Corgi launched a dedicated AI liability product on May 5, 2026 and closed a $160 million Series B at a $1.3 billion valuation the next day, roughly four months after a Series A that valued it at $630 million.

The product covers hallucination, algorithmic bias, training data disputes, adversarial attacks, synthetic media and autonomous system failure. None of those perils has a credible loss triangle behind it, which is the pricing problem the capital is betting on.

Key Takeaways

  • A $160 million Series B at a $1.3 billion valuation, led by TCV with 18 additional investors, taking total funding above $268 million and doubling the $630 million Series A mark in about four months.
  • Corgi writes on its own paper, licensed since July 2025, so its actuarial team sets reserve adjustments directly rather than negotiating them with a fronting partner when AI development surprises.
  • Severity benchmarks already run to ten figures. Anthropic settled copyright claims for $1.5 billion and Universal Music filed a $3.1 billion suit in January 2026, against 164 active AI copyright cases tracked in 2026.
  • Correlation, not frequency, is the rating-plan problem. An agent with 85% per-step reliability achieves only 20% end-to-end success across a ten-step workflow, so pricing modules independently and summing understates the aggregate.
  • The demand catalyst is a contested deadline. The EU AI Act's August 2, 2026 high-risk obligations remain legally in force after the second political trilogue failed on April 28, 2026, but a deferral to December 2027 is still on the table.

What Was Actually Launched

Corgi is a licensed carrier rather than a managing general agent, approved in July 2025 and writing on its own paper. That is the structural detail with actuarial consequences: underwriting authority, claims adjudication and reserve methodology sit in one place, so when AI liability development moves, the reserve adjustment does not require a capacity provider's agreement.

The AI coverage is a modular add-on to the existing Technology Errors and Omissions policy rather than a standalone form, using affirmative language inside the E&O structure across six categories: model performance and hallucination, algorithmic bias, training data disputes, adversarial attacks and model theft, synthetic media liability, and autonomous system failure.

The $160 million round was led by TCV with 18 additional investors, bringing total funding above $268 million against $40 million of annual recurring revenue in the first year after licensing. The valuation doubled from $630 million in roughly four months.

It enters a market that is splitting rather than growing. Verisk's ISO generative AI exclusion endorsements took effect January 1, 2026, with most state approvals clearing in thirty to sixty days, so traditional carriers are removing the exposure from commercial general liability at the same time specialty markets write it affirmatively. Legacy carriers stepping back concentrates the risk in a small set of writers with no loss history to calibrate against.

Pricing a Peril Class With No Triangles

Generative AI in commercial form is about three years old and enterprise deployment is younger, so the reporting cycle has not produced development to build paid or incurred triangles from. Better data collection does not fix that in the near term.

The signal data that does exist sets the severity range rather than the frequency.

Loss Category Data Point Source
Hallucinated legal citations 125+ filings with fabricated case references identified in early 2025; courts imposed $10,000+ sanctions in at least 5 cases ComplianceHub, court records
AI copyright litigation 164+ active AI copyright cases tracked as of 2026 AI Lawsuit Tracker
Training data IP settlements Anthropic: $1.5B settlement; Universal Music: $3.1B lawsuit filed January 2026 Court filings
Enterprise AI failure losses 64% of companies with $1B+ revenue report $1M+ losses from AI failures EY survey 2025
Medical device AI recalls 1,357 FDA-authorized AI devices; 60 involved in 182 recalls FDA database
AI defamation claims Multiple lawsuits filed against Meta, OpenAI for chatbot-generated false statements about named individuals Court filings, Damien Charlotin database

The exposure base is where the pricing work starts, because the traditional ones do not discriminate. A $10 million revenue company routing all customer interaction through an agentic chatbot and a same-revenue company using AI for internal document summarization carry the same payroll and the same premium under a conventional base.

The alternatives being tested measure the deployment instead: inference volume, the split between customer-facing and internal-only surfaces, model capability tier benchmarked against published hallucination leaderboards, a regulated-industry multiplier for healthcare, financial services and legal, and governance attestation scaling the load factor on the insured's model inventory and human review practice.

Correlation is the harder half, and it is quantified. An AI agent with 85% per-step reliability achieves only 20% end-to-end success across a ten-step workflow, so failure probability compounds far faster than a linear exposure model predicts. An insured running autonomous agents in a regulated industry is exposed simultaneously to hallucination, bias and autonomous failure claims, and Corgi's self-service module selection means the rating plan has to price each module while carrying that dependence. Summing independent module rates prices the portfolio short.

The cyber precedent sets expectations for what happens next. AIG wrote the first internet security liability policy in 1997, and for a decade carriers priced cyber on competitor benchmarking and judgment rather than credibility-weighted experience. The line reached $15.3 billion of global premium by 2024, but the route ran through a repricing cycle in which premiums rose more than 30% annually between 2020 and 2022 as ransomware losses overwhelmed the initial rate assumptions.

The American Academy of Actuaries has documented that history, and the Chicago Fed identified the absence of standardized actuarial tables as the persistent structural gap. AI failure modes are less well defined than a data breach, so the same gap is wider here.

The Demand Case Rests on a Deadline That Is Still Moving

Organic demand for AI liability would build slowly. What accelerates it is regulation, and the regulation is not settled.

Annex III of the EU AI Act classifies AI systems used for risk assessment and pricing in life and health insurance as high-risk, triggering obligations under Articles 9 through 15 covering risk management, data governance, technical documentation, human oversight and conformity assessment. The compliance date is August 2, 2026.

The European Commission proposed deferring it to December 2027 through the Digital Omnibus on AI. The second political trilogue failed on April 28, 2026, which leaves August 2 legally in force without making it certain. An insurance line whose growth thesis rests on a compliance deadline is exposed to that deadline moving, and a deferral to December 2027 removes sixteen months of demand from the pricing assumption.

There is a second-order effect inside the compliance itself. Article 10(5) creates an exception allowing processors to use special category data for bias testing, which is a sensible provision that also generates a documented record of the results. Where that record shows disparate impact, it is subpoenable, so the act of complying manufactures evidence for the bias claims the coverage responds to.

Domestically the picture is steadier but earlier. The NAIC Model Bulletin has been adopted by more than half of US states since December 2023, and the AI Systems Evaluation Tool is in a twelve-state pilot running through September 2026. Grant Thornton's 2026 survey found only 24% of insurers confident of passing an independent AI governance review, which is a demand signal and an underwriting problem at once: the same governance deficit that makes the coverage attractive is what makes the insured's loss more likely.

Further Reading

Sources