for any AI workflow: does the person approving the output know the work well enough to spot what is wrong?

Agentic AI reaches production when leaders define the outcome first, retire the old process and pay for change management.

A life and retirement carrier, a large insurance broker and an AI vendor compared notes on getting agents out of pilots. Principal Financial’s example came from pension risk transfer, where a single participant can be covered by many plan documents and rules.

• Start from the center of the maze. Sri Reddy of Principal said to define the outcome first and work backward, because many existing processes rest on old assumptions and self-imposed limits. To see them, Reddy had leaders from an unrelated business review the operation for six weeks, with no legacy assumptions.
• Turn off the old way. When Principal deployed AI, it shut down the previous process, because running both side by side breeds resistance and doubt. Phillip Maljevic of Alliant added the other half: build the new system so useful that people want it.
• Agents need a map of the business. Chirag Jindal of Plutus argued that most insurance work crosses CRM, policy and claims systems, so agents need a shared business vocabulary and a layer that tracks work across systems. Without them, AI solves small moments while token costs rise and value stays unclear.
• Do not cut change management. Maljevic said it is often the first line removed from a statement of work, and it should not be. Alliant gathered its claims leaders in one room to find shared workflows and created an AI center of excellence. Principal gave AI training to its entire workforce.

My take: The change management point hit home, because I write statements of work, and I have seen that line treated as optional. It is usually the difference between a system people use and one they work around. A simple test for your next AI project: is change management in the budget, and would it survive the first round of cost cuts?

Carriers are not getting closer to AI at scale. Some are slipping backward, and the fix is to put AI inside the systems where work already happens.

Christina Colby of Guidewire shared a benchmark of property and casualty carriers around the world. In February they rated their AI maturity at about three on Gartner’s five-stage scale. Asked again in September, they had slipped back slightly. Their reasons: data quality, a lack of people who can build agents, and resistance to change. In Colby’s view, these are the same barriers as any large transformation.

• Three tests for embedded AI. It must be contextual, seeing only what the person it works for is allowed to see. It must be embedded in the workflow, and not every step needs an agent; rules and traditional models still suit predictable decisions. And it must be accountable: observable, auditable and continuously evaluated.
• The apprenticeship paradox. In underwriting, senior staff tend to feel threatened while juniors welcome AI as a shortcut to experience. In claims, Colby often sees the reverse: senior handlers are glad to shed routine work, while juniors feel uneasy with decisions they lack the experience to check, and early data points to rising attrition among them.
• Build it inside the core system. Shigeo Kuwabara of GenerativeX, after 35 years at Tokio Marine, now works as a forward-deployed engineer who sits with the carrier’s team. The demo showed a claim moving from first notice to adjuster assignment in about two seconds inside the existing claims system, including spotting a $1,200 gap between a repair estimate and the standard schedule.

My take: In the Q&A, Kuwabara said coding agents have matured so much that a career of business knowledge is now a bigger advantage than coding experience. The apprenticeship paradox is the one leaders seem least prepared for. If AI does the groundwork juniors used to learn from, how do they build judgment? Worth asking in any team: who is learning the work that AI now does the first draft of?

A pilot that works is not the same as a pilot that scales. Agentic AI pays off when it enriches decisions, runs on shared plumbing and has an owner.

The panel brought together a model provider, a cloud provider and two carriers. Tony Jacob of OpenAI noted that new frontier models now arrive about every six weeks, far faster than the industry adopts them.

• From chatbot to agent. Jacob drew the line clearly. A chatbot answers a question about a flooded basement. An agentic service handles the whole claim: it pre-fills the notice, gathers the details, assigns an adjuster and lines up repair partners. The next step is several agents coordinating with each other.
• Enrich, do not just compress. Patrick Gallic of Tokio Marine described a team that cut submission to quote from about three hours to a few minutes. The real change, though, was pulling in data they never had before, such as weather forecasts and ownership records. Speed alone just competes with outsourcing.
• The pilot that could not scale. Gallic shared a submission intake project that reached about 90% accuracy and delighted the business. But it was built on five direct model calls with no evaluations, routing or cost controls, so it could not be copied to other lines. Scaling needs model testing, routing, cost tracking and a named owner from the start.
• Keep your options open. The panel urged one shared layer for governance, observability and token costs, so carriers can pick the best model for each task and change providers as prices and capabilities shift. Gallic added that most of a process can stay rule-based, calling AI only for the steps that need judgment.

My take: Gallic’s honest story is the most useful kind of conference example: a success that taught a lesson. The pilot worked, but nobody had designed it to be repeated. A good test before any pilot goes live: if this had to run in 100 places next year, what would break first?

Not all AI is a large language model. For predicting claims, purpose-built machine learning on curated data still does the heavy lifting.

A broker and an analytics firm described using predictive models in workers’ compensation and liability claims. The aim is not faster paperwork but lowering the total cost of risk: premiums plus claims.

• Spend scarce attention on the right claims. Peter Enos of Gallagher said the model flags claims likely to grow, such as one reserved at $10,000 that is on track to settle for $250,000. Focusing resources there has helped clients cut claim frequency and severity by 30 to 50%. It also takes emotion out of the conversation, which helps at renewal.
• Better intake, better predictions. Chris Connell of Mobotory said a 150-character claim description is a weak input. Structured detail, such as which knee was injured, makes predictions far more reliable. Connell also noted that legal costs make up about a third of claim costs in the data the firm sees, from roughly 12% at the best performers to 50% at the worst.
• Test your AI vendor. Ask whether they can show how accurate they are at different claim sizes, whether they build their own models or wrap a general one, and whether they curate their data. If you could get the same answer from a chatbot yourself, why pay for it?

My take: This was a vendor-led panel, so the numbers deserve a closer look before anyone relies on them. But the buying advice was sound, and it applies in every industry I have worked in. Enos framed it best: augmented intelligence, not artificial, because the tool should make a person better at the job. Before you buy, ask the vendor: how do you measure your own accuracy, and can I see it?