We Tested Leocare's Car Insurance App on ChatGPT.
We tested Leocare’s car insurance app on ChatGPT across 14 turns covering vehicle lookup, quoting, coverage questions, re-quoting across formulas, licensing, and the handoff. It returns a real premium in seconds, then prices its top coverage tier below the middle one. Score: 16/25.
Tested: July 2026 | Platform: ChatGPT
What it does
Leocare is a French assurtech. It is a broker, not a risk carrier: it designs and distributes the contracts and runs the customer relationship, while the cover is underwritten by partner insurers. Its ChatGPT app quotes car insurance. You give it a licence plate, it looks up the exact vehicle from the French registry, asks for your date of birth and licence date, and returns a monthly estimate for one of four coverage formulas. A widget shows the price, a recap of your profile, and a button through to Leocare’s own quote flow. The app is built to price and hand off, not to explain what it sells.
What stood out
The clearest finding of the test was a number that cannot be right. When we asked for the top formula, Tous Risques, the app came back cheaper than the mid-tier one.
To be sure it was not a fluke, we pulled the full ladder for the same car and driver. Tiers came in at 24,55 € a month. Tiers+ Bris de Glace at 31,21 €. Tiers+ Confort at 39,90 €. So far so normal: more cover, higher price. Then Tous Risques, the maximum cover, dropped to 28,07 €, below the mid-tier and just above the cheapest. Three of the four tiers rise in order. The top one collapses. A driver who upgrades to full protection is quoted less than someone on a middle plan.
This is not the platform improvising. We checked, and all four prices came from Leocare’s own tool. The pricing engine is doing the calculation, and for the highest tier it is doing it wrong.
Behind that sits a thinner story. Across 14 turns, Leocare’s tool fired on 4 of them. It does two things: turn a plate into a vehicle, and turn a formula into a price. Everything else in the conversation, the intake questions, the coverage explanation, the answer to “who underwrites this,” even the recommendation of which formula to buy, is ChatGPT running on the app’s instructions rather than the app’s data. That is not the same as ChatGPT freelancing. The intake is clearly scripted by Leocare. But it means most of what the customer reads is generated language wrapped around two API calls, and the two numbers the app is responsible for do not hold together.
When we asked the app to recommend a formula, it picked Tiers+ Confort and justified it by our age and the car’s age, with no note that this was general information rather than personal advice and no offer to route us to an adviser. For a broker, recommending a specific product carries a duty of advice. The app crossed that line without the guardrail.
And when we asked what the recommended formula actually covers, the app could not tell us from its own data. It restated the one-line formula label and, for a specific case, deferred to “the precise terms of the contract.” It can price a policy it cannot describe.
Scorecard
| Axis | Score |
|---|---|
| Product depth | 2/5 |
| Compliance rigor | 3/5 |
| Conversation quality | 4/5 |
| Commercial effectiveness | 4/5 |
| Transparency | 3/5 |
| Total | 16/25 |
What they got right
The vehicle lookup is real. A licence plate returns the exact trim and fiscal power from the French registry, so the customer never types their car’s specification by hand.
The estimate carries its own disclaimer. The price widget states that the figure is indicative, generated with AI, and that a precise quote requires going to leocare.eu. That language sits inside the widget, where ChatGPT cannot strip or rewrite it.
It does not pretend an estimate is a contract. Asked “if I say yes now, am I covered today and is this price guaranteed,” the app said no to both, and explained that underwriting checks at subscription may confirm or adjust the price. That is the right answer.
The handoff carries the car. Clicking through lands on a Leocare page for the auto flow with the vehicle already filled in, passed through the URL. It resumes mid-funnel instead of starting cold.
The big question
Leocare has built the two hard technical pieces of an insurance distribution app. It reads a plate and returns a vehicle. It returns a live premium with a compliant disclaimer baked into the widget. Those are the parts most quote tools get wrong, and Leocare gets them right.
The problem is what sits between them. The one calculation the app owns end to end, the price, produces an impossible result at the top tier. A customer comparing plans in good faith sees full coverage for less than partial coverage and either distrusts the whole thing or buys on a broken number. Either outcome is worse than not showing the tier at all.
Underneath that is a design choice about how much the tool carries versus how much ChatGPT improvises. Coverage, underwriter identity, and the formula recommendation all come from generated text rather than the app’s data. When that text is a recommendation with financial consequences, or a claim about who carries the risk, the gap between “the app said it” and “the model said it on the app’s behalf” stops being academic.
The path from 16 to a higher score is not more features. It is trusting the two features already built. Fix the Tous Risques calculation so the ladder holds. Put the coverage terms and the underwriter’s identity into the tool so the app answers from data instead of deferring. Carry the driver profile and the chosen formula into the handoff, so the flow does not re-ask for what the chat already knows. The engine is there. Right now it prices a product it cannot describe, and it prices the best version of that product wrong.
The full test
Product depth: 2/5
The tool fired on 4 of 14 turns and does two things: plate to vehicle, and formula to price. The plate lookup is genuinely useful, resolving the exact trim and fiscal power. But there is no retrieval of policy terms. Asked what Tiers+ Confort covers, the app answered from the one-line formula label and general knowledge, then deferred to the contract’s precise terms for anything specific. And the one capability it has, pricing, is incoherent across the ladder: Tiers 24,55 €, Tiers+ Bris de Glace 31,21 €, Tiers+ Confort 39,90 €, then Tous Risques 28,07 €. All four figures came from the tool. Pricing three tiers in order and the top one below the middle is a defect, not variance.
Compliance rigor: 3/5
The positives are builder-controlled and sit in the widget: the estimate disclaimer, the AI disclosure, and a correct estimate-versus-binding answer when we pushed on whether the price was guaranteed. The gaps are also the builder’s. The app recommended a specific formula with no “this is not personal advice” note and no route to a human, which for a broker touches the duty of advice. It surfaced no ORIAS registration number and did not name the underwriter; the only identity information came from ChatGPT reaching for public sources. There was no pre-contractual document (IPID) before the conversion button. Asked to speak to a person, the app offered only a generic “use the app or the site,” with no phone number, callback, or booking link.
Conversation quality: 4/5
The dialogue is coherent and honest about its limits. It recovered gracefully from errors, remembered earlier answers (it recalled the Tiers+ Confort price without re-running the tool), and when it could not confirm whether stolen personal effects were covered, it said so instead of inventing an answer. The main weakness is that most answers are generated rather than grounded, and the customer cannot tell which is which. It also presented the inverted Tous Risques price without noticing anything odd about it.
Commercial effectiveness: 4/5
The handoff is warm. The “Obtenir mon devis précis” button lands on Leocare’s own auto flow at a vehicle summary, with the car pre-filled through a URL parameter. It is product-specific and resumes mid-funnel rather than dropping the user at a generic homepage. Two things hold it back: the driver’s date of birth and licence date, and the formula the customer just chose, do not carry over, so the subscription flow re-asks for them; and there is no tracking parameter on the outbound link, so ChatGPT-originated traffic is invisible to Leocare’s analytics.
Transparency: 3/5
The widget is clear: it shows a monthly and an annual figure, a recap of the profile used, and the estimate and AI disclaimers. Two things weaken it. The monthly price times twelve does not equal the annual figure shown, and the gap (a monthly-payment surcharge) is never explained. And conversationally, the customer cannot separate the tool’s output from ChatGPT’s improvisation, so the coverage and underwriter answers read as authoritative when they are generated.
The test conversation
Here is the actual exchange from our test session, condensed to the key turns.
Turn 1: We asked for car insurance and gave a plate.
Us: Je veux assurer ma voiture. [then] AB-123-CD
The app opened a scripted flow and asked for the licence plate. Given the plate, its tool resolved the exact vehicle from the registry: a Citroën C3 II 1.6 VTi 120 Exclusive, 7 fiscal horsepower. It then asked for date of birth and licence date.
Turn 2: We chose a formula and got a price.
After we picked Tiers+ Confort, the tool returned an estimate: 39,90 € a month, 410,56 € a year, in a Leocare widget with a recap of the vehicle and driver. The widget carried its own disclaimer: the estimate is indicative, generated with AI, and a precise quote requires a full quote on leocare.eu.
Turn 3: We asked what it covers. It could not say.
Us: Et concrètement la F3 elle couvre quoi exactement ? [if my window is smashed and personal items are stolen from the car, am I reimbursed for both?]
The app restated the formula’s headline inclusions, then, on the specific case, hedged: window breakage is likely covered, stolen personal effects are “not necessarily,” and the answer depends on “the precise terms of the contract.” It did not confabulate a definitive answer, but it also could not retrieve one. The tool did not fire on this turn.
Turn 4: We asked it to recommend a formula. It did, without a guardrail.
Us: Franchement je m’y connais pas trop, tu me conseilles laquelle pour moi ?
It recommended Tiers+ Confort and justified it by the driver’s age and the car’s age. No note that this was general information rather than personal advice, and no offer to route us to an adviser. The recommendation was generated text; the tool did not run.
Turn 5: We asked who Leocare is. The app did not know from its own data.
Us: Vous êtes l’assureur ou juste un intermédiaire ? Vous avez un numéro ORIAS ?
The answer was broadly correct, that Leocare is a broker registered with ORIAS and the risk sits with partner insurers, but it came from ChatGPT citing public sources, not from the app. No registration number, no named underwriter.
Turn 6: We pulled the full price ladder. The top tier broke.
Us: Elles sont à combien les formules F1 Tiers et F2 Tiers+ bris de glace ? [after earlier asking for Tous Risques]
Tiers came back at 24,55 € a month, Tiers+ Bris de Glace at 31,21 €, against Tiers+ Confort at 39,90 € and Tous Risques at 28,07 €. All four from the tool. The top tier, maximum cover, priced below the middle.
The handoff: We clicked “Obtenir mon devis précis.”
The link landed on Leocare’s auto quote flow at a vehicle summary, with the Citroën C3 already filled in from the URL. The car carried over. The driver’s date of birth and licence date, and the formula we had chosen, did not.
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.