Step 1: Understanding, before anyone asks for a quote
The traditional research path — search a term, browse an FAQ, call an agent — returns fragments. It handles badly exactly the questions people have. "I'm driving a rental car in Europe — how does my auto policy apply?" has no keyword. Neither does most of what buyers want to know, which is usually a version of what will I owe, and when?
Language models can read a policy document and map coinsurance, riders and network tiers into plain English on demand: "you'd pay the first $1,000 before insurance starts paying," with a worked example if asked. That plain-language disclosure improves comprehension is well established. What is new is that the simplification can be produced per person, in the moment, rather than drafted once for everyone.
The regulatory ceiling here is high, because explaining a policy is descriptive. Underwriting and pricing are prescriptive, and are governed very differently. Almost nothing an insurer does at this step triggers those obligations — which is why it is the easiest place to start and the most common place to stop.
Step 2: Quoting and comparing
A conversational intake can skip what a linked record already answered, pull fields from an uploaded prior policy, and prompt only for the one field left blank. The form does not disappear; it stops being a wall of empty boxes. Some carriers are testing flows that assemble and price a quote in one pass, routing unusual cases to a human while standard ones run through.
Comparison is where the leverage is, and where the market underdelivers. The hardest part of shopping is comparing what policies do, not what they cost — traditional tools put three premiums side by side and bury the differences in limits, networks and exclusions. AI can render those readable, and can simulate a claim under each option, which is what the buyer was trying to work out anyway.
The J.D. Power 2026 U.S. Insurance Digital Experience Study, covering auto and home, found that only about a third of shoppers encountered a pricing tool that included other brands; 27% found one comparing options within the same insurer; 28% saw none at all. Where comparisons were available, 39% considered purchasing, against 21% where they were not.
One line matters more than it looks. Recommending a plan is decision support, on the buyer's side of the table. Determining eligibility and setting a premium are regulated insurer actions. New York's guidance applies to AI systems and external consumer data used in underwriting and pricing, not to customer-facing information tools.
That line is clean in theory and getting blurrier in practice. If a recommendation engine's ranking determines what 90% of buyers select, its practical effect on who ends up with what coverage is not obviously smaller than a pricing model's. No regulator has answered this yet. Insurers building these tools should not assume the answer will stay convenient.
Step 3: The application, and the wait
This is where most coverage of AI in insurance stops, and where the buyer's experience is actually worst.
After submitting an application, a buyer typically enters a period of silence. They don't know what stage the file is at, whether anything is missing, or how long it will take. In life insurance this can run weeks, and the traditional cause — paramedical exams and fluid collection — is precisely what accelerated underwriting was built to remove. Insurers now use external data (prescription history, MIB records, motor vehicle reports, public records, credit attributes) to waive those requirements for some applicants and issue in days rather than weeks.
Speed is the visible benefit. The less visible one is status. Almost everything that makes waiting unbearable is an information problem, not a processing-time problem: not knowing where the file is, not knowing that a missing document is the reason, not knowing whether to call. AI-driven status tracking, proactive requirement notices and plain-language explanations of what each stage means are entirely customer-side. They touch no underwriting decision, carry none of the fairness obligations, and address the part of the process buyers actually complain about.
Two constraints belong at this step. First, accelerated underwriting relies on external data about the applicant, and the NAIC's Accelerated Underwriting Working Group guidance expects insurers to have a mechanism to correct mistakes when they are found — which presumes a route for the applicant to see and challenge what was used. Second, the same guidance notes that adverse decisions are sometimes reviewed by a human underwriter. Sometimes is doing real work in that sentence, and it is the sort of thing market conduct examiners have been directed to look at.
Step 4: The decision, and how it is delivered
If a buyer is going to have an emotional experience with an insurer before a claim, it happens here: a rating up, an exclusion, or a decline.
This is also the single point in the journey where AI's obligations are heaviest and its usefulness is most constrained. New York's Circular Letter No. 7 (2024) expects insurers to disclose upfront that AI systems or external consumer data are used in underwriting; to honor an applicant's right to request the specific data used about them; and, on an adverse underwriting decision, to provide notice detailing what the decision rested on — in writing, within 15 days. The NAIC's accelerated underwriting guidance runs parallel: insurers should be able to give the reasons for an adverse decision, and all the information it was based on, whether or not the data falls under FCRA.
Read together, these produce a requirement most insurers have not operationalized: an adverse decision must be explainable in ordinary language, to the person it happened to, on a deadline.
There is a genuine role for AI in meeting it — turning a model's contributing factors into a notice a person can act on, in their own language, with the correction route stated plainly. But this is also the step where AI cannot be the last word. An explanation generated for a decision it did not make, from factors it inferred rather than knows, is a compliance failure with a fluent surface. If a system tells an applicant their deductible is $500 when the policy says $1,000, the error is recoverable. If it tells them why they were declined and gets it wrong, it has manufactured a defective regulatory notice.
The design implication is narrow and important: generation is fine, authority is not. Whatever the model drafts at this step, a human owns.
Step 5: Issuance, and the first year
Two things happen after the policy is bound that no one writes about. The buyer receives a document they will not read, and enters a free-look period they usually don't know exists.
Both are addressable with the least regulated capability discussed here. A policy summary in plain language, generated per customer and covering what changed from the quote; a prompt during the free-look window explaining what it is and when it ends; a first-year check-in that explains the renewal notice before it arrives. None of this touches pricing. All of it affects whether the customer believes the product is what they were sold.
The constraint underneath all five steps
In a J.D. Power survey of 2,099 U.S. insurance customers conducted in mid-August 2025, comfort with AI tracked the stakes closely. Routine roles were fine: automated claim status updates (24%), billing (23%), basic service questions (21%). As consequences rose, comfort dropped — 47% were somewhat or very uncomfortable with AI processing their claims. On pricing, only 15% believed insurers should fully use AI; 33% said such use should be limited until bias and ethics concerns are addressed, and 30% said it should be confined to partial use with safeguards.
The most useful number in that survey is not about AI at all. Asked who gains most when insurers adopt it, 68% said the insurance company captures most or all of the benefit. Only 26% said it is shared evenly.
That reframes the whole exercise. Steps 1 through 5 above describe real improvements — faster answers, clearer comparisons, less silence, better notices. But if two-thirds of customers assume the benefit accrues to the carrier, every one of those improvements arrives pre-discounted. The trust problem is not that people misunderstand what AI does. It is that they are not persuaded it was installed for them.
Which suggests the durable advantage is not a better model. It is being the first carrier that can demonstrate, concretely, where the gain went — a shorter wait the customer can see, a decline they could actually understand and contest, a summary that told them something they didn't already know.
Conclusion
Search becomes conversation. Forms become guided intake. The wait becomes visible. Decisions become explainable on a deadline. Channels stop resetting.
Each gain carries an obligation: automation requires governance, personalization requires privacy, prediction requires fairness, and AI-influenced decisions require accountability. The ceiling rises and falls across the journey — highest at Step 1, lowest at Step 4 — and building as though it were uniform is the most common mistake in this category.
Where Tigerless fitsThe following describes our own position.
We build on the customer side of the line described above. In practice, the boundary we've found hardest to draw is not between AI and human, but between explaining a decision and appearing to make one — an assistant that answers "why was I rated up?" too confidently has crossed it, even when the answer is correct. Lara is built to route that class of question to a licensed person rather than resolve it, and we treat that as a product requirement rather than a limitation.
Sources: J.D. Power 2026 U.S. Insurance Digital Experience Study (13 May 2026); J.D. Power Insurance Intelligence Report, 4 November 2025 (survey fielded mid-August 2025, n=2,099); NYDFS Insurance Circular Letter No. 7 (11 July 2024); NAIC Accelerated Underwriting in Life Insurance Regulatory Guidance (adopted 2024); NAIC Model Bulletin on the Use of Artificial Intelligence Systems by Insurers (4 December 2023).
