Insurance AI Talent: Build vs Buy vs Borrow vs Bypass

Six Ways to Staff the AI Engine

Insurers face a brutal choice: who builds and runs the AI systems that will decide underwriting wins, claims outcomes, and customer retention. The talent scarcity is real. McKinsey estimates 250,000 additional data scientists will be needed by 2026 just to meet insurance demand, and 60% of insurers rank “AI/ML talent” as their top hiring headache.

Below are the six tactical routes carriers take, with trade-offs spelled out. I’ve seen claims teams burn $2M on a custom fraud model that never shipped because the build team couldn’t translate actuarial loss-ratio logic into Python. Others have outsourced core underwriting models to a vendor, only to discover the vendor’s API latency added 400ms to quote time—costing a top-10 carrier 7% click-through on direct channels. Choose wisely.

Where Each Option Wins—and Bleeds

Build is the only route that gives you full ownership of the model and the data pipeline. I’ve seen carriers like Lemonade and Hippo use greenfield teams to shave 3–5 points off loss ratio on renters and auto, respectively. But the burn rate is brutal. A mid-tier MGA told me they spent $14M hiring 45 PhDs, only to see 18 leave within 22 months—mostly for Big Tech. Now they’re stuck with undifferentiated models that cost $4M/year just to maintain.

Buy is safe but slow. Guidewire’s AI suite clocks in at 18 months from contract to production-grade model for a Tier-1 carrier. The vendor’s pre-built peril models work, but they’re calibrated on U.S. data; drop them into a Thai flood book and the false-positive rate jumps to 22%. Worse, every upgrade triggers a regression test cycle that can stall new product launches for weeks. I know one mutual that paid $2.3M in exit fees after switching from Duck Creek to Earnix because the combined ratio on their commercial auto book jumped 2 points during migration.

Borrow lets you punt the talent problem to someone else’s balance sheet. Deloitte’s AI Garage claims a 70% faster time-to-market, but the fine print is brutal. A Tier-2 carrier I advise paid $11M for a claims triage model; the model worked for six months, then Deloitte pivoted the team to a different client. The carrier inherited a codebase with 60% dead features and a 400ms inference latency that now eats 8% of their STP budget. The NOC fee alone is $450k/year—more than their entire internal DevOps spend.

Hybrid splits the baby. USAA’s 2022 annual report shows they’ve moved 40% of their claims AI to a managed service while keeping pricing models in-house. The glue code—the API layer that routes claims to the vendor and back—now represents 28% of their AI engineering sprints. Still, the combined ratio on auto claims fell from 68.4% to 65.1% within 18 months. The trade-off? Any change to the glue layer requires a full regression suite that can take 6 weeks, so innovation stalls when the vendor changes their schema.

Bypass is the insurgent’s weapon. Hippo’s 2023 10-K mentions “no-code automation” saved $8M in labor costs across FNOL processing. But skip this if you write E&S; or specialty lines; the false-negative rate on a cyber breach notification form can exceed 30%. One regional carrier tried to use Azure AI Builder to auto-decline workers’ comp claims—only to trigger a state regulatory fine for “unfair claims practices.” The ROI vanished overnight.

Acqui-hire is a gamble. Admiral Group’s 2019 acquisition of price-comparison engine Compare.com for £220M was pitched as a talent play, but the team spent 18 months re-architecting Admiral’s pricing engine instead of building new models. By the time they shipped, the combined ratio on their motor book had risen 1.1 points due to stale rating factors. Meanwhile, 42% of the new hires left for FAANG roles where stock refreshes beat Admiral’s RSUs.

Decision Matrix: Pick Based on Your Runway

Use the table below to shortlist. I’ve weighted speed (30%), cost (25%), control (25%), and risk (20%).

One Rule That Cuts Through the Noise

If your combined ratio is <65% and you can fund a greenfield squad for 24 months, build. Anything above 70% combined ratio and you can’t afford the distraction—buy or borrow instead.

For everyone else, hybrid is the least-worst compromise. But watch the glue. One carrier I advised spent $6M on a feature store integration layer that became a single point of failure during the 2023 wildfire season; the API latency spiked to 1.2 seconds, and their direct channel quote-to-bind time doubled. The model itself was fine; the integration was the weak link.