What AI Agents Are Actually Doing in Insurance Underwriting

Will Drewes

Founder, Fern Strategy · 3 min read

The numbers that exist, and the ones that don't

If you've been tracking AI adoption in insurance underwriting, you've probably seen the same aggregate claims recycled across analyst decks: processing times cut by 70%, loss ratios down 18.5%, combined ratios improving 3 to 6 percentage points. Those figures come from McKinsey, BCG, and Deloitte. They're real enough as directional signals. But they're industry-wide averages, not company-specific results you can pressure-test.

Named, auditable outcomes are harder to find. Here's what the public record actually shows.

Zurich Insurance: processing time cut in half

Zurich implemented an AI underwriting engine focused on policy processing consistency and regulatory compliance. The reported outcome: 50% reduction in policy processing time. That's a meaningful operational number, not a customer-experience metric. Faster processing in underwriting means faster premium binding, fewer handoffs sitting in queues, and less manual re-keying between systems. The compliance angle matters too. Consistency across underwriters is one of the harder problems in regulated lines, and it's where rules-stable, input-messy workflows tend to benefit most from automation.

The limitation worth naming: the public source doesn't pin an exact implementation date, so it may predate the last 18 months. The outcome itself is credible and specific.

Lemonade: 90-second quotes at scale

Lemonade's AI-driven risk assessment delivers quotes in under 90 seconds for standard personal lines policies. This is a speed and volume metric, not a loss ratio figure, but it reflects something structurally important. Lemonade built its underwriting stack around automation from the start rather than layering AI onto legacy processes. The 90-second number holds because the rules don't move. The inputs (applicant data, property details) are messy and varied. That's exactly the condition where agents outperform static rules engines.

For incumbents, the lesson isn't to copy Lemonade's consumer model. It's that the speed gap between AI-native and legacy underwriting is now visible to customers and brokers, not just to operations teams.

What the aggregate data signals

The industry benchmarks, even without named company attribution, point to a consistent pattern. Insurers using AI in underwriting are reporting 3 to 6 percentage point improvements in combined ratios and loss ratio reductions around 18.5% in segments where the tooling is mature. These aren't rounding errors. A 3-point combined ratio improvement on a large commercial book is a material underwriting profit swing.

The gap between those aggregate results and the thin public record of named case studies is itself informative. Most of the real outcomes are sitting in private vendor case studies or internal reports that never get published. Carriers that are ahead aren't advertising it.

What this means for underwriting leaders

The public data confirms the direction but obscures the specifics. That creates a build-vs-buy problem with limited reference points. Before committing to a platform or a custom build, the questions worth answering are: Where exactly in your underwriting workflow do stable rules meet messy inputs? What does your current audit trail look like for automated decisions, and will it hold up to a regulatory review? How would you explain a declined risk that an agent scored?

Zurich's processing improvement and Lemonade's quote speed are real outcomes, but they're outputs of a governed, auditable foundation. The carriers that will struggle are the ones that deploy fast and document late. Regulators in this space are not slowing down, and the examination question isn't whether you used AI. It's whether you can show your work.