The underwriting gap — where the insurance models haven't arrived yet

The insurance landscape for Physical AI becomes clearer when you map it on two axes: coverage availability and model maturity.

Automotive sits in the top-right: high coverage, mature models. Car insurance is mandatory in most jurisdictions. Decades of actuarial data. Telematics-based pricing using real-time driving behavior. Progressive wrote over $6 billion in net premiums in a single month in 2025. The models are deep, the data is rich, the products are everywhere.

Traditional machinery sits in the top-left: coverage exists but models are outdated. Commercial General Liability (CGL) policies have covered manufacturing equipment for decades. There's actuarial data from thousands of incidents. But the models were built for deterministic machines — PLCs, relays, hydraulic systems. They don't account for AI-based failure modes: neural network misclassification, adversarial inputs, emergent fleet behaviors.

Cyber insurance sits in the bottom-right: coverage is low relative to the risk surface, but models are maturing rapidly. Global cyber premiums reached an estimated $16.5 billion in 2025 and are projected to exceed $42 billion by 2030. Carriers are investing heavily in AI-driven underwriting and real-time threat assessment. It's a young market, but it's growing faster than almost any other insurance category.

Physical AI sits in the bottom-left: low coverage, immature models. That quadrant finally moved in October 2025 — but not far. China Pacific launched the first policy written exclusively for humanoid robots, with Ping An and PICC following within weeks. Axis Insurance's autonomous robotics programme began covering AI perception failures, control-takeover attacks, and production losses when a software update takes a fleet offline. Zurich and YAS added embedded robotics cover in Hong Kong in June 2026. What none of them has is depth: no decades of actuarial history, no published failure-rate baselines, no standardised diagnostic-coverage reporting to price against. These are first movers writing policy ahead of the data — a market opening, not a market that works. Most commercial cover available today is still an endorsement on existing machinery coverage that doesn't address the new failure categories.

And the movement runs both ways. While specialty carriers wrote the first robot policies, mainstream carriers moved to write the risk out: Berkshire Hathaway, Chubb and Travelers filed to exclude AI-related liability from general liability policies, and regulators approved more than 80% of those requests. The specialty market is opening a door at the same rate the standard market is closing one.

That bottom-left quadrant is where the gap lives. And it's not a small gap. EY noted in 2025 that humanoid robots will fundamentally reshape commercial insurance. Accenture found that 71% of insurance executives envision deploying autonomous mobile robots within 5-10 years. QED Investors argued that a robot incapable of being insured cannot exist outside a controlled environment.

The industry sees the market. It cannot yet price it. Pricing requires data — failure rates, diagnostic coverage ratios, incident causation patterns, mean time to dangerous failure — that the Physical AI industry hasn't generated yet, because the safety infrastructure that would produce it is nowhere near installed at scale. Even NVIDIA's Halos, shipping since June 2026, is still early-access on a handful of platforms.

The underwriting gap is not just an insurance problem. It's a deployment bottleneck. Until robots can be insured at reasonable premiums, operators face unquantified liability. Until operators face quantified liability, safety investment looks like a cost rather than a prerequisite. Breaking this cycle requires the safety data layer that neither the insurance industry nor the robotics industry has built yet.