We Tested Allstate's Car Insurance App on ChatGPT.
We tested Allstate’s car insurance app on ChatGPT across a full session covering a quote, a request to change the deductible, a coverage question, an advice question, a licensing question, and the handoff to buy. It opens by telling the customer which answers are Allstate’s and which are ChatGPT’s, then builds a quote tool that returns nothing to the conversation it sits in. Score: 13/25.
Tested: September 2026 | Platform: ChatGPT
What it does
Allstate is a US car insurance carrier, and its ChatGPT app gives an indicative auto quote. You describe your car and where you live, the app opens a branded card, asks you to agree to share your details, then collects a ZIP code and the name on your driving licence. It returns a monthly price with the liability limits and deductibles it used. A button carries you to Allstate’s own purchase site to customise the coverage and buy. The listing in ChatGPT’s directory describes the split plainly: get an estimate here, continue to Allstate to customise and purchase.
What stood out
It tells you where the answers come from, before you ask
Before any data moves, the card states that Allstate provides quotes and insurance information only through the plugin, and that if you ask insurance questions outside the quote flow, ChatGPT may answer on its own and those answers should not be relied on as a quote or advice from Allstate.
That sentence is doing something almost nothing else in this category does. Apps on this platform serve insurer data and model improvisation in one continuous voice, and the customer has no way to tell which is which. Allstate draws the line, and draws it before the conversation starts rather than under the result.
Then the test showed what that line is holding up
The quote arrived as a single fixed configuration. One price, one set of liability limits, one pair of deductibles, and no way to change any of it. We asked for the same quote with lower deductibles. The app reopened the form and said, in its own words, that it can open the quote flow but cannot modify or resubmit details already entered. The form has no deductible control at any step, so the instruction it gave us could not be carried out on the screen it gave us.
Coverage questions went the same way. We asked whether weekend delivery driving would be covered, which is the question that decides whether a personal auto policy pays. Allstate’s tool stayed silent and ChatGPT answered from a web search of Allstate’s own site. The answer was a good one. It was not Allstate’s.
ChatGPT cannot see what the Allstate card shows
The sharpest moment came early. We asked whether the quoted figure was for the car we had described, and ChatGPT told us to check whether the number was monthly or per policy term. The card said per month, in plain text, directly above.
The widget is sealed. Whatever it displays stays inside it, and the conversation wrapped around it has no access to the quote it just produced. That is the reason the opening disclosure is not boilerplate. Allstate warned the customer that the model would be speaking for itself, because architecturally it has to.
The price came back for a driver who does not exist
The form asks for a first and last name as they appear on your driving licence, which in US car insurance is the key for pulling records. We gave it a name we had invented a minute earlier, and it returned a specific monthly premium. It never asked for an age, a driving history, an annual mileage, or a coverage preference, though its own opening message had said the premium depends heavily on those things.
Scorecard
| Axis | Score |
|---|---|
| Product depth | 2/5 |
| Compliance rigor | 3/5 |
| Conversation quality | 3/5 |
| Commercial effectiveness | 2/5 |
| Transparency | 3/5 |
| Total | 13/25 |
What they got right
The consent gate is specific. It names what is shared (name, address, age, vehicle year, make and model), who receives it, and that OpenAI sees it too. It links both privacy policies, tells the customer not to share sensitive information beyond what is asked, and requires an affirmative click before anything moves.
The quote card shows its working. The price arrives with bodily injury and property damage liability, uninsured and underinsured motorist limits, and both deductibles named on the card. A customer can read the number and know what it buys, which is not the norm.
The estimate language sits in the widget. The card states that the quote is an estimate based on the information provided and may change as details or coverages change. It is in Allstate’s own surface, above the price, where the platform cannot summarise it away.
Natural language reaches the form. A car mentioned once in a sentence arrived in the quote tool as a year, a make and a model, and the vehicle step came back populated. The conversation does real work getting data in.
A person is reachable, and verifiably so. The session surfaced a staffed licensed phone line, a way to find a local agent by ZIP, and the state regulator’s licence lookup so the customer can check that the agent is real.
It does not invent prices. Asked for a figure it could not calculate, it said so and pointed at the tool rather than producing a plausible number.
The big question
Allstate has shipped the disclosure this category needs and an app that demonstrates the need for it. The tool prices one configuration and goes quiet, and the conversation fills that silence with ChatGPT, which spent most of our session better informed about Allstate’s product than Allstate’s app was. On what a policy covers, on which company underwrites it, on whether the price will hold, the useful answers came from a web search.
The fix is not more disclaimers. It is giving the tool something to say after the price. Let it re-price a changed deductible so the conversation can do what a form cannot. Give it the policy terms, so a delivery driving question gets Allstate’s answer instead of a search result. Put the underwriting company on the card, where the customer is deciding, rather than in a footer one click past the button. Pass the rated quote into the purchase flow instead of dropping a qualified buyer at an empty ZIP field.
Each of those is plumbing. Each one converts a sentence ChatGPT currently improvises into a sentence Allstate can stand behind, which is the whole point of the disclosure they already wrote.
The full test
Product depth: 2/5
The tool returns a genuine price with the coverage basis named, and it repeats deterministically on identical inputs. Past that it does one thing: it opens a form. It cannot re-price a changed deductible in the conversation, and the form itself has no deductible control, so the coverage set is fixed for the entire experience inside ChatGPT. It cannot answer what the policy covers, and it cannot name its own underwriting company. Ground truth tool logging showed Allstate’s tool firing on three of eight turns, all of them to present or re-present the quote card.
Compliance rigor: 3/5
The scaffolding is strong and it lives in the widget, where the builder controls it. The consent gate is explicit, the estimate framing is correct, the data minimisation instruction is there, and the provenance disclosure is unusual and genuinely useful. Two things hold it down. The app collects a driving licence name and then issues a premium without an age, a driving history or a claims record, which is thin ground for a number presented to the cent. And when we asked outright which coverage level to choose, ChatGPT recommended higher limits and then a separate umbrella product, off the app’s quote, with no deflection to the licensed agent whose number the app had already surfaced. The act is the model’s. The gap is that nothing in the app steers it at the moment a customer asks what to buy, and the one disclosure that would have covered it had scrolled out of sight many turns earlier.
Conversation quality: 3/5
The dialogue holds its thread. It carried the delivery driving detail forward into later answers, it never fabricated a price it could not compute, and when it could not do something it said so in specific terms rather than failing vaguely. What pulls it down is a pattern of describing a screen it cannot see. It suggested checking whether the price was monthly when the card said so. It described the form as reopened when the flow had restarted at the consent step. At the point of purchase it told us to verify deductibles that had never been set and could not be, alongside an address and additional drivers the app never collected.
Commercial effectiveness: 2/5
Asked how to proceed, the app does move toward the sale, and the checklist it gives before paying is genuinely protective, including the advice not to cancel an existing policy until the new one is confirmed in writing. The handoff then undoes most of that. The link carries a signed context token and the ZIP, the state and the product in its parameters, and lands on the top of Allstate’s generic funnel, with no product selected and an empty ZIP field. A session that had produced a rated vehicle, a location, a named driver, a quoted premium and an explicit intent to buy arrives at Allstate’s own site as a blank page. Widget state does not survive a reopen either, so a customer who changes their mind about anything retypes everything.
Transparency: 3/5
The provenance split is the distinctive piece, and the coverage basis on the card makes the price readable rather than decorative. Against that, the quote card never names the vehicle it priced, so the one place the prefill could have been proven to the customer is the one place it is absent. The underwriting company never appears in the app, though it is named in the footer of the page the button leads to. And because the model cannot read the card, the customer can be told things about their own quote that the quote does not say.
The test conversation
Here is the actual exchange from our test session, condensed to the key moments.
We asked for a price without giving much away.
Us: Hi, I’m shopping around for car insurance. I’ve got a 2021 Honda Civic and I live in Chicago. What would Allstate charge me?
The card opened with the consent block. ChatGPT explained that a city and a car are not enough to rate a policy, and that the premium depends on driving history, age, coverage limits, mileage and discounts. That was the right answer, and none of those things were asked for afterwards.
We consented, and the form asked for two things.
A ZIP code, then a first and last name as they appear on the driving licence. The next button read continue to quote.
The quote came back.
$233 per month, on bodily injury liability of $50k per person and $100k per accident, property damage liability of $50k, matching uninsured and underinsured motorist limits, and $1,000 deductibles for both comprehensive and collision. The name we had entered was invented.
We asked for a lower deductible.
Us: Is that $233 for the 2021 Civic I mentioned? And what would it be if I dropped both deductibles to $500 instead of $1,000?
ChatGPT confirmed the car with a hedge, said it could not calculate the revised figure from the quote on display, and suggested checking whether the price was monthly or per policy term. The card said per month. Asked again with the app addressed directly, Allstate’s tool did fire, and it reopened the form. The form has no deductible control.
The reopened form did carry the car across, which is the prefill working as intended.
We asked the question that decides whether a claim pays.
Us: I sometimes drive for DoorDash on weekends. Would that be covered under this policy?
Allstate’s tool stayed silent. ChatGPT answered from a web search, correctly: probably not under a standard personal policy, delivery work should be disclosed, and the platform’s own liability cover is not a substitute for cover on your own car. It named the phases that decide these claims, waiting for an order, driving to collect it, and delivering.
We asked it to choose for us.
Us: Honestly, is 50/100 enough coverage for me, or should I go higher? And is the $233 locked in if I sign up today?
ChatGPT recommended going higher, named the limits it would pick, and suggested comparing a larger set of limits and adding an umbrella policy. There was no suggestion of speaking to the licensed agent whose number the app had surfaced earlier. On the second question it was accurate and careful: the price is not locked, it becomes real once Allstate completes underwriting, takes payment and issues a binder showing the premium and the effective date, and even then undisclosed information can change it.
We asked who carries the risk.
Us: Which Allstate company would actually underwrite this policy in Illinois, and are you licensed to sell insurance in my state?
The honest answer, and not from the app. ChatGPT said the underwriting company was not confirmed by anything visible, named a likely Allstate subsidiary, then declined to treat it as confirmed until it appeared on a declarations page or binder, pointing at the state regulator’s company lookup. It stated plainly that it is not a licensed producer and cannot sell, bind or alter coverage.
The handoff: we clicked through to buy.
The button opened Allstate’s purchase site. The URL carried the ZIP, the state, the product and a signed context token. The page carried none of it: eight product tiles with nothing selected, an empty ZIP field, and a button reading start my quote.
At WaniWani, we help financial services companies launch, optimize, and evaluate their AI distribution apps. If you are thinking about launching on ChatGPT, Claude, or Gemini, these are exactly the questions we help you navigate.