AI for insurance agents: 9 prompts, what never to paste, and the AI rule 24 states and DC adopted

How can insurance agents use AI? For the writing, from redacted notes

Among agencies using AI, 49% use it for marketing content, 43% for coverage form analysis and 35% for contract reviews. That's from the Big "I" 2026 Agency Universe Study, as reported by Insurance Journal on 24 September 2026. The table maps the nine prompts on this page to those jobs and the daily ones around them, with what you remove before pasting.

Five of the nine need no client data at all. The other four work as well with [client] and [policy] in place of the real values, because the model needs the facts of the coverage, not the identity of the person. That habit is also the cheapest answer to the barrier agencies cite most after know-how: 48% of agencies named security and privacy, per the same study.

The other kind of AI on the first page of results faces the client: a chatbot or voice agent that answers policyholder questions, books appointments and takes call notes. Nothing on this page needs one. A bot also changes who sees the data, because every answer a client types into it goes to the bot's vendor before anyone at the agency reads it. If a client types a license number or a health detail, that's Section 3K Nonpublic Information under Model #668. Section 3P of the same model defines a "Third-Party Service Provider" as a Person that "contracts with a Licensee to maintain, process, store or otherwise is permitted access to Nonpublic Information". Captive agents should ask the carrier before adding one, since the NAIC AI bulletin places AI governance on the insurer. Drafting, with a person reading every line before it goes out, keeps that person between the model and the client.

Which AI rules apply to insurance agents? The NAIC bulletin binds carriers, not producers

The NAIC's AI model bulletin applies to insurers, and on the NAIC's own map dated 1 April 2026 it had been adopted in 24 states and the District of Columbia. California, Colorado, New York and Texas run their own AI rules for insurers. The other 22 states, Florida, Georgia and Ohio among them, had adopted neither. Adoption came fast and then slowed: 21 of the 25 adoptions landed in 2024.

The bulletin, adopted by the NAIC on 4 December 2023, describes the agent's own ground in its background section: "AI techniques are deployed across all stages of the insurance life cycle, including product development, marketing, sales and distribution." Yet the word "producer" doesn't appear anywhere in the adopted text, and neither does "agent" (we searched the PDF on 3 October 2026). Every expectation lands on the insurer and its written AI program. For a captive agent, that makes the carrier's AI program the rule book, including which tools you may paste client files into. For an independent agent, the bulletin reaches you only through whatever each appointing carrier writes into its own terms.

The rule that does name producers is a different NAIC model, the Insurance Data Security Model Law (#668), adopted in 28 jurisdictions, Puerto Rico among them, on the NAIC's 1 April 2026 map. The map carries its own caveat: it "does not reflect a determination as to whether the pending or enacted legislation contains all elements of the model." The model applies to licensees. Its Section 9 exceptions turn on agency size and on whose security program covers an agent. The NAIC text (adopted in the 4th quarter of 2017, technical edit 2025) reads:

The average independent agency reported 9.9 staff in 2026, right at the line, up from 8.2 in 2024, per the Big "I" study. That figure is a mean, and it counts staff, while the model law counts "employees, including any independent contractors". States can change the threshold and the wording when they adopt the model, and the NAIC map does not say which did. This page quotes the NAIC text. It isn't legal or compliance advice. Whether Section 4 applies to a given agency, captive or independent, is a question for your state's department of insurance or the agency's own counsel.

What are the best AI prompts for insurance agents? Five for renewals, quotes, coverage, claims and contracts

The best prompts for insurance agents hand the model the coverage facts, ban the promises an agent can't make, and leave a placeholder where the client's identity would go. Each prompt below runs in ChatGPT, Claude or Gemini as written. Replace the brackets and keep the negative lines.

Prompt 1, the renewal with a rate increase: "Write a renewal email under 120 words from me, an insurance agent, to [client]. Their [auto/home] policy renews on [date] at $[new premium], up from $[old premium]. The carrier's stated reasons: [paste]. Explain the change in plain words, offer a review call, and list two things that could lower the premium, such as a higher deductible or a bundle. Do not promise a discount, a rate or a coverage change."

Prompt 2, the quote comparison: "Compare these two quotes for the same client in a table: carrier, premium, deductibles, each coverage limit, notable exclusions and endorsements. Use only what is in the quotes; write not stated where a quote is silent. Then write three plain sentences on the real differences. Quotes: [paste, with names, birth dates and license numbers removed]." The "not stated" line stops the model filling a gap from general knowledge. That's how a comparison table goes wrong.

Prompt 3, the coverage explainer: "Explain this clause from a [HO-3 / BOP / personal auto] policy form to a client with no insurance background, in under 100 words, with one example of a loss it covers and one it does not. End with: your policy documents govern what is covered. Clause: [paste the form wording]." Never ask the model whether a specific loss is covered. That's the carrier's claim decision, and a wrong yes in writing from an agent is an errors and omissions problem.

Prompt 4, the claim status update: "Turn these adjuster notes into a calm update for [client] under 90 words: where the claim stands, the next step and its date, and what the client needs to send. No guesses about the outcome or the payment amount. Notes: [paste with the claim number, injuries and medical details removed]."

Prompt 5, the contract insurance requirements: "Read the insurance section of this contract my commercial client must sign. List every required coverage, limit, additional insured, waiver of subrogation and primary and noncontributory wording in a table, quote the exact line for each, and flag anything their current policies may not meet: [list current coverages and limits]. Contract section: [paste]." Contract reviews were the third most common AI job in the Big "I" study, at 35%. This is what one looks like.

Can AI write insurance marketing and prospecting emails? Yes, with four prompts and no client data

Yes. Marketing content is the most common AI job among agencies using it, at 49% in the Big "I" 2026 study. None of these four prompts needs a client record, so they're the safe place to start.

Prompt 6, account rounding: "Write a 90 word email to a client who has [auto] with us but not [home or umbrella]. Explain in one sentence what the second policy protects that the first does not, mention that bundling can lower the total cost without naming a figure, and ask for a 10 minute review. No pressure language."

Prompt 7, commercial prospecting: "Write a cold email under 100 words to the owner of a [restaurant / contractor / dental practice] in [city]. Name one exposure this kind of business commonly underinsures, in plain words, and ask for a 15 minute call. Peer to peer tone, no statistics I have not given you, no claims about price."

Prompt 8, the social post: "Write three versions of a post for [Facebook / LinkedIn] about [hurricane season / teen drivers / renters insurance]: 40 words, 90 words, and one that opens with a question. One call to action, no emoji, no promise of savings or coverage."

Prompt 9, discovery questions: "I have a first call with a [new homeowner / small business owner]. List the 12 questions an independent agent should ask before quoting, grouped by property, liability and life changes, each with one line on why it matters to the quote." Captive agents can swap "independent agent" for the carrier's product line. The questions change less than you'd expect.

Run the longer ones (prompts 2 and 5) in two different models when the stakes are high and read only the rows where they disagree. On a social post, one model and your own edit are enough. AI for small business owners has the general agency jobs, bookkeeping and job ads, and AI for real estate agents covers the closing side of the same client.

What should an insurance agent never paste into ChatGPT?

Never paste what the NAIC data security model law calls Nonpublic Information. That's a client's name or other identifier together with a Social Security number, a driver's license number, an account or card number, an access code or password, or biometric records. It's also any health information about a client or their family other than age or gender. That's Section 3K of Model #668. It doubles as the practical list of what never to paste, whatever line an agent writes.

In an agency, that list covers more paper than it sounds. Life and health applications carry medical history. Motor vehicle reports carry license numbers. Claim files carry injury details, and payment screens carry account numbers. Before any of it touches a chat, replace each identifier with a placeholder ([client], [license ending 1234]) and remove the health detail entirely. Every prompt above still works.

Then check the tool's training default, because on consumer plans it's on. OpenAI's help center says chats on individual plans such as Plus may be used to train its models unless you switch that off. Business and Enterprise are excluded by default. Anthropic moved consumer Claude accounts to train unless you opt out in its August 2025 terms update. Claude Team and Enterprise list model training as none by default. Captive agents have one more step first: the carrier's approved tool list, which the carrier's own security program sets. What never to paste into an AI chat has the general list beyond insurance.

Which AI tool is best for insurance agents? It depends on the size of the agency

For a solo or captive agent, the best AI tool is a $20 a month chat plan, ChatGPT Plus or Claude Pro, with training switched off and the carrier's rules checked first. For an agency office, it's a business plan with training off by default and an admin console. That's easier to document if the agency has to show how it chose a vendor. In a state that adopted Model #668 with the NAIC's line at ten employees, Section 4F describes that choice as vendor due diligence. Your state department of insurance can say whether it applies to your agency. Prices as of October 2026:

Whizi is the publisher of this page, and it loses the agency row. It has no team accounts or admin console. Its privacy policy names no SOC 2 report or HIPAA terms, so an agency that has to document how it chose a vendor gets a cleaner file from ChatGPT Business or Claude Team. It fits the solo agent who wants more than one model. Starter has ChatGPT Luna and Claude Haiku, and Pro at $29.99 a month adds Claude Sonnet and ChatGPT Sol. A Project keeps an agency brief of up to 32,000 characters and 10 pinned files on every chat. A message sent to Claude still reaches Anthropic and one sent to GPT reaches OpenAI, per Whizi's data and privacy docs, so the same list of what never to paste applies unchanged.

None of these tools is an agency management system. They draft. Your AMS or CRM keeps the policy record, the carrier portal issues the policy, and nothing here quotes a real premium. ChatGPT Business vs Plus has the seat math for a two to five person office.

Will AI replace insurance agents? Not on the government's projection

No. The US Bureau of Labor Statistics projects insurance sales agent jobs to grow 3% from 2025 to 2035, "About as fast as the average for all occupations", from 572,600 jobs in 2025. Median pay was $62,280 in May 2025, and 12% of agents are self-employed, per the Occupational Outlook Handbook.

The handbook also states what a chat app can't take on: "Agents must be licensed in the states where they work." The license, the appointment and the errors and omissions policy all sit with a person, and every prompt on this page leaves the decision with that person. What changes is the share of the week spent writing. An agent who drafts renewals, comparisons and prospecting emails from prompts 1 to 9 keeps the calls and the coverage advice, which is the part clients pay an agent for.