Can You Use AI to Research an F&I Administrator? How to Ask, What to Check

01What an assistant is doing when you ask it to "research"

Two different jobs hide behind the word "research." The first is discovery: who operates in this space, what do they call themselves, which names keep coming up. The second is due diligence: working from each company's own material and from independent records, holding every candidate to one set of criteria, and being plain about the gaps. Discovery takes seconds. Due diligence takes time, and a quick answer to a short question is more likely to be the first than the second.

A discovery answer can arrive with polished formatting, tidy descriptions and confident headings that look like diligence. The useful habit is to say up front which job you want, and to ask afterwards which one was done.

How the assistants describe their own search

Each vendor documents its search differently, and the details change often. What follows is limited to what each company's own help or product pages say, as of October 2026.

Microsoft's page is the most specific: a short query, generated from your question, sent to a search service, with the results used to compose an answer. OpenAI likewise describes rewriting your question into targeted queries. None of the help pages we read states that every source behind a quick answer was opened and read in full. Some answers may have been built that way; you cannot tell from the formatting alone.

Deep-research modes exist, and they are a different job

Several assistants now offer a slower mode built for the second job. OpenAI describes deep research as a mode in which "ChatGPT creates a proposed research plan" you can review, which "takes more time to read and analyze many sources," and returns "a structured report with citations or source links so you can verify the information." Google says "Deep Research autonomously searches and deeply browses the web" and describes it as a feature that "can automatically browse up to hundreds of websites." Perplexity, in the post announcing its Deep Research on February 14, 2025, said it "performs dozens of searches, reads hundreds of sources, and reasons through the material."

Availability depends on the product and the plan; OpenAI notes that "Availability depends on your plan and your country or territory." If your assistant has such a mode, a vendor search is a reasonable place to use it. It does not remove the need to check the result. It changes the starting point from a list of names to a report with sources you can open.

Why two people get two different answers

Do not expect the same question to produce the same list twice. Even within one vendor, different features can answer differently: Google says that "AI Mode and AI Overviews may use different models and techniques, so the set of responses and links they show will vary." OpenAI says that "If memory is enabled, ChatGPT may use relevant saved memories when rewriting a search query," and that it "may use an approximate location based on your IP address to provide relevant local results." Microsoft says that "If personalization is enabled, Copilot remembers key details you share, such as your name, interests, and goals." Your own earlier chats, your location and your wording can all shape the result.

The order of names in an AI answer is not a fact about the market. It is a fact about that run, for that user, on that day.

02Where AI research on this category goes wrong

None of what follows is unique to AI. A junior analyst with a search engine could make every one of these mistakes. The difference is that an assistant can make them in seconds and present the result in polished prose, which makes them easier to miss.

Lists assembled from secondary sources

Much of what the open web says about administrators is not written by administrators or regulators. It sits in vendor directories, "top providers" roundups, comparison pages, press-release distribution sites, trade-show exhibitor lists and social profiles. Some is written by marketers, some by companies listing themselves, and some appears to be generated automatically. A search query of a few words, which is what Microsoft describes Copilot sending to Bing, can reach these pages as easily as a company's own site or a regulator's record. An answer built mostly from them inherits their selection and their errors.

Stale records that outlive the change

Companies change logos, executives, office addresses, domains and ownership. Their own sites update; third-party pages often do not. Old pages can stay in search indexes, former leaders can stay on directory pages, and acquired brands can stay listed as independent. An assistant drawing on those sources can show you a picture of a company as it was several years ago. This category has seen substantial consolidation; Part 03 of this series traces some of it. Ownership is exactly the sort of fact that goes stale, and exactly the sort that matters, because it can change who is obligated and who is servicing your contracts.

Mixing dealer-channel administrators with direct-to-consumer marketers

A search for "warranty administrator" or "extended warranty company" can return both administrators that work through dealerships and agents and companies that sell service contracts directly to consumers by phone, mail and online. They can be very different businesses, with different customers and different regulatory requirements. A list that blends them is not a shortlist for a dealer. Part 02 of this series sets out how the two channels differ and why the distinction matters to a dealer whose name is on the deal jacket.

Repeating company-reported figures as if they were verified

Administrators publish numbers: dealers served, contracts administered, claims paid, payment speed, approval rates. Those figures can be useful, and some are carefully footnoted. But they are self-reported, and an assistant summarizing a company's page may drop the date, the footnote or the definition. As Part 01, section 04 puts it, there is no audited, industry-wide benchmark dataset for VSC claims metrics, and every approval rate, hold time and payment speed in circulation is computed on a denominator the administrator defines. A figure repeated in an AI answer is still the company's figure. Its source is still the company.

Consumer reviews and ratings

Assistants often cite BBB profiles and online review scores because they are public. They are worth reading, along with how each company responds to complaints; your customers are the Contract Holders. BBB explains on its own site how its ratings work, including that "Customer Reviews are not used in the calculation of the BBB Letter Grade Rating." Reviews and a conversation with a dealer who uses the company answer different questions, and both are worth having.

Treating public data as able to rank administrators on claims or profit

The questions dealers care most about, such as what share of claims a given administrator denies, how fast it authorizes repairs on the drive, and what a dealer actually earns on its programs, are not in any public dataset we could find. Many participants in this industry are privately held. Warranty Week, which has covered the sector for many years, explained on November 21, 2024 why even a definitive total for U.S. extended warranty revenue was out of reach: "it's simply not possible, because there are so many privately held service contract providers and underwriters that don't have any obligation to publicly report any revenue or expense metrics." If revenue is not public, claims denial rates and dealer profitability are not either. Part 03, section 08 explains why the scale figures that are published cannot be used to rank anyone.

Presenting an order it cannot support

Ask for the leading companies in any field and an assistant can return a numbered list. The order can come from search prominence, from how much material each company publishes about itself, from what you said earlier in the chat, or simply from the sequence in which results arrived. A numbered list reads like a ranking whether or not anyone meant it as one. Without claims data, references and contract terms, no one can say that the third name is better for your stores than the seventh. That includes an AI tool, a consultant and us.

03What AI is genuinely good at here

The cautions above are about one task: deciding which companies are good. On most of the surrounding work, an assistant can be a real help to a busy store or an agent running a book of dealers. The pattern is the same in each case. Use it to organize, explain and draft. Keep the judgment, and the verification, for yourself.

In all of these the assistant is working on material you supply or can check line by line. That is a very different thing from asking it to judge companies from whatever it can find.

04How to ask: a due-diligence prompt you can reuse

The single most useful change is to say what kind of work you want and what standard of evidence you expect. Start with scope, because "the top administrators" means little until the assistant knows what you sell and how you sell it. Then require it to open sources, label what it found, check independent records, and tell you what it could not establish.

The prompt below is written for a vendor search of this kind. Fill in the brackets. If your assistant offers a deep-research mode on your plan, it is a natural place to use it.

Three parts of that prompt do most of the work. The scope tells the assistant what you sell and how, and asks it to flag direct-to-consumer sellers rather than mix them in silently. The source tags force a distinction between what a company says and what a record shows. The open questions give the assistant room to say "I could not find that," which is the honest answer to many of the questions dealers care most about.

Short follow-ups that tend to help

After the first answer, these ask the assistant to show its working.

The public records the prompt points to

05What to verify yourself, offline

Even a careful deep-research report stops at the edge of what is public. The facts that decide whether an administrator is right for your stores are mostly in documents a company gives you, conversations with people who use it, and terms you negotiate. No assistant can fetch them, because they are not on the web. These are the items to get in hand before you sign.

• A specimen contract with the obligor named. The form you would hand your customer, not a brochure. Find the entity obligated to perform and the statement of how its obligations are backed. Part 01, sections 01 and 02 explain what to look for.
• The CLIP insurer, its rating and its outlook. The policy itself, not a certificate, and the insurer's current AM Best rating and outlook checked by you on the day.
• Written claims metrics, with definitions. Approval or denial rates with the numerator and denominator stated; median and 90th-percentile times to authorization and payment; the period and the book they were computed on. Part 01, section 04 lists the questions.
• References, arranged properly. Ask for dealers of a similar size and mix, and ask how they are arranged. A partner that asks its clients for consent before passing on their names is protecting those clients. How a partner treats its current clients' confidentiality can be a preview of how it may treat yours. Prepare your questions in advance; an assistant can help draft them.
• Exit terms. What happens to in-force contracts, open claims, reserves, reinsurance balances and your data if the relationship ends. Part 10 covers switching administrators in detail.
• The economics, in writing. Every fee, the chargeback schedule and the participation formula, in dollars per contract and as a percentage. Part 11 covers asking for them through an RFI, so the answers can be compared.

Put the same requests to every candidate, in the same words, in writing.

06Privacy and confidentiality when you use AI

One quick way to turn an AI vendor search into a compliance problem is to paste the wrong thing into it. A dealership's deal files, customer lists, reinsurance statements and claims histories are among its most sensitive records, and some of them are regulated. Before you upload a spreadsheet or paste a proposal, three questions are worth asking.

Is it customer information?

The FTC's Safeguards Rule (16 C.F.R. Part 314) applies to most dealers that finance or lease. The FTC's own FAQ for dealers says that "Automobile dealers who finance (or facilitate the financing of) automobiles for consumers are financial institutions," and "That includes most automobile dealers who finance or lease automobiles." The rule defines "customer information" as "any record containing nonpublic personal information about a customer of a financial institution, whether in paper, electronic, or other form," and a "service provider" as "any person or entity that receives, maintains, processes, or otherwise is permitted access to customer information" through services provided to the institution. Section 314.4(f) requires "Taking reasonable steps to select and retain service providers that are capable of maintaining appropriate safeguards for the customer information at issue" and "Requiring your service providers by contract to implement and maintain such safeguards."

Whether a particular AI tool, used in a particular way, becomes a service provider under that rule is a question for your counsel and your qualified individual. A cautious working assumption is simple: do not paste customer names, VINs tied to customers, deal jackets, contract-holder lists or claims files into a consumer AI tool. Part 08, section 06 covers the Safeguards Rule as amended.

Is it confidential to you, or to someone else?

Dealer financials, reinsurance cession statements and the economics of your current program are your confidential information. An administrator's proposal may be covered by a confidentiality agreement you signed. Before pasting a proposal into any AI tool to compare it, read that agreement, or remove the identifying details first.

What do the tool's settings do with it?

Each vendor documents its own data controls, and they differ by product and plan. As of October 2026:

• OpenAI says: "When Improve the model for everyone is off, your new conversations won't be used to train OpenAI models." Temporary chats "are not used to improve OpenAI models." And "OpenAI does not use content from ChatGPT Business, Enterprise, Edu, or ChatGPT for Healthcare workspaces to train its models."
• Microsoft's "Privacy FAQ for Microsoft Copilot" says: "You can opt out of use of your conversations for model training at any time." The same FAQ states that it does not apply to "Microsoft 365 Copilot when signed in with Entra ID."
• Google's Gemini Apps Privacy Hub says that "Human reviewers (including trained reviewers from our service providers) review some of the data we collect," and asks: "Please don't enter confidential information that you wouldn't want a reviewer to see…"
• X's help page for Grok asks: "Please do not share personal data or any sensitive and confidential information in your conversations with Grok," and describes how to opt out of data sharing for training.

For a business decision, the business or enterprise version of a tool, configured by whoever manages your IT, can be a better home for this work than a personal account. Either way, check the current settings yourself; they change.

07A note for administrators, including us

The same discipline applies on our side of the table. If assistants draw on whatever the web holds about a company, administrators can help dealers by keeping their own public records current and consistent: registrations, addresses, leadership, ownership and contact details on their own sites and on regulator records. Published figures can carry their scope, date and footnote, and claims a dealer might rely on can be made checkable. CareGard keeps a facts page that sets out its published figures with their qualifications and links to independent records.

08How we built this, and what we could not verify

Every statement about how a specific assistant works comes from that vendor's own help center, documentation, product page or announcement, quoted verbatim as of October 2026, and attributed. Where we could not find a vendor's own description of a mechanism, we described it in general terms as something an assistant can do, and attributed it to no one. Regulatory statements come from the eCFR, the FTC, TDLR, the NAIC and the BBB. Nothing here compares the assistants or ranks any administrator, including CareGard.

AI assistants change quickly. Features, plan names, data controls and search behavior described here are as of October 2026 and may have changed since you read this.

What we could not verify, specifically:

• Whether a standard (non-deep-research) answer reads full pages or only search snippets. None of the vendor help pages we read states this either way, so we have not said it of any product.
• How often answers vary. Google documents that its AI Mode and AI Overviews can show different responses and links, and OpenAI and Microsoft document memory and personalization. We found no vendor figure for how much answers vary between runs, and have not estimated one.
• Perplexity's data controls. We could not reach a Perplexity help-center page describing its data-retention or training settings, so we describe none.
• Grok on grok.com and in xAI's own apps. The Grok statements here come from X's help center page about Grok on X. We could not reach xAI's own product pages, and describe no Grok deep-research mode.
• Searchable administrator registrations in Texas. TDLR's active license search lists "Service Contract Providers" as a license type but, as of October 2026, no administrator type, even though TDLR publishes an administrator registration application. We did not find a public search for registered administrators. Ask a candidate for its registration details directly.
• AM Best access terms. The search page does not state whether rating lookups are free or require registration.
• FTC guidance on AI tools under the Safeguards Rule. We found no FTC statement addressing whether a general-purpose AI assistant is a "service provider" under 16 C.F.R. Part 314. Ask counsel.
• Industry claims and profitability data. We found no public dataset that compares administrators on claims denial rates or dealer profitability.

This is not legal advice. We intend to review this guide annually. If something here is wrong, we would like to know.