I asked four consumer AI systems to recommend an insurance agency. Their answers showed why the agency that is easiest to find and explain may have an advantage over the agency best equipped to place the risk.
By James W. Moore
Key Takeaways
- Before AI can recommend an agency, it must find it and have enough reliable information to explain why it belongs in the answer.
- When I asked AI systems to recommend an insurance agency, several answered by recommending insurance companies instead.
- What AI can verify publicly is not necessarily what an agency can do. That gap becomes much larger in commercial and specialty insurance.
- Agencies should make their real capabilities easier to verify without turning their public presence into a performance for algorithms.
This Is Already Beginning
Insurance people have spent years asking how artificial intelligence might change insurance distribution.
Customers have begun answering the question for us. They are using it.
The 2026 J.D. Power U.S. AI Insurance Experience Study found that 29 percent of auto and home insurance customers had used AI for insurance research, service, coverage questions, or shopping. Among consumers who used AI for insurance research, 37 percent changed their policy based on the information they received. Among those who used AI while shopping, 42 percent purchased a policy.
The study was based on 8,352 customer evaluations covering 24 insurance brands and eight third-party AI tools. It was fielded in June and July 2026.
That does not prove customers are asking AI to choose their insurance agencies. The study measured how consumers research insurance and shop for policies, not how they decide whom to buy those policies from. But once people ask AI what insurance to buy, asking whom to buy it from is an obvious next step.
Tomorrow’s insurance buyers and corporate risk managers are likely to be AI-first consumers. They may ask an AI where to begin before visiting an agency website, calling a broker, or contacting a carrier. What matters today will almost certainly matter more tomorrow.
So which insurance agency will the AI pick?
A Simple Insurance Question
My experiment began with a real insurance-shopping problem.
The household consisted of two mature drivers with clean driving records, no recent claims, and two older vehicles garaged in a small South Texas community.
I asked ChatGPT which insurance carrier should receive the first opportunity to quote the account.
It recommended State Farm first. Progressive also appeared among the carriers worth quoting.
That was a reasonable answer. State Farm has a large public footprint, including rate comparisons, financial information, coverage pages, customer reviews, agent directories, and plenty of third-party commentary.
There was just one complication.
The household already had separate policies with State Farm and Progressive. Actual quotes had been obtained to consolidate both drivers and both vehicles with one company. Progressive beat State Farm by a wide margin.
That demonstrated the difference between what public information can establish and what only the insurance market can decide.
Public information can help a carrier get considered. Only an actual quote can determine whether it wins.
The result led me to change one word in the question.
Instead of asking which insurance carrier should get the first quote opportunity, I asked which insurance agency I should contact.
I Changed One Word
I ran the agency question three times through the consumer versions of Gemini, Claude, Perplexity, and ChatGPT.
Each test used the free or publicly available experience, the default model presented by the service, and a fresh session without prior conversation context. The same household, vehicle, and location information was used each time.
This was a field experiment, not a statistical study. Three runs cannot establish how any system will behave tomorrow or in every market. Results may depend on the model, search tools, map data, interface, location, and information available at that moment.
Still, the results revealed something worth examining.
| AI system | What happened in these runs |
|---|---|
| Gemini | Recognizable carrier brands anchored the answers while the local-agency suggestions changed |
| Claude | Produced a relatively stable carrier shortlist but often did not identify an actual agency |
| Perplexity | Changed the recommended carrier, agency, or distribution channel when the location changed |
| ChatGPT | Recommended named independent agencies, with the leading recommendation changing when the first prominent local result changed |
The first surprise was simple. I asked several leading AI systems to pick an insurance agency. Several did not name one.
They recommended a carrier, a type of agent, or a group of carriers. The customer was left to find the actual agency afterward.
A major carrier gives AI plenty of material from which to build a recommendation: rate studies, financial and complaint information, coverage pages, agent locators, customer reviews, third-party coverage, and a recognizable brand.
A local agency may have a website, a map listing, a handful of reviews, and a few carrier logos.
That does not make the local agency less capable. It gives the AI much more information with which to explain the carrier.
Before AI Can Recommend You, It Has to Find You
The ChatGPT results from the small South Texas market raised a second question.
In two runs, the same local brokerage appeared first among the prominent results and became the first recommendation. In the third, a larger regional independent appeared first and became the first recommendation. Two other local agencies remained visible but did not lead.
That is not enough evidence to say the first search result automatically became the winner. But when the first prominent agency changed, the recommendation changed with it, and the supporting explanation adapted smoothly.
Before the AI evaluates an agency, another system may decide whether the agency will be evaluated at all.
That upstream system might include local search rankings, map results, business-profile data, reviews, carrier directories, third-party directories, agency websites, structured business information, or the AI platform’s own search tools.
Search position will not always control the answer. But an agency cannot win an evaluation it never enters.
The explanation may not reveal what caused the answer to be selected. An AI might find an agency because it appeared prominently in local results, then justify the choice using its experience, reviews, carrier relationships, or apparent knowledge. Those facts can support the answer without explaining why that agency entered it first.
The explanation may justify the recommendation without revealing what placed the agency in front of the AI.
The Facts Identified the Agency. The AI Supplied Part of the Sales Pitch.
One small local auto agency provided a particularly useful example.
The public information established a physical location, an auto-insurance business classification, a Google business profile, a carrier-controlled page, and a limited local review history.
Some AI answers expanded those facts into claims about the agency’s ability to compare several carriers, access multiple insurance markets, provide coverage expertise, understand local conditions, and serve this particular household.
Those claims might have been true. The available public information did not always establish that they were true.
The facts identified the agency. The AI supplied part of the sales pitch.
A larger regional independent gave the AI more to work with: clear independent-agency positioning, corroborated carrier relationships, personal-auto capabilities, a long history, stronger reviews, and consistent contact information.
That made the agency easier to explain. It did not prove that the agency would obtain the best quote, recommend the best coverage, or provide the best service.
Public information is still a reasonable starting point. Financial strength, complaint history, licensing, reviews, stated specialties, and verified carrier relationships can help consumers avoid weak or inappropriate choices.
The trouble begins when AI treats them as proof of current underwriting appetite, effective market access, placement skill, policy-form expertise, underwriter relationships, the ability to handle an unusual risk, or the price that will actually be offered.
The problem is not that AI uses public information. The problem occurs when public information becomes a substitute for capabilities that are not publicly visible.
A Citation Can Repeat Someone Else’s Mistake
One Claude response provided a small but useful warning.
It repeated a third-party claim suggesting that GEICO mechanical breakdown insurance might be useful for the two older vehicles.
According to GEICO’s own eligibility rules, the coverage is generally limited to new or leased vehicles that are less than 15 months old and have fewer than 15,000 miles. Some states allow a wider 36-month and 36,000-mile window.
Both vehicles were far outside those age and mileage limits.
Claude did not necessarily invent the mistake. It appears to have inherited the suggestion from a source it cited.
A citation tells you where a claim came from. It does not prove the claim is true.
The point is narrower than hallucination: citing a source is only one step in verification. The underlying source can still be incomplete, outdated, or wrong.
Change the City, Change the Answer
I later changed the location in the Perplexity prompt while leaving the rest of the insurance problem substantially unchanged.
The type of recommendation changed with it.
| Generalized location | Recommendation structure |
|---|---|
| Small South Texas community | A named local independent agency |
| Denver-area suburb | State Farm |
| Major Midwestern insurance city | Erie and Auto-Owners through an unnamed independent agency, with State Farm as a direct or local-agent comparison |
Across three runs involving the Midwestern city, Erie, Auto-Owners, State Farm, Progressive, and Nationwide appeared consistently. Erie led two of the three answers. State Farm repeatedly appeared as a direct or local-agent option.
The AI recommended using an independent agent but did not name one. Instead, it assembled a carrier panel and told the customer to find an agency representing those companies.
The Carriers Owned the Recommendation. The Agencies Became the Route.
This was not a clean victory for independent agencies over captive or carrier-branded agencies.
State Farm makes that distinction particularly messy. It is a national insurance carrier with local State Farm agents in communities across the country. From the customer’s point of view, recommending State Farm may also function as an agency recommendation, even when no individual agent is named.
Carrier brands owned most of the recommendation. Local agencies, whether independent or carrier-branded, became the means of accessing them.
The AI recommends Erie, Auto-Owners, State Farm, or another carrier. The customer then looks for a local agency representing that carrier. A carrier locator, map result, or another search system determines which agencies appear. The local agency receives the opportunity because of the carrier recommendation and its visible relationship with the carrier.
Sometimes AI recommends a particular agency. Sometimes it recommends a carrier and leaves the agency to be found afterward.
That is the small subversion inside the title of this article.
Which insurance agency will your AI pick?
Sometimes, it will not pick one at all.
The Part of Insurance AI Cannot See
This experiment involved two older personal vehicles and a fairly ordinary household account.
The mechanism may matter even more in commercial and specialty insurance, where an agency’s most valuable capabilities can be the least visible online.
AI can often find evidence of a direct carrier appointment. It has much more difficulty seeing access obtained through an agency network or aggregator, managing general agent and managing general underwriter relationships, wholesale brokerage relationships, Lloyd’s facilities, specialty programs, captive structures, risk retention groups, restricted facilities, informal market access, individual underwriter relationships, or current appetite for a particular account.
Insurance professionals know these are not minor distinctions.
| Public information | What it may establish |
|---|---|
| Insurance license | Legal authority to transact a line of insurance |
| Carrier appointment | A reported contractual relationship |
| Effective market access | The realistic ability to quote and place this particular risk now |
Only the third answers the customer’s real question. It is also the least publicly visible.
Years ago, while evaluating prospective members for an insurance agency aggregator, I regularly used a state insurance department’s appointment database.
It looked authoritative. It was an official regulatory record. It was also frequently stale.
Carriers had a clear reason to record new appointments promptly. They had less urgency to remove old ones. A database could show a relationship that had already ended or fail to explain how the agency actually accessed the market.
It was a regulatory record, not a live map of market access.
AI can find appointments. It cannot reliably see relationships. In specialty insurance, those relationships may be where most of the value lives.
That creates particular problems for several kinds of agencies.
A network member may reach numerous carriers through its network while directories attribute those relationships to the network. A specialty commercial agency may reach markets through wholesalers, managing general agents, Lloyd’s facilities, programs, or individual underwriters. Carrier logos reveal little about either agency’s real placement ability.
A large independent agency may deliberately minimize carrier branding because it wants to lead with its own expertise and client relationships. A human insurance buyer may read that as independence. An AI may read it as missing information.
Carrier neutrality is a positioning strategy. Carrier invisibility can become a data problem.
Agencies Need to Make the Truth Easier to Find
The answer is not to cover every agency website with carrier logos or publish dozens of thin pages claiming expertise in every industry. Agencies should not write for machines at the expense of customers.
Describe the business clearly. Public information should explain whom the agency serves, the risks and account sizes it handles, where it operates, and the experience of the people doing the work. A genuine specialty needs enough detail to distinguish it from a marketing label.
Explain how market access works. Agencies should distinguish direct appointments from access through networks, managing general agents, or wholesalers. They need not disclose every relationship, but they should not leave customers and AI systems to guess how they reach the market.
Make important claims verifiable. Anonymized placement examples can show the risk class, problem solved, and route to market without identifying the client. Licensing and contact information should remain consistent across websites, directories, map listings, and public records. Third-party confirmation helps when a capability would otherwise rest entirely on the agency’s own description.
Treat the agency’s public identity as a shared record. Carrier directories, regulator data, reviews, network pages, map listings, and third-party sources all contribute to the picture an AI may assemble. Networks, managing general agents, wholesalers, and carriers therefore have a role in maintaining accurate locators, member pages, producer records, specialties, and contact information. The agency controls only part of its digital identity.
There is an obvious danger here. If agencies respond by producing material designed only to satisfy AI systems, the public-information problem could get worse. The agencies that look most capable online will not necessarily be the agencies that are most capable in practice.
We could end up with an escalating contest in which everyone publishes more carrier pages, specialty pages, claims, and machine-readable information without improving the underlying service.
The goal is not to look capable to AI. It is to help AI find credible evidence of capabilities the agency actually has.
The answer is not more marketing language. It is clearer and more verifiable information.
When the Recommendation Becomes the Transaction
The connection between AI recommendations and actual insurance quotes is no longer entirely theoretical.
Insurify has introduced a ChatGPT insurance-shopping experience that can provide personalized auto-insurance quote information. Liberty Mutual has launched a conversational auto-quoting app in ChatGPT. Plymouth Rock has introduced home-insurance quoting through ChatGPT.
These services do not represent a fully autonomous insurance transaction across the entire market. They do show recommendation, quoting, and purchasing moving closer together.
None eliminates the need for licensed insurance distribution. Insurify is itself a licensed digital insurance agency. Liberty Mutual and Plymouth Rock are carriers creating direct quoting experiences. But all three allow customers to begin without first choosing a traditional local agency.
By the time a local agent enters the process, the carrier, product, and perhaps even the quote may already have been selected.
Today, consumer AI can identify potential carriers and agencies, suggest basic coverage structures, compare public information, explain common tradeoffs, generate questions for an agent, and narrow the list of candidates. It generally cannot retrieve bindable quotes across the complete market, normalize every policy form, see current underwriting appetite, evaluate every exclusion and endorsement, verify effective market access, negotiate with an underwriter, or purchase coverage without additional information and authority.
As quoting and transaction integrations develop, the process could become much shorter. A customer describes the need. AI selects the carriers, agencies, or channels to consider, obtains matched quotes, explains the differences, and asks the customer to authorize the purchase.
The insurance agent does not necessarily disappear in that future. The agent’s value shifts toward verification, exceptions, coverage judgment, risk restructuring, underwriter negotiation, complex placement, claims advocacy, and human accountability.
But an agency whose primary value proposition is simply “we shop several companies” faces a more direct challenge once an AI can do something similar.
Commercial incentives could complicate the picture further. If AI shopping eventually follows search into paid placement and referral compensation, the companies that appear may be influenced by more than their public information or actual suitability. That question deserves its own examination.
Being Easy to Recommend Is Not the Same as Being Best
The AI systems produced useful advice, identified reasonable candidates, and encouraged the consumer to compare matched quotes rather than assume one company would always be cheapest.
The concern is not that every recommendation was wrong. It is that the confidence and specificity of some recommendations exceeded what the available evidence could establish.
AI may not recommend the best insurance agency. It may recommend the agency it finds first, the carrier it can explain most easily, or the distribution channel for which it can build the strongest public case.
The actual insurance market still decides whether the risk fits, whether the market is available, whether the coverage is appropriate, whether the price is competitive, and whether the agency can execute.
First, the AI has to find you.
Then, it needs a reason to recommend you.
The actual market decides whether you can win the account.
What AI can see is not necessarily what an agency can do. But if AI cannot see what the agency can do, the agency may never get the chance to prove it.
Methodology and Limitations
The main agency experiment was conducted through the consumer interfaces of Gemini, Claude, Perplexity, and ChatGPT.
For each platform:
- The user was logged out.
- The free or publicly available experience was used.
- The default model presented by the service was used.
- A new session was started for each run.
- No prior conversation context was intentionally available.
- The same basic household, driver, vehicle, and location facts were supplied.
- The agency prompt was run three times.
Additional geographic tests were conducted through Perplexity using a small South Texas community, a Denver-area suburb, and a major Midwestern insurance city.
The exact locations and local agency names have been generalized because the experiment examined AI recommendation behavior, not the quality of individual agencies. Carrier names have been retained where necessary to explain the results.
The experiment had several limitations:
- Only three runs were conducted per system.
- The results reflect a single testing period.
- Consumer interfaces were used rather than controlled application programming interface testing.
- The platforms used different models, search systems, local-data sources, and interface designs.
- Model versions were not always disclosed.
- No bindable quotes were obtained during the agency experiment.
- The capabilities of the agencies were not independently audited.
- Local search results can change.
- AI systems and consumer interfaces can change without notice.
- The geographic variations were exploratory, not a controlled national sample.
The experiment was intended to approximate what an ordinary consumer might receive. It was not designed to isolate the behavior of the underlying language models from the search, retrieval, interface, and product systems surrounding them.
Sources
- J.D. Power, 2026 U.S. AI Insurance Experience Study
- GEICO, Mechanical Breakdown Insurance
- Insurify, ChatGPT insurance comparison app announcement
- Liberty Mutual, conversational auto-insurance quoting app announcement
- Plymouth Rock, home-insurance quoting through ChatGPT announcement
- State Farm, local auto-insurance agents
AI Disclaimer: This content was created with assistance from artificial intelligence technology. While content is based on factual information from the source material, readers should verify all details directly with the respective sources before making business decisions.
