Generative AI in the insurance industry

Generative AI, the class of systems that write text, summarize documents and produce images, is being used inside insurance companies and brokerages to read policies, draft correspondence and summarize claim files. For a business buyer, the bigger change is on the policy: since January 2026, ISO has offered optional generative AI exclusions for the commercial general liability form, and carriers are attaching them at renewal. The tools help the insurer; the exclusions can reduce what you are covered for.

What counts as generative AI

The NAIC's 2023 Model Bulletin defines generative AI as AI systems that generate content, such as data, text, images, sounds or video, that is similar to but not a direct copy of existing content. That differs from the predictive models insurers have used for decades to score risk (see How AI is used in commercial underwriting). A predictive model returns a number. A generative model returns language, which is easier to use and easier to get wrong.

How insurers and brokers use it

The common applications are in the office, not in the decision:

• Reading and summarizing. Models digest long policy forms, loss runs, engineering reports and claim notes, so an underwriter or adjuster starts with a summary and citations back to the documents.
• Drafting. First drafts of quotes, coverage letters, certificates, renewal proposals and marketing copy. A human reviews and sends.
• Intake and chat. Assistants answer policyholder questions about billing, coverage and claim status, and collect first notice of loss. Quality varies; the good ones hand off to a person.
• Knowledge search. Brokers and underwriters ask questions of internal guidelines and manuals in plain language.
• Claims support. Summaries of medical records, transcripts of recorded statements, and draft reserve rationales. See How AI is changing commercial claims handling.
• Code and data work. Insurers' technology teams use it to translate legacy systems and clean data.

Notice what is missing: few insurers claim to use generative models to set prices or decide coverage. Regulators expect the opposite discipline for any AI system that makes or supports a regulated decision, as the NAIC bulletin and state rules described in AI regulation in insurance: the NAIC bulletin and state rules make clear.

Where it goes wrong

1. Fabricated content. Models can state something confidently that the document does not say. In coverage work that means a wrong exclusion, a wrong limit, or a wrong endorsement number. A broker who forwards a generated coverage summary without checking it against the policy is exposed to an errors and omissions claim.
2. Data leakage. Pasting a client's submission into a public chatbot can expose confidential information. Carriers and brokers with serious programs use enterprise tools with contractual limits on data use.
3. Hidden bias. The training data may reflect past practices. Testing and human review matter more when a generated summary influences a decision about a person.
4. Over-reliance. An adjuster who accepts a generated summary of a 600-page file may miss the one page that matters.
5. Fraud on the other side. Deepfaked invoices, voices and videos are now a mainstream funds-transfer and claims risk. See Deepfakes and AI-driven fraud: what cyber and crime policies cover and fraud analytics: how insurers detect fraud.

The 2026 ISO generative AI exclusions

This is the part with direct consequences for your policy. Verisk's ISO introduced three optional exclusion endorsements for commercial general liability, available January 2026, according to Insurance Journal (August 17, 2026) and Insurance Business (September 16, 2026):

Insurers choose whether to file and attach them, and may do so state by state, at renewal. Insurance Business reported 2,369 generative AI exclusion filing records across 49 states. Because the forms are optional, some policies have them and some do not, and carriers may also use their own wording. Read the actual endorsement, not a summary, including the definition of generative AI and the phrase "arising out of," which courts often read broadly.

What the exclusions are aimed at: a business that uses a chatbot to answer customers, a content tool that produces ads or images, or an AI-generated product design, and then faces a claim for defamation, copyright-style advertising injury, or physical injury that the carrier says the AI caused. Under a standard CGL, those claims might have been argued under Coverage B or A. The exclusion is meant to end that argument.

Hypothetical example

Say a regional retailer uses a generative tool to write product descriptions and a customer-facing chatbot. The chatbot tells a shopper that a space heater is safe to leave on overnight, and a fire follows. At the renewal before, the CGL had no AI language. At the renewal after, it carries a broad generative AI exclusion. The carrier's coverage position may now be that the loss arose out of generative AI. Whether that holds depends on the exact wording, the state, and the facts, which is why the exclusion needs a coverage lawyer's reading before a claim, not after.

What to do about it

1. Check renewal paperwork for CG 40 47, CG 40 48, CG 35 08 or manuscript AI language. Ask the broker to list every endorsement that changed.
2. Inventory your AI use: customer-facing chatbots, content generation, hiring tools, code assistants, anything embedded in your product. Underwriters now ask.
3. Look at other policies for a home. Technology errors and omissions, cyber and standalone AI liability products may respond to some of what the CGL no longer does, with different triggers; see Technology E&O insurance for software, SaaS and IT firms, Cyber insurance vs tech E&O: which policy pays when and AI liability insurance: how insurers are treating artificial intelligence risk. None is an automatic substitute.
4. Negotiate where it matters. If your business depends on AI features, ask for a narrower endorsement such as the Coverage B-only form, or a carve-back for your named uses.
5. Set internal rules: human review of AI output that reaches customers, logging, and vendor contract terms that allocate responsibility.

Key takeaways

• Insurers use generative AI mostly to summarize, draft and search, not to set prices or decide coverage.
• The risks are fabricated content, data leakage and over-reliance, so human review matters.
• ISO's optional generative AI exclusions (CG 40 47, CG 40 48, CG 35 08) became available in January 2026 and are being attached to CGL policies at renewal.
• Read the endorsement itself; its definition and "arising out of" wording decide how far it reaches.
• Inventory your AI use and check tech E&O, cyber and AI-specific policies for the gap.