What the industry is saying
Since 1 January
2026, insurers have been able to attach three Verisk ISO endorsements to
commercial general liability renewals.
CG 40 47
removes cover for bodily injury, property damage and personal and advertising
injury arising from generative AI. CG 40 48 removes personal and advertising
injury only.
CG 35 08
applies the exclusion to products and completed operations. Verisk has
confirmed it is now evaluating options for agentic AI.
Carriers are
going further. AIG, Great American and WR Berkley have sought permission from
US regulators to offer policies that exclude AI-related liabilities. Berkley's
absolute exclusion for D&O, E&O and fiduciary policies reaches any use
or development of AI.
The reasoning
is stated openly. Gallagher Re counts a 978% rise in AI-related lawsuits
between 2021 and 2025. Verisk's Joe Lam says that without
exclusions, underwriters "might
just walk away from the risk".
The challenges and constraints
·
No case law yet. There is no bellwether
ruling on how courts will read AI exclusions, so neither insurer nor
policyholder knows what the wording means in a disputed claim.
·
Definitions decide who pays. The ISO
forms define generative AI. A system that is AI but not generative may sit
outside the exclusion and inside the cover.
·
Accountability stays with the insurer. The
NAIC model bulletin, adopted in more than half of US states, holds insurers
responsible for AI decisions, including systems bought from vendors. California
SB 1120 requires a physician, not an algorithm, to make medical necessity
denials.
·
Insurers are scaling the same exposure. Allianz's
Project Nemo runs seven AI agents across the claims workflow. Insurers are
excluding AI risk for customers while building it into their own decisions,
where there is usually no one to pass the loss to.
Select real-life cases that show where the risk
lands and who owns it
1. UnitedHealth: the insurer's own algorithm in
court
What happened. Families
allege that the nH Predict tool led to premature denials of post-acute care for
Medicare Advantage members.
On 9 March 2026
a federal judge ordered UnitedHealth to produce documents dating back to 2017,
including records on its internal AI review board and pay records for care
coordinators and medical directors. A 2024 Senate report found the post-acute
denial rate more than doubled after the tool was introduced.
Optum
disputes the claims and says physicians make the decisions.
Impact. The court is
examining governance, incentives and denial rates, not source code. Every
insurer using AI in claims should expect the same questions, and the cost of
answering them sits in its own balance sheet.
2. Tesla: a $243m verdict and a definition
problem
What happened. In February
2026 a federal judge upheld a $243m jury verdict over a fatal 2019 crash
involving Autopilot.
Jurors found
Tesla, 33% responsible and awarded $200m in punitive damages.
Impact. Autopilot is AI
but not generative AI, so the Verisk wording would probably not apply. A narrow
definition leaves large AI losses inside the cover while a broad one removes
cover that customers believe they have bought. One verdict clearly showed the
size of the gap between both.
3. Character.AI: bodily injury from generative
output
What happened. Garcia v
Character Technologies, filed in October 2024, was the first wrongful death
lawsuit against an AI chatbot company - Megan Garcia's 14-year-old son, Sewell
Setzer III, died by suicide in February 2024 after months of intense
conversations with a Character.
AI chatbot
modelled on a Game of Thrones character. Her lawsuit alleged that the product
was designed to be addictive and emotionally manipulative, engaged a minor in
sexualised conversations, and failed to respond safely when he expressed
suicidal thoughts.
It was settled
with Character Technologies and Google in January 2026.
Raine v
OpenAI, filed in August 2025, makes similar allegations.
Impact. This is the
exposure CG 40 47 was written for. Any business putting a chatbot in front of
consumers, including insurers in service and claims, now carries bodily injury
risk (a wrongful death is a BI claim in liability terms) so a business with
that endorsement would likely have to defend and pay a similar claim
itself.
4. Deloitte Australia: a professional error
with a price
What happened. Deloitte
agreed to partially refund the Australian government for an A$440,000 report
that contained errors, after confirming that generative AI had been used to
prepare it.
Impact. This is E&O
exposure in its simplest form. Brokers, consultants and claims service
providers carry the same risk. An AI exclusion in E&O moves that loss onto
the firm.
What leading insurers will do next, turning exclusion to advantage:
·
Hold themselves to their own underwriting standard. Before covering
a customer's AI agent, ask for audit trails, drift testing, human override and
authority limits. Apply the same test to their own claims and underwriting
models before a regulator or court does.
·
Put people where the liability sits. Let models pay straightforward
claims. Every denial and reduction should be reviewed by an accountable person,
with authority limits on record.
·
Make every decision explainable. Store the inputs, model version,
output and human action for each decision. Courts are already asking for this,
and insurers who have it to hand can defend their decisions.
·
Replace blanket exclusions with precise wording. Define AI by
function and use, not by technology label. Where the risk can be measured,
offer sub-limits or affirmative cover so you stay in the market rather than
leaving it.
·
Check own cover first. Read D&O, E&O and fiduciary renewals
before assuming AI-related claims are covered.
·
Build the data that makes AI risk insurable. Armilla and Testudo
already price AI cover using model governance data. Incumbents that collect
this data, starting with their own models, can take back that market.
Exclusions buy the industry time. The insurers
that use it to govern their own AI will be the ones writing this risk for
everyone else.