On AI, Insurance, and the Questions I Can't Set Down

I was recently approached to interview with a well-funded AI insurtech startup. Good funding, serious customers, compelling product. I'm a backend engineer actively moving from consulting into product roles, and this checked a lot of boxes.

But it's 2026. You can't evaluate an AI company without carrying the weight of what we've already built and what it's done.

The company's pitch is coherent. Insurance underwriting drowns in manual process — submissions reviewed in order of arrival, portfolios managed through spreadsheets and institutional memory. Their platform uses agentic AI to triage submissions, generate quotes, and align decisions to portfolio strategy in real time. Customers report an 89% reduction in time to quote. One company consolidated nine systems into one. The people building it seem genuinely motivated.

Insurance is one of the most consequential points of contact between institutions and ordinary people — the mechanism by which risk is either absorbed or pushed back onto those who can least afford it. The tools that preceded AI here already showed the problem. Credit scores as a proxy for risk. ZIP codes as a proxy for behavior. Pricing models that correlate with race and income without ever naming them. AI doesn't create that problem. It industrializes it, accelerates it, and makes it harder to audit.

This platform frames its value in terms of "high-appetite risk." The goal is to help carriers identify and bind more business that fits their portfolio strategy. That's a rational business objective. The person on the other end of the quote — the small business owner, the homeowner — is not in that frame. When the system gets better at identifying high-appetite risk, what happens to the people who don't qualify? Do they find coverage elsewhere at higher prices? Do they go uninsured? Does the optimization of one portfolio aggregate, across the industry, into a coverage gap that falls disproportionately on people with less bargaining power?

I don't know. I'm not sure the company does either. I'm not sure anyone is tracking it.

The companies that built dominant social platforms weren't trying to fracture political epistemology or damage teenage mental health. They were trying to connect people, and they were very good at it. The harm came from scaling a system optimized for one outcome — engagement — without building in the friction to ask what else the system was doing, and at whose cost. The intention was fine. The architecture was the problem. And the architecture got locked in before the consequences were legible.

I don't see that conversation happening in insurtech right now. I see efficiency metrics and customer testimonials and a compelling narrative about replacing legacy systems that genuinely aren't serving anyone well. What I don't see is serious public engagement with what the platform optimizes for when it's working exactly as designed — and who absorbs the cost of that optimization.

There's a labor question worth naming too. The jobs this technology is replacing aren't warehouse jobs. They're underwriters — knowledge workers with years of domain expertise. The AI maturity model this company describes culminates in a system that automatically quotes deals within predefined guardrails, with underwriters auditing randomly for quality control. That's not augmentation. That's replacement with an oversight layer on top.

I'm still in the process. I don't know where I'll land. But I think these questions are worth asking out loud — because I suspect I'm not the only engineer who has found themselves here, interested in the opportunity and uncertain about the territory.