Who Pays When the Agent Is Wrong, Part Two

October 6–7, 2026

Anthropic has introduced Claude for Google Workspace in public beta, bringing a Claude sidebar into Google Docs, Sheets, and Slides for Pro, Max, Team, and Enterprise subscribers. Users can work directly on open files rather than copy content between applications, while separate beta connectors let Claude create and edit Google files from its own chat interface. Capabilities include drafting and restructuring documents, creating spreadsheet formulas and pivot tables, and building slides that follow an existing deck’s layouts. For enterprise buyers, the significance goes beyond convenience: this gives organizations another AI option inside Google’s productivity environment, weakening the assumption that their productivity-suite vendor must also supply their primary assistant.

Key takeaways:

October 5, 2026

Cohere launched North 2, updating its enterprise agent platform with a redesigned orchestration harness, cross-session memory, and granular administrative controls. According to the company, agents can retain context between sessions and draw on shared knowledge and assets, while administrators can manage user quotas, rate limits, organization-wide spending caps, model access, and activity monitoring. The platform supports cloud, on-premises, and air-gapped deployments. The important enterprise signal is the combination: useful digital workers need continuity, but persistent agents also need explicit economic and administrative boundaries. North 2 positions those controls as core platform capabilities rather than problems customers must solve after deployment.

Key takeaways:

October 5–6, 2026

VentureBeat’s latest reporting highlights a consequential gap between enterprise data-governance investments and the information agents actually use. In its VB Pulse survey, roughly two-thirds of companies were building governed semantic layers, yet only 20% of that group said those layers were their agents’ main source of business context. Separately, 64% of August respondents reported tracing a confidently wrong agent answer to missing or inconsistent business context during the previous six months. A semantic layer provides agreed definitions for business terms and metrics; querying raw tables or loading more information into a prompt does not automatically preserve those definitions. The digital-workforce implication is straightforward: governance infrastructure delivers little protection if agents do not use it in their execution paths.

Key takeaways:

Insurance usually trails risk by years. This time it's keeping pace with it, and that alone is worth noting. At least six specialist carriers built and launched agent-liability products in 2026: Klaimee, a Y Combinator-backed startup offering self-serve coverage for first- and third-party agent losses; Testudo, a Lloyd's-backed managing general agent targeting U.S. mid-market enterprises; HSB, the Munich Re subsidiary that introduced AI liability coverage for small and mid-sized businesses in March; Mayflower Specialty and Hadron, bundling AI exposure into existing D&O, employment practices, and errors and omissions coverage; Agent Insured, entering the EU market in the third quarter; and Corgi, packaging AI liability alongside cyber and D&O coverage for startups.

That's not a niche experiment. It's a specialty insurance category forming from scratch inside a single calendar year, which is itself a signal about how real the underlying demand has become. "Who Pays for Autonomy" covered the vendor-side version of this question, the war between seat-based and outcome-based pricing. Insurance is the other half of the same conversation: not what an agent costs to run, but what it costs when it's wrong, and who absorbs that cost when it happens.

Here's the part that complicates the story. While specialty insurers were building new products to cover AI liability, the broader commercial insurance market was moving in the opposite direction. W.R. Berkley confirmed it's now writing "absolute AI exclusions" into its policies, language that removes coverage for "any actual or alleged use, deployment, or development of AI" entirely. Verisk's ISO division, whose policy language underpins much of the commercial insurance industry, released three generative AI exclusion endorsements, CG 40 47, CG 40 48, and CG 35 08, effective January 1 of this year, carving AI-related harm out of standard general liability coverage by default.

So the market isn't converging on a single answer. It's splitting. Standard commercial policies are actively excluding AI exposure at the same moment specialty carriers are racing to sell narrower, purpose-built coverage for exactly that exposure. For a buyer, that means the baseline assumption has to flip: AI liability isn't covered unless you've gone out and bought a product specifically built for it, and most organizations haven't yet realized their existing general liability policy quietly stopped covering it sometime around New Year's Day.

The numbers on the uncovered side are already substantial. One in five commercial insurers reported an AI-related loss in 2025, and only about half of those losses were fully covered under existing policies, according to research cited by law firm Honigman. Intellectual property and copyright claims now make up more than 23% of generative AI litigation, and the broader litigation count tells its own story: U.S. generative AI lawsuits exceeded 700 by early 2025, up nearly tenfold since 2021.

Set that against the trajectory "The Liability Question" documented, Gartner's projection of more than 2,000 legal claims against enterprises tied to AI agent harm by the end of this year, and the shape of the problem comes into focus. Claims volume is climbing fast. Coverage is arriving more slowly, and where it exists, it's covering roughly half of what's being lost. That gap doesn't close itself, and until it does, every enterprise running agents in production is self-insuring the difference whether it intended to or not.

It's worth being specific about what these new products cover, because the coverage is narrower than the marketing suggests. HSB's product, confirmed for small and mid-sized businesses, covers bodily injury, property damage, and personal and advertising injury tied to AI use, filling a real gap in standard general liability policies that exclude AI-generated advertising, marketing, and social content. Timothy Zeilman, HSB's global head of product ownership, framed it plainly: "AI insurance helps remove that uncertainty by filling the gaps in coverage, so businesses can stay ahead of emerging risks." Munich Re's separate aiSure program targets higher-limit coverage for AI vendors themselves rather than the enterprises deploying their tools. Testudo and Klaimee are building toward broader first- and third-party agent-loss coverage, though neither has published premium figures or coverage limits publicly yet, which is itself worth noting: pricing transparency usually follows claims history, and claims history is exactly what this market doesn't have.

What's still largely missing is coverage for what one industry guide calls "autonomous-agent torts": harm caused by an agent acting on its own initiative rather than through a human decision a traditional general liability policy was built to cover. Standard commercial general liability policies assume a human caused the harm. An agent that deletes a database or executes an unauthorized transaction on its own doesn't fit that assumption cleanly, and most of this year's new products are still built closer to the general liability model than to a true agent-autonomy model. The gap "The Liability Question" identified at the legal standard level, reasonable oversight asking whether a human could have intervened, is the same gap the insurance market hasn't finished building a product around yet.

This is the question at the center of the issue, and the honest answer is: not with confidence, not yet. The methodology gaining traction is something researchers are calling trace-economic underwriting, pricing a policy based on what a specific agent did in a specific sequence of actions rather than a flat assessment of what kind of agent it is. Two 2026 papers testing this approach against synthetic loss scenarios found it reduced pricing error dramatically compared to flat-rate models, from roughly 17,700 down to 569 arbitrary loss units in testbed conditions. That's a meaningful result, but it's a result from a controlled simulation, not from priced and paid claims in the real market. No insurer in this space has published actual claims or payout data yet, because the policies are too new and the loss history that would let an actuary price with real confidence doesn't exist.

That's not a criticism of the insurers building these products. It's the same problem agent governance faces everywhere else in this series: the infrastructure to measure risk accurately, whether that's continuous monitoring, a named accountable owner, or now an actuarial model, has to exist before the risk can be priced with precision, and right now most of that infrastructure is still being built in parallel with the products meant to depend on it.

There's a complication specific to this market that most general liability underwriting has never had to deal with: the vendor whose agent caused the loss might not exist by the time the claim gets resolved. "The Vendor Consolidation Wave" covered Gartner's estimate that only about 130 legitimate agentic AI vendors exist out of the thousands claiming the category, with more than 40% of agentic AI projects projected for cancellation. An insurer pricing a policy against an agent built on a specific vendor's platform is implicitly pricing the risk that the vendor itself survives long enough to patch the flaw that caused the claim, provide the forensic data needed to establish what happened, or simply still be a going concern when the claim is filed. Vendor instability isn't a side issue for this market. It's baked into the underwriting risk itself, in a way traditional liability insurance for software never had to account for.

The governance side of this is getting sharper scrutiny too. Delaware's McDonald's derivative litigation this year extended officer oversight duties to cover AI governance decisions specifically, and the SEC's 2024 enforcement actions against Delphia Asset Management and Global Predictions for overstated AI claims, sometimes called AI-washing, show regulators are already treating inflated AI capability claims as a disclosure problem with real teeth. Both trends point toward D&O policies, not just general liability, becoming a second front in this same coverage question, and several of this year's new products, Mayflower Specialty and Hadron's bundled approach especially, are already being built with that in mind.

There's a quieter implication worth drawing out. "The Continuous Audit" and "The Single Point of Accountability," the first two issues in this series, argued that real-time monitoring and a named owner are the infrastructure Level 5 maturity requires. That same infrastructure is exactly what an underwriter needs to price this risk with anything better than a flat guess. An organization that can document continuous monitoring and a clear chain of accountability for every agent in production isn't just closer to defensible under the reasonable oversight standard. It's a materially better underwriting risk, and it's a reasonable bet that as this market matures, premiums will start reflecting that difference directly, the same way cyber insurance premiums began rewarding documented security controls once that market matured past its own early, undifferentiated pricing.

That creates a genuine incentive alignment this series hasn't had until now. The governance work this series has been arguing for isn't just a compliance cost or a liability shield. It's about to become a line item insurers can see and price, which means the organizations doing the work early will likely pay less for the same coverage than the ones that wait until a claim forces the question.

Six insurers built agent-liability products in a single year while the standard commercial market simultaneously excluded that same risk from its default policies. That's not a mature market reaching equilibrium. It's two halves of the insurance industry moving in opposite directions at once, leaving a gap that one in five commercial insurers has already fallen into at least once. The pricing methodology that might eventually close that gap, trace-economic underwriting, shows promise in testbed research but hasn't been proven against real claims, because the real claims history this market needs doesn't exist yet.

None of that means the answer is to wait. It means the organizations that build real-time monitoring and named accountability now, the infrastructure this series has argued for from its first two issues, will be the ones positioned to buy this coverage on the best terms once the actuarial models catch up to the risk. Everyone else will be self-insuring the gap in the meantime, whether they've acknowledged that or not.

Understanding exactly where your organization's uninsured AI exposure sits is the first step toward closing it, whether through better governance, the right coverage, or both. The Complete Agentic AI Readiness Assessment includes a framework for mapping your current liability exposure against what today's insurance products cover, so you can see the gap clearly instead of discovering it through a claim. Get your copy on Amazon or learn more at yourdigitalworkforce.com. For organizations evaluating agent liability coverage or preparing to present AI risk to underwriters, our AI Blueprint consulting helps build the monitoring, ownership, and documentation case that turns a harder-to-price risk into a defensible, insurable one.