Everyone's sharing that MIT report about 95% of AI pilots failing. I get it—the hype cycle is exhausting. But here's what I learned building AI infrastructure for life and annuity carriers: The failures aren't about AI. They're about treating AI like it's 2010 software.
POCs fail because we're asking the wrong question. Stop asking 'Can AI do this?' Start asking 'Can we operationalize this?' The carriers who figure this out will capture their share of the $1.1 trillion McKinsey says is on the table.
Life and annuity carriers aren't retreating—they're doubling down. According to Wipro-Morningstar research, U.S. insurers plan to more than double AI's share of IT budgets from ~8% to ~20% in the next 3–5 years. LIMRA and EY project similar growth.
Yet most remain stuck in the same cycle: dozens of proofs of concept, very few production deployments.
The POC-to-Production Gap
Deloitte's 2024/25 survey found 76% of insurance executives have implemented GenAI somewhere—but most remain in pilot stage. The blockers? Poor data foundations, legacy IT, and weak business-tech alignment.
Here's the counterintuitive truth: These aren't bugs, they're features of our industry. Insurance runs on complexity—multi-state rules, decades-old systems, unstructured data. Any AI solution that doesn't account for this reality is dead on arrival.
POCs shine in controlled labs with clean data and simple use cases. But production means processing unstructured legacy data, orchestrating across multiple systems, and proving every decision with auditable lineage.
The Regulatory Reality Check
Regulators aren't waiting for us to figure this out.
NAIC's Model Bulletin from December 2023 demands comprehensive AI governance programs. As of March 2025, 24 states have already adopted it, with more following suit. FINRA's 2024-2025 guidance reminds firms that existing rules fully apply when using GenAI technologies. Their 2025 Report emphasizes implementing governance programs that identify risks and prohibited use cases.
State-level momentum is accelerating too. Colorado and New York have their own AI frameworks. Iowa became the first state to define "bias" in AI systems.
But here's what the doomsayers miss: MIT's own data shows companies using external AI partners succeed 67% of the time versus 33% for internal builds. Why? Because partners have already built the compliance layer. They've done the unglamorous work that POCs skip.
These regulations aren't barriers—they're your competitive moat. The carriers who build compliance into their AI DNA will move fastest while others scramble to retrofit.
And this brings us to why the traditional playbook keeps failing.
Why Traditional Approaches Fall Short
I've watched carriers try three paths, and they all lead to the same dead end:
DIY Labs: "We'll build it ourselves." Six months later, you've got data scientists creating brilliant models that your compliance team won't approve and your IT team can't deploy.
Point Solutions: "Let's start small." So you buy an underwriting bot. Then a claims tool. Now you've got three silos that don't talk to each other, can't share learnings, and each require separate compliance reviews.
Foundation Model Theater: "We've trained an insurance-specific LLM." Sounds impressive until GPT-5 drops, open models get 10x cheaper, and you're locked into yesterday's technology.
The cost reality check hits hard when you look at actual numbers.
Here's the reality: GPT-4o-mini costs 94% less than the original GPT-4. That's like your AWS bill dropping from $10,000 to $600 overnight. And it's still among the more expensive options. Open models like Llama and Mistral, available through platforms like IBM watsonx.ai, can cost even less when self-hosted or accessed through cloud providers.
The cost differential has reached a tipping point. For high-volume, routine tasks—which represent 80% of insurance AI use cases—you don't need premium models. Document classification, data extraction, basic Q&A—these work perfectly well with efficient models at a fraction of the cost.
But the real issue isn't price—it's lock-in. When you build on someone's proprietary "insurance-trained" model, you can't switch when better, cheaper models emerge. And they emerge every six months. You're betting your AI strategy on a depreciating asset.
Think of AI models like rental cars. Why buy a Ferrari when you need different vehicles for different trips? You want a truck for heavy document processing, a sedan for routine queries, and maybe that Ferrari for complex synthesis. That's model-agnostic architecture—use the right tool for each job, and switch whenever something better comes along.
The intelligence that actually matters isn't in the foundation model. It's in your document ingestion, your embeddings, your retrieval strategy, and your ability to route queries intelligently. Why run everything through GPT-4 when document processing works fine with Llama, query understanding uses Granite, and only final synthesis needs premium models?
So if the models aren't the bottleneck, what is?
The Missing Middle Layer
The bottleneck isn't model quality—GPT-4 is plenty smart. It's the operational layer between AI and your production systems.
What the industry needs:
Orchestration, not integration. A middleware layer that connects to existing systems through APIs and connectors. Policy admin systems stay, AI enhances them.
Shared intelligence with isolated execution. Common product knowledge and capabilities should be reusable, but each carrier needs their own isolated environment with separate data stores and audit trails.
Audit by design. Every AI decision must be traceable to its source—document, embedding, model version. Not just logs, but structured, queryable databases that satisfy regulators.
Reusable agents. Build once, deploy across channels. The same intelligence should serve advisors, portals, and customer service without rebuilding for each interface.
This operational layer is the difference between perpetual pilots and production AI. Most carriers haven't built it yet—which explains the 95% failure rate.
The Path Forward
Gartner says GenAI is entering the "trough of disillusionment." Good. That's when the theater ends and the real work begins.
The carriers investing 20% of IT budgets in AI will separate into two camps: those collecting POCs like trophies, and those building operational AI infrastructure.
The 95% failure rate isn't a verdict on AI—it's a verdict on our approach. While everyone debates whether we're in a bubble, the smart money is on the unglamorous stuff: middleware, orchestration, governance. The plumbing that makes AI work.
The question isn't whether AI will transform insurance. It's whether you'll be ready when it does.
