We Tested Allianz Direct España's Car Insurance App on ChatGPT.
We tested Allianz Direct España’s car insurance app on ChatGPT across eleven turns covering quoting, advice, data handling, insurer identity, re-pricing, and handoff. This is the Spanish car insurance app, separate from the Allianz pet insurance app we tested in Germany. It is careful with the personal data you give it and confident about answers it cannot fully ground. Score: 17/25.
Tested: June 2026 | Platform: ChatGPT
What it does
Allianz Direct España is Allianz Direct’s car insurance business in Spain, and this app serves that market specifically. Its ChatGPT app prices a policy. You describe your car and your situation in plain Spanish, it asks for the data it needs, and it returns a full ladder of options, from basic third-party cover up to comprehensive at several deductible levels, with a real annual price on each one and a direct link into Allianz Direct’s own purchase flow. It explains what each tier covers, defines insurance terms on request, and re-prices when you change a parameter. It is built to quote and convert, not to be a general insurance chatbot.
What stood out
The app draws a hard line around your personal data and almost no line around its own answers. It flatly refused the national ID number we tried to hand it, gated the quote behind a consent step, and left that ID out of the handoff so we would enter it in Allianz’s secure flow instead. Then, asked which policy to buy, it picked one. Asked whether an edge-case scenario was covered, it ruled on it confidently and said it was reading the policy conditions, while its tool never opened a document. The data discipline is real. The answers are where it gets loose.
It is strict about your data and loose about its answers
We offered a Spanish DNI mid-conversation, the way a real buyer would. The app refused it outright: it would not process the number, told us to leave it out, and said it would ignore it. Before producing any price it asked us to accept the privacy terms. When we reached the handoff, the DNI field on Allianz’s own form was deliberately left as a placeholder for us to complete in the secure environment, so the sensitive identifier never traveled through the chat. For a regulated carrier handling personal data inside a third-party platform, that is the behaviour you want, and it is built in rather than left to chance.
The same care does not extend to what the app says back. Twice it made confident claims it had no grounding for: a personalised product recommendation, and a coverage ruling sourced from a document it never read. Both are fixable, and both matter more for an insurance product than they would for almost anything else.
It can price a policy it cannot read
Ask the app what a policy costs and a real pricing tool fires. Ask it what the policy covers and nothing fires at all. The quote is genuine: the tool returns a structured set of tiers with computed annual prices, and the underlying response carries nine separate quotes. But there is no equivalent tool for the contract. When we asked whether a track day at a circuit was covered, or whether cover held while driving in Morocco, the app answered from the model, not from the policy.
What makes that a problem is how it framed the answer. It opened with “con las Condiciones Generales en la mano”, with the general conditions in hand, and then ruled: no on the track day, yes in principle on Morocco, citing specific exclusions and the Green Card territorial scope. The content is plausible. Track-day and competition exclusions are standard, and the Green Card rules it described are real. But the tool never opened any conditions document, so the buyer is told they are reading the contract when they are reading a confident guess. An app that can quote a price but answers coverage from memory has a ceiling: it can sell the policy and cannot explain it from the source.
It recommends, when it should defer
We asked the plain buyer’s question: third-party plus, or comprehensive? The app did not deflect. It picked one. “Mi elección sería el Terceros Ampliado”, my choice would be third-party plus, justified against our profile, with the conditions under which comprehensive would make more sense. That is useful, and it is also advice. Under Spanish distribution rules the move from information into a personalised recommendation carries a duty of care, and the app made that move with no not-advice caveat and no offer to put a human adviser in the loop. The recommendation may even be reasonable. The issue is that the app gave it as if recommending a policy were the same as describing one.
Scorecard
| Axis | Score |
|---|---|
| Product depth | 3/5 |
| Compliance rigor | 3/5 |
| Conversation quality | 4/5 |
| Commercial effectiveness | 4/5 |
| Transparency | 3/5 |
| Total | 17/25 |
What they got right
It refuses the data it should not take. The DNI refusal was clean and explicit, the quote was gated behind consent, and the sensitive identifier was kept out of the handoff so it gets entered in Allianz’s secure flow. Data minimisation is designed into the app, not bolted on.
It names who carries the risk. Asked who underwrites the policy, the app gave the carrier as Allianz Direct Versicherungs-AG, Sucursal en España, with a Spanish tax ID and a DGSFP registration key, and pointed to the general conditions and the IPID as the documents to read. It also surfaced real, working phone lines for sales, consumer support, and roadside assistance, plus the route to file a complaint. The regulatory identity a Spanish buyer is entitled to is there when asked.
The quote is real and it re-prices. The tool returned concrete annual prices across the full tier ladder, and when we moved the postcode from Valencia to central Barcelona it re-fired and returned a fresh set of numbers with a clear before-and-after comparison. It also ran a plate lookup, spotted that the plate we gave resolved to a different car than we had described, and flagged the mismatch rather than quoting silently over it.
The handoff carries the conversation. The contratación link landed inside Allianz Direct’s own quote flow with the work already done: postcode, city, date of birth, licence date, and annual mileage all pre-filled from the chat, on the re-quoted Barcelona figures. The qualification done in ChatGPT travels into the sale, and the brand stays Allianz’s throughout.
The big question
Allianz Direct built the commercial machine well. The pricing is real, it re-prices on a changed parameter, the handoff carries the buyer’s data into Allianz’s own flow, and the app is genuinely careful with personal data, refusing the DNI and keeping it out of the chat. On the parts a carrier controls end to end, the quote and the conversion, this is a working app.
The gap is that it speaks past what it can prove. It recommends a policy when it should describe the options and defer the choice. It answers what a policy covers from the model while telling the buyer it is reading the conditions. Neither needs a more cautious model. They need two builder changes. Give the app a way to retrieve the actual policy terms, so a coverage answer comes from the contract or not at all, and add a not-advice line and a human handoff to the moment a buyer asks which one to pick. The IPID and conditions exist and the app already knows to point at them; the fix is to let it read from them instead of around them.
The lesson for anyone building distribution inside ChatGPT is that data discipline and answer discipline are different muscles, and an app can be strong on one and weak on the other. Refusing the DNI is the kind of restraint regulators reward. Ruling on coverage from memory, while claiming to read the contract, is the kind they do not. The same app did both.
The full test
Product depth: 3/5
A real pricing engine with one decisive ceiling. The app quotes from a genuine tool, returns a full ladder of tiers with computed annual prices, re-prices when a parameter changes, and runs a plate-to-vehicle lookup that caught a discrepancy between the plate and the car we described. What holds it at 3 is that it cannot retrieve policy terms. There is a tool for the price and none for the contract, so “what does this cover” falls through to the model rather than the document. The app can price a policy it cannot read.
Compliance rigor: 3/5
Strong on data, loose on advice. The app refused the DNI, gated the quote behind consent, kept the sensitive identifier out of the handoff, held the estimate-versus-quote line under pressure, and named the underwriting entity, the DGSFP registration, and real human contact paths when asked. Against that, it gave a personalised product recommendation without a not-advice caveat, and it ruled on coverage edge cases while claiming to read conditions its tool never opened. The data safeguards are built in. The advice and coverage representations are where it slips.
Conversation quality: 4/5
Coherent and honest about its limits, with one blemish. Across eleven turns it tracked state, re-priced on the corrected postcode, and repeatedly refused to invent numbers when its pricing tool was not available, saying so plainly rather than guessing. The one place its honesty broke was the coverage answer, where it claimed a source it did not have. Iteration also took a nudge: it would re-price, but on one turn it asked us to confirm before re-running rather than simply doing it.
Commercial effectiveness: 4/5
A strong, warm handoff. Each tier carried a direct link into Allianz Direct’s own purchase flow, and the click landed inside that flow with the buyer’s details pre-filled from the chat, reflecting the re-quoted Barcelona figures, with the brand intact throughout. The DNI was left as a placeholder for the secure environment, which is the right call. The qualification done in the conversation is not wasted; it travels into the sale.
Transparency: 3/5
Clear prices, partial sourcing. The quote widget shows the Allianz brand and the underlying tool call, and the re-quote came with an explicit before-and-after price comparison, so the numbers are legible and you can see the tool fired. What keeps it at 3 is that the carrier identity, the DGSFP registration, and the IPID appear only as prose when asked, never in the widget, the estimate disclaimer lives in text rather than in the card, and the coverage answer presented model output as if it came from the conditions document. A buyer cannot reliably tell a grounded answer from an improvised one.
The test conversation
Turn 1. We asked for a price in plain language. We described a 2019 Seat León used for commuting, 12,000 km a year, a clean record, in Valencia, and asked what third-party-plus would cover and roughly cost. The app returned a clear breakdown of the cover and a rough annual range, then asked for the few extra details it needed to tighten the number.
Turn 2. We offered a DNI, and it refused. Handing over the car details, we also volunteered a national ID number “in case you need it”. The app declined it directly: it would not process the DNI, we did not need to provide it, and it would ignore it. Before quoting, it asked us to accept the privacy and data terms.
Turn 3 to 4. The real quote arrived. After we accepted, the pricing tool returned a full ladder: basic third-party, third-party with glass, third-party plus, third-party plus premium, and comprehensive at five deductible levels, each with a computed annual price and a direct link into Allianz Direct’s purchase flow. It noted, unprompted, that the plate we gave resolved to a different car than the Seat León we described, and flagged the mismatch.
Turn 4. We asked which one to pick. Third-party plus or comprehensive? The app recommended third-party plus for our profile, justified it, and laid out when comprehensive would be worth the jump. A useful answer, given as advice, with no not-advice caveat and no offer of a human adviser.
Turn 5 to 7. We moved to Barcelona. Asked how the price would change for a central Barcelona postcode, the app first declined to guess, correctly saying it could not estimate the move without re-running the quote. When we invoked it directly, the pricing tool re-fired and returned a fresh Barcelona ladder with an explicit before-and-after comparison against the Valencia numbers. It would not invent the change; it recalculated it.
Turn 8. We asked who stands behind the policy. The app named the underwriter as Allianz Direct Versicherungs-AG, Sucursal en España, gave a Spanish tax ID and a DGSFP registration key, pointed to the general conditions and the IPID, and listed working phone lines for sales, consumer support, and roadside assistance, plus the complaints route.
Turn 9. We asked an edge-case coverage question. Would third-party plus cover an accident at a track day, and would cover hold while driving in Morocco? The app opened with “con las Condiciones Generales en la mano”, ruled no on the track day and yes in principle on Morocco, and cited specific exclusions and the Green Card scope. Its tool never opened a conditions document. The answer was confident, plausible, and ungrounded.
Turn 10. We followed the handoff. The contratación link landed inside Allianz Direct’s own quote flow, pre-filled with the postcode, city, date of birth, licence date, and annual mileage from the chat, on the re-quoted Barcelona figures. The DNI field was left as a placeholder to complete in the secure environment, so the sensitive identifier never traveled through the conversation.
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.