Insurance AI Talent: Build vs Buy vs Borrow vs Blend
Over the past three years, I’ve watched claims teams, underwriting desks, and actuarial departments rush to embed AI into workflows. The biggest bottleneck isn’t the tech—it’s the people. Insurers can’t hire fast enough, and the talent that does exist costs what feels like actuarial prices.
That’s why the “build vs buy” debate no longer applies. Today, carriers must choose among a menu of talent strategies: build in-house, buy via third-party AI vendors, borrow from academia or consortia, or blend multiple approaches. Each path carries different risks, timelines, and price tags.
Six AI Talent Acquisition Strategies Compared
| Strategy | Speed to Deploy | Control & IP | Cost (3-Year TCO) | Risk | Best For |
|---|---|---|---|---|---|
| 1. Build In-House (Greenfield Lab) | 12–24 months | Full ownership of models and data pipelines | $8M–$20M (salaries, GPUs, ops) | High execution risk—attrition, culture clash, model drift | Large P&C carriers with >$10B GWP and mature data engineering |
| 2. Acqui-Hire (Roll-Up) | 3–9 months | Full IP transfer; talent stays 1–2 years post-close | $30M–$150M (earn-outs included) | Integration risk—cultural mismatch, retention drop-off | Tier-1 insurers needing domain-specific AI fast (e.g., climate risk) |
| 3. Buy Off-the-Shelf (Vendor SaaS) | 4–8 weeks | Limited—vendor owns model weights; carrier owns outputs | $500K–$3M/year (per line of business) | Vendor lock-in, black-box models, rising subscription fees | Mid-market carriers with <$5B GWP and tight budgets |
| 4. Borrow (Academic/Industry Consortia) | 6–12 months | Shared IP; carrier gets non-exclusive license | $1M–$5M (consortium fees + internal integration) | Slow governance cycles, IP dilution, limited exclusivity | Specialty lines (cyber, parametric) or regional mutuals |
| 5. Blend (Hybrid MGA/TPA Partnership) | 6–18 months | Carrier owns data; partner owns model; co-developed IP | $2M–$8M (shared dev cost, revenue split 60/40) | Governance complexity, data-sharing friction, profit-share disputes | MGAs expanding into AI-enabled underwriting without heavy CapEx |
| 6. Buy Talent as a Service (TaaS—Contractors + Platform) | 4–12 weeks | Carrier owns outputs; contractors own model code | $200K–$1.5M/year (variable headcount) | Knowledge leakage risk, contractor churn, compliance gaps |
Strategy 1: Build In-House (Greenfield Lab)
When I talk to CIOs at carriers like Allstate or Liberty Mutual, they describe building AI as a “moonshot.” They’re right. A dedicated AI lab with 50–100 full-time staff—data scientists, ML engineers, and domain actuaries—can take 18 months to deliver a production-grade underwriting model that reduces loss ratio by 3–5%. But most labs don’t hit that target. The attrition rate for AI talent in insurance hovers around 25% per year, and the ones who stay often jump to fintech or big tech for 30–50% higher compensation.
The real killer is model drift. A model trained on 2021 claims data may perform poorly in 2024 due to inflation, climate events, or regulatory shifts. Maintaining it requires a dedicated team of five to eight engineers—an ongoing cost most carriers underestimate. At one insurer I worked with, the annual cost of model drift remediation reached $1.2M.
Trade-off: Full control comes with full operational drag. You own the tech debt and the talent churn.
When to choose: Only if you have $10B+ in premium, a mature data platform, and a board willing to fund a multi-year transformation. Even then, most carriers hedge by pairing in-house labs with vendor partnerships.
---Strategy 2: Acqui-Hire (Roll-Up)
This is the nuclear option. In 2022, Hiscox acquired Codec.ai, a climate risk modeling startup, for $45M. Goal: embed parametric trigger models for wildfire and flood. Within six months, Hiscox launched a parametric product in California—something it couldn’t have built internally in under two years.
The catch? Codec’s team stayed for 18 months. After that, attrition spiked. Three senior data scientists left for Hippo (now at $150K base + equity). Hiscox had to backfill at a 40% salary premium.
The financial risk is even worse. A 2023 Oliver Wyman study found that 68% of AI acqui-hires in insurance fail to deliver the projected ROI within 24 months. The main culprit: integration of legacy systems. Most insurers use 10+ core systems (policy admin, claims, billing). Merging a startup’s cloud-native stack with on-prem mainframes is a nightmare.
Trade-off: Speed vs. cultural suicide. You get talent, but you also get their habits—and their exit plans.
When to choose: If you need a specific AI capability immediately (e.g., AI-driven fraud detection for workers’ comp) and have the budget to absorb risk. Not for incremental improvements.
---Strategy 3: Buy Off-the-Shelf (Vendor SaaS)
For mid-market carriers like Kin Insurance or PURE Group, buying AI from vendors like Ghost Automations or Drishti is the pragmatic path. You can deploy a claims triage bot in six weeks and reduce adjuster workload by 30%. The cost? $1.2M/year for a 50,000-policy book.
But here’s the dirty secret: most vendors don’t improve over time. In 2023, I reviewed 12 AI claims platforms. Only two had model refresh cycles faster than annually. The rest relied on outdated training data, leading to a 15% false-positive rate on bodily injury claims. Carriers end up paying for “AI” that’s no better than a rules engine.
Another risk: subscription creep. A vendor might start at $200K/year for auto damage assessment, then upsell to $1.5M/year when you want to add subrogation prediction. By year three, your AI budget is 300% higher than projected.
Trade-off: Speed and predictability, but vendor dependency and rising TCO.
When to choose: If your combined ratio is already tight (<100) and you lack AI talent. Also ideal for lines where regulatory risk is high (e.g., life insurance underwriting).
---Strategy 4: Borrow (Academic/Industry Consortia)
Consortia like The Insurance AI Lab (backed by Swiss Re and Munich Re) or the IAA’s AI working group offer a middle path. You get access to cutting-edge models without hiring a full team. For $2M over three years, a regional carrier can license a climate peril model developed at ETH Zurich.
The downside? You’re sharing the IP. In 2023, a mutual insurer in the Midwest used a consortia model to predict hail damage. When it tried to commercialize a spin-off product, the consortium blocked it—citing IP sharing agreements. The carrier had to pivot to a vendor model instead.
Consortia also move slowly. The average governance cycle is 9–12 months, which doesn’t align with the quarterly cadence of most insurers. If you need to adjust a parametric trigger for a new flood zone, you’ll wait until the next consortium meeting.
Trade-off: Shared innovation at the cost of exclusivity and speed.
When to choose: If you’re in a specialized line (e.g., crop insurance) or a mutual with limited budget. Also useful for R&D-heavy carriers like Aviva testing climate models.
---Strategy 5: Blend (Hybrid MGA/TPA Partnership)
This is the strategy that’s quietly reshaping specialty lines. Take WeatherMod, an MGA focused on hail damage. It partners with Lemonade to access Lemonade’s AI claims engine, but keeps its own underwriting data private. The result: WeatherMod can underwrite hail risk with a 20% lower loss ratio, while Lemonade gets a new distribution channel.
The partnership model works because it splits risk. The MGA handles distribution and customer acquisition, while the carrier (or insurtech) provides the AI engine. In a recent deal, The Hartford partnered with an AI TPA to automate workers’ comp claims. The TPA’s model reduced indemnity payments by 12%, but The Hartford had to share 15% of premium savings with the TPA.
The biggest friction point? Data sharing. Carriers are reluctant to hand over bordereaux and loss runs to third parties. Many partnerships collapse over NDAs and data-use clauses. I’ve seen deals fall apart when a carrier demanded ownership of the final model—something the TPA refused.
Trade-off: Speed and cost-sharing, but governance nightmares.
When to choose: If you’re an MGA expanding into AI-enabled underwriting or a carrier looking to pilot AI in a specific line without heavy CapEx. Especially effective for cyber, parametric, or niche commercial lines.
---Strategy 6: Buy Talent as a Service (TaaS)
This is the Wild West of AI talent strategy. Platforms like Toptal, Upwork, or specialized insurtech contractors (e.g., Insurtech.ai) let you spin up a team of 10 contractors in two weeks. Cost: $200K/year for a senior ML engineer ($120/hr) plus $50K/month for a platform license.
But here’s the problem: contractors don’t build for longevity. They optimize for the project at hand. I’ve seen contractors build a fraud detection model in 90 days, then leave—leaving the carrier with no documentation and a model that fails every quarterly audit. The carrier had to rebuild it internally at $300K.
Another risk: compliance. Contractors often lack understanding of insurance regulations (e.g., NAIC model laws, GDPR). One insurer I worked with had a contractor build an AI claims model that accidentally violated New York’s Regulation 187, resulting in a $2.3M fine.
Trade-off: Flexibility at the cost of institutional knowledge and compliance risk.
When to choose: For short-term pilots, crisis response (e.g., post-Hurricane Ian claims surge), or when you need niche skills (e.g., LLM fine-tuning) without long-term commitment.
---Which Strategy Wins—and When
I’ve seen carriers pick the wrong strategy and pay the price. Here’s my rule of thumb:
- If you’re a Tier-1 carrier ($10B+ GWP) with a mature data platform and a board willing to fund a multi-year transformation: Build in-house, but pair it with vendor partnerships to mitigate model drift. Allstate and Liberty Mutual are doing this—with mixed success. Expect 20% attrition in the first year.
- If you need a specific AI capability ASAP (e.g., AI-driven fraud detection for workers’ comp) and have the budget to absorb risk: Acqui-hire. But only if you can integrate the team quickly. Hiscox’s Codec.ai deal worked because the target was small and focused. A sprawling AI platform acquisition? Disaster.
- If your combined ratio is already tight (<100) and you lack AI talent: Buy off-the-shelf. But negotiate hard on model refresh cycles and data ownership. Lemonade’s success with its claims engine proves this works for greenfield carriers, but mid-market incumbents often get locked into bad contracts.
- If you’re a mutual or regional carrier with limited budget: Borrow via consortia. But only for R&D-heavy lines (e.g., climate risk). Don’t expect commercial-grade models.
- If you’re an MGA or TPA expanding into AI-enabled underwriting: Blend. Partner with a carrier or insurtech to share costs and risk. WeatherMod’s hail model is a textbook example. But choose your partner carefully—data-sharing disputes sink more deals than bad models.
- If you need a short-term pilot or crisis response: Buy talent as a service. But treat it as a stopgap, not a strategy. Contractors will leave, and the model will rot.
Final Verdict: The Blend Strategy Is the New Build vs Buy
There’s no universal winner. The right strategy depends on your scale, budget, and risk tolerance. But one trend is clear: the best carriers are blending multiple approaches.
For example, Chubb builds its own AI models for high-value commercial lines (e.g., D&O underwriting) but buys off-the-shelf for personal auto claims. Meanwhile, Progressive relies on in-house talent for pricing models but partners with MGAs for niche products like rideshare insurance.
The key is to treat AI talent like venture capital: diversify your bets. Build where you need control, buy where you need speed, borrow where you need expertise, and blend where you need flexibility.
And remember: no strategy eliminates the talent war. The real battleground isn’t the tech—it’s the people who build it.