Insight or Instinct? What It Actually Takes to Let Data Lead in Underwriting
In this article:
- Why the commercial case for data investment in underwriting has fundamentally changed, and how to make it internally
- How the best carriers are treating automation as a talent strategy, not an efficiency play
- Why simulating certainty with AI is more dangerous than acknowledging uncertainty, and where to draw the line
- What the gap between data richness and decision quality is costing carriers, and what closes it
Setting the scene
Insurtech Insights USA returned to the Javits Center in New York City this June, drawing thousands of insurance executives, investors, and technologists across two days of sessions spanning embedded insurance, climate risk, AI infrastructure, and distribution strategy. One fault line ran beneath nearly every conversation: the industry has more data than it has ever had, and is still not making decisions as well as it should. The decisions made in the next two years will draw a hard line between carriers that price risk with confidence and those that are still guessing at scale.
One session addressed this directly. Insight or Instinct? Using Data to Sharpen Underwriting Decisions at Scale convened a candid panel moderated by Ruth Foxe Blader of Citrine Venture Partners, featuring Bryan Brizzi, Chief Digital Officer at Crum & Forster; Caolan Kovach-Orr, Head of AI and Automation at Arch Capital Services; Vicki Garrett, a 30-plus-year life insurance underwriter at Swiss Re; and Don Keleman, Managing Director US at Trendtracker. Not a technology showcase. A working conversation between operators with real books of business, offering unfiltered perspective on what the shift to insight-led underwriting actually requires.
Data is no longer a cost center. It is a revenue argument.
For most of its history inside insurance carriers, data governance was a defensive investment: better reporting, cleaner audits, more consistent submissions handling. The commercial case was absent, and the people building data programs knew it. Bryan Brizzi, CDO at Crum & Forster, named this directly. When he led a data governance initiative at a prior organization, a business leader asked the question that always comes: is this going to allow us to sell more insurance? At the time, he could not say yes. Two years later, his answer is unambiguous. "With better, cleaner, quality data, we can write more policies."
That shift changes the internal argument entirely. Data quality initiatives and governance programs that were previously framed as infrastructure costs now have a revenue line attached to them. Carriers that made those investments early are beginning to show it in underwriting performance. Those that did not are funding catch-up from a position of competitive disadvantage.
What this demands in practice: Stop pitching data investment as operational hygiene. The CFO and CUO case is now written premium, loss ratio improvement, and speed to bind. Quantify the impact for your specific book. Programs that cannot connect to a revenue or risk outcome will keep losing the budget conversation.
The AI conversation is skipping a step. That is why deployments underperform.
The industry has spent two years jumping straight to the model, the automation layer, the decision-support tool, without reckoning seriously with the infrastructure beneath it. The result is a recognizable pattern: AI deployments that underperform not because the technology is insufficient, but because the data feeding it is not. Vicki Garrett, with more than three decades in life underwriting at Swiss Re, put it plainly: the question is not whether the tools are capable. It is whether the data ecosystem underneath them is coherent enough to make the outputs reliable. Submissions data, loss history, and external enrichment sitting in separate systems with no common data model does not get resolved by the AI layer on top. It gets amplified.
A 360-degree view of risk is not a software feature. It is the outcome of sustained investment in data connectivity and quality control. The gap between knowing that and having done it is where competitive divergence accumulates.
What this demands in practice: Before the next AI evaluation, audit the data architecture honestly. The carriers extracting the most from AI investment have already resolved the connectivity problem. For those that have not, that work is the highest-return investment currently available.
Existing governance frameworks are exactly what AI requires. Apply them.
Caolan Kovach-Orr, who leads AI and Automation at Arch Capital, offered a necessary corrective to current market enthusiasm. "You still can't just throw data into an AI model and expect anything valuable to come out." The rigor required for responsible AI deployment is identical to what model governance has always demanded. What has changed is the scale of the failure mode when that work is skipped.
A traditional predictive model that produces flawed output does so at a pace that allows detection. A generative AI system can produce confident, directionally wrong outputs at scale, faster than any prior generation of tools. The actuarial and data science functions inside carriers already know how to govern models. That discipline needs to be extended to AI outputs with the same rigour applied to pricing and loss development models. Treating generative AI as exempt from those standards is a risk that will show up in combined ratios before it appears in any review log.
What this demands in practice: AI in underwriting belongs inside the same governance structures applied to any actuarial model. Defined validation protocols before production. Ongoing performance monitoring. Clear escalation paths when outputs fall outside expected ranges. The teams building durable AI capabilities are the ones extending existing frameworks before the failure modes become visible, not after.
The goal of AI in underwriting is orientation, not certainty.
The panel pushed on where AI-generated insight is genuinely useful versus where it creates a false sense of precision more dangerous than acknowledged uncertainty. Don Keleman, Managing Director US at Trendtracker, recentered the conversation:
"Strategy is not an exact science. If you ask McKinsey to help with your strategy, it's close, but it's directional. And that's the point."
Much of the frustration with AI in underwriting comes from a mismatch in expectations: tools deployed in search of certainty, returning probability instead. But strategic decision support has never operated to a certainty standard. It operates to an orientation standard: am I better informed, faster, with greater confidence in the direction of my judgment?
The carriers making genuine progress use AI to surface patterns, reduce analytical burden, and compress the time between signal and decision. The judgment still belongs to a human. The question is how prepared that human is when they exercise it.
What this demands in practice: Define the confidence framework before deployment. For high-frequency, lower-complexity risks, automation thresholds can be set high. For complex or high-exposure risks, AI output is structured input to a human decision, not the decision itself. A tiered approach by decision type is what defensible AI governance in a regulated industry actually looks like.
Automation is a talent strategy. The best carriers have made that reframe.
Experienced underwriters are scarce. The contextual judgment built over decades of pattern recognition does not live in any system and does not transfer easily. That reality defines what automation is actually for.
Its most valuable use is not replacing experienced judgment. It is protecting the conditions under which that judgment can be applied where it matters. In most carrier operations, senior underwriters spend a disproportionate share of their time on high-volume, lower-complexity work: routine renewals, straightforward new business, administrative processing. Every hour consumed there is unavailable for the complex, novel, or high-exposure risks where experience genuinely changes the outcome.
The carriers getting this right have stopped measuring automation by cost per decision and started measuring it by capacity redeployed toward higher-value work.
What this demands in practice: Map the underwriting workflow against two dimensions: decision volume and the degree to which experienced judgment changes the outcome. The highest-return automation targets sit at the high-volume, lower-complexity intersection. Automate there first, define where the freed capacity goes, and hold the program accountable to that redeployment.
Performance visibility is an early warning system. Most carriers are running it as a reporting function.
Ask underwriting leadership whether they have real-time visibility into loss ratio by segment, pricing adequacy by line, or submission-to-bind dynamics by channel. The answer is almost always no, and not because the data does not exist. The gap is between data sitting somewhere in the organization and intelligence continuously surfaced in a form leadership can act on.
The consequences are predictable. Loss ratio deterioration surfaces in quarterly results, by which point the decisions driving it are months old. Pricing inadequacy compounds quietly across renewal cycles. Leakage persists until someone commissions an analysis on a cycle entirely disconnected from when the problem began. A performance dashboard is not a reporting enhancement. It is an early warning system. Most carriers have the former when they need the latter.
What this demands in practice: Underwriting performance visibility requires business ownership, not technical delivery alone. Loss ratio by segment, rate adequacy trends, new versus renewal mix: these must be defined by underwriting leadership and surfaced continuously as live inputs to decisions. If performance data arrives quarterly, the question is not how to improve the report. It is how to compress the lag to the point where course-correction is still possible.
Intelligence at market speed: what closing the gap requires
Every insight from this session points to the same conclusion. The competitive advantage in underwriting is shifting from data access to decision velocity: the ability to act on the right information faster and with greater confidence than the market around you.
Markets reprice faster than annual review cycles can track. New risk categories emerge and evolve faster than frameworks built for stable risk classes can absorb. A carrier whose strategic intelligence operates on a quarterly cycle, assembled manually by teams already managing the book, is structurally behind. That gap compounds.
This is where Trendtracker operates. Built for strategic decision-makers, Trendtracker is an AI-powered strategic intelligence platform that brings everything together in one place: market shifts, regulatory developments, competitive moves, and emerging risk signals, continuously monitored and synthesized into structured intelligence that underwriting and strategy teams can act on. No more scattered sources. No more manually piecing together context from disparate reports, inboxes, and industry publications. The intelligence insurers need to make confident decisions, organized and ready when it matters.
The panelists at Insurtech Insights USA named what every leader in that room recognized: intelligence that sharpens judgment without replacing it, that reduces noise without losing signal, and that arrives in time to inform a decision rather than explain an outcome. That is what Trendtracker delivers.
The carriers that close the gap between data richness and decision quality first will write better business, manage portfolios with greater precision, and respond to market change before their competitors see it coming. In an industry built on pricing uncertainty, that is the most durable advantage available.
