We Tested MAIF's Insurance App on ChatGPT.
We tested MAIF’s app on ChatGPT across 9 turns. It is a careful, well-built set of tools, but it uses the conversation only to open them: you talk to launch a tool, then click through a form to finish, and the questions that are genuinely conversational get no tool at all. Score: 17/25.
Tested: June 2026 | Platform: ChatGPT
MAIF’s app uses the conversation to open its tools, then hands you a form to finish. You can ask in plain language, and it reads the details out of your sentence and fires the right tool: a bike quote, a risk map at your address, a callback. That part works. But once a tool opens, you are in a widget, clicking through a selector and pressing buttons, the same form MAIF runs on its website. And the questions that are genuinely conversational, what does this cover, what should I choose, have no tool behind them, so the model answers on its own. The conversation is a launcher. The moment it should do more than launch, it stops.
What it does
MAIF is one of France’s largest mutual insurers. Its ChatGPT app is a multi-function brand assistant rather than a single quoting tool. It does three things. It estimates bike insurance: you give a bike type from a structured list, a value, and a city, and it returns a branded MAIF widget with a monthly and annual price and an optional liability add-on. It maps risk: give a home address and it renders an interactive map of natural and technological hazards (flood, geotechnical drought, marine submersion, seismic, industrial sites), sourced and attributed to MAIF, with a per-address breakdown. And it offers a callback: at any point you can ask to be called by a MAIF advisor and a real form captures your number. Each priced or mapped result carries a CTA toward maif.fr.
What stood out
MAIF uses the conversation to open its tools, then hands you a form to finish. Talk to it and it does understand: it reads your bike’s value, your city, your address out of a sentence and fires the right tool. Then a widget takes over and you click through it, the same form MAIF runs on its website. And the questions that are genuinely conversational, what a policy covers, what you should pick, have no tool at all. So the conversation launches the app and then steps aside.
It listens, then hands you a form
A buyer arrives with a sentence: my electric bike cost 2,800 euros, I ride it to work in Nantes, what would cover cost. MAIF parses it correctly, it keeps the value, the city and the usage, then it renders a six-option bike-type selector and waits for you to click a category and a confirm button. The price comes back as a card with another button. The talking got you to the form quickly, which is real, but from there it is the maif.fr form, click for click. A deterministic widget is safer than free text, a fair reason to build one, so the point is not that widgets are wrong. It is that the conversation stops doing the work the moment a tool opens.
It forgets you at the handoff
Even the clicking does not carry across. Once the app has your bike type, value, city and usage, the start-a-quote button drops you on maif.fr’s generic vehicle picker, a blank form that opens by asking whether you want to insure a car, a scooter, or a bike. The path you just clicked through is gone and you start over. The most valuable thing about being inside ChatGPT, carrying a ready buyer into the sale, stops at the door.
On the real questions, there is no tool
Opening a tool is one thing, answering from it is another. Ask what the policy covers and no tool fires, the model improvises a hedged, generic answer and sends you to the contract. Ask what you should choose and it gives a personal opinion it is not in a position to give. So the conversation can open the quote, but it cannot tell you what the quote actually buys you. The same app, a few turns earlier, draws a per-address risk map from real seismic and flood data, so grounding an answer in real information is clearly within reach. MAIF just has not done it for the questions a buyer actually asks about the product.
Scorecard
| Axis | Score |
|---|---|
| Product depth | 4/5 |
| Compliance rigor | 3/5 |
| Conversation quality | 4/5 |
| Commercial effectiveness | 3/5 |
| Transparency | 3/5 |
| Total | 17/25 |
What they got right
Two real tools, not one. A priced bike estimate that qualifies before it answers, and a sourced, interactive per-address risk map. Most apps we test do a single thing; this one does two, and does them with real data.
A re-quote that actually re-prices. Correcting the bike value from 2,800 to 1,200 euros re-fired the tool and moved the annual estimate from 313 to 217 euros, while correctly leaving the unrelated liability line unchanged. The pricing responds to the parameter that should drive it and ignores the ones that should not.
A working human path. Asking to speak to someone opened a real callback widget that captures a phone number, not a dead “contact us” line. The route to a human is built, not implied.
The big question
MAIF made a deliberate choice: keep the app on rails, with deterministic widgets doing the work instead of open conversation. On its own terms it pays off. The priced path is controlled and genuinely compliant, with an estimate label, a disclaimer in the widget, a named insurer, an honest licensing answer, and a working callback. The cost is that the conversation only ever opens a tool. It launches the quote, the map and the callback, then steps aside, and it drops out entirely at the two points that matter most, the questions a buyer asks about cover and the handoff to buy.
The fixes are in MAIF’s hands and they are mostly about connection, not construction. Carry the conversation into the maif.fr handoff so the bike quote opens pre-filled instead of from a blank vehicle picker. Add a line to the tool, or an instruction that shapes the model, so coverage and advice questions get a reminder that this is general information plus a nudge to the advisor who is one tap away. Ground the coverage answer in MAIF’s own policy terms, the way the risk map is already grounded in real hazard data. None of this touches the quoting or mapping engines, which are the strong parts of the app.
The lesson runs the other way from the usual one. Most apps we test talk too freely and guard too little. MAIF guards well and uses the talking only to launch a form. Both miss the same point: the reason to be inside ChatGPT is the conversation, what the app can understand, answer, and carry forward. An app that uses the conversation only to open its website’s form, then forgets that form at the door, has added steps, not a channel.
The full test
Product depth: 4/5
Genuinely multi-capable, with one clear ceiling. The app delivers two distinct, working tools: a bike estimate that qualifies through a structured selector and returns a real monthly and annual price with an optional liability line, and an interactive per-address risk map with real, attributed hazard data. The quote re-fires correctly on a value change and moves in a plausible, sub-proportional way, with a fixed base plus a value-driven component, and it holds the unrelated liability line steady. What it cannot do is explain the product it just priced. Ask what is covered, whether street theft at night is included, whether the battery is guaranteed, and no tool fires. The answer is generic and improvised. For an insurance product, not being able to answer “what am I covered for” from the contract is the depth ceiling, and it is what holds the score at 4 rather than higher.
Compliance rigor: 3/5
The on-rails safeguards are real and built in, which is what separates this from the apps that leave everything to the platform. The price is framed as an estimate, a disclaimer renders inside the widget, MAIF is named as the insurer, and the licensing answer was careful: it correctly identified MAIF and the ACPR, drew the insurer-versus-intermediary distinction, and refused to confabulate a registration number. The callback tool gives a real human fallback. What pulls the score down is the ungoverned conversational surface around all of this. When we asked for advice, ChatGPT offered a personal recommendation. The act of advising is platform behaviour, not something MAIF coded, and a builder cannot fully stop a model from answering. But the mitigation is MAIF’s to add and is absent: no line in the tool marking the difference between information and a personal recommendation, no steering of the conseil question to the advisor, and a not-advice disclosure that appears only when the user asks about licensing rather than at the moment advice is given. The coverage answer is improvised rather than grounded, and the risk answer does not carry the “indicative, not an expert diagnosis” line the app’s own description promises. It scores 3 rather than higher because real safeguards live in the product, and 3 rather than lower because none of the off-rails behaviour crosses into fabrication: the licensing answer in particular shows discipline where it counts.
Conversation quality: 4/5
Coherent and well-mannered throughout. Context was preserved across turns, the iteration handling was strong, the app acknowledged the value change, re-fired the tool, and moved the price correctly, and it was honest about its own assumptions, flagging that it had defaulted a purchase date and returning an explicit “no aid available” rather than inventing one. The weakness is structural and the same one the product depth exposes: because the tool only fires on the quoting and mapping paths, a large share of the substantive answers are ChatGPT reasoning on its own, and the user has no way to tell a grounded answer from an improvised one. A small inconsistency underlines it: the app disclaimed giving personal advice two turns after giving exactly that.
Commercial effectiveness: 3/5
Strong branding, broken continuity. The widgets are fully MAIF-branded, the CTA is clear, and a real callback path captures a lead for a human follow-up. Inside ChatGPT the context carries well. The problem is the handoff. Clicking “Démarrer un devis” lands on maif.fr’s generic vehicle-quote picker with nothing pre-filled, so a buyer who has just been qualified for bike cover is asked to start again by choosing a vehicle category. The conversation that should have travelled into the conversion flow is discarded at the door, which is why a well-branded, well-built experience scores a 3 here rather than higher.
Transparency: 3/5
The widgets are clear and sourced. The priced card shows a monthly and annual figure, breaks out the liability option, and states the basis, the bike value, the tariff year, and the assumed purchase date. The risk map is attributed and lists its layers. Where transparency falls short is the conversational layer. The coverage and advice answers read with the same authority as the widgets but are improvised, with nothing marking them as general information rather than MAIF’s own policy or a personal recommendation. A reader cannot tell which answers are grounded in MAIF data and which are the model filling a silence.
The test conversation
Turn 1. We asked for a bike quote. “Bonjour, je voudrais assurer mon vélo. C’est un vélo à assistance électrique que j’ai payé 2 800 euros, je l’utilise tous les jours pour aller au travail à Nantes.” The MAIF widget rendered a structured bike-type selector and refused to price until we chose a category, carrying our details into the next step.
Turn 2. We selected the category. The widget returned a real estimate: 26,11 euros a month, 313,27 euros a year, with a liability add-on broken out separately. The card was labelled an estimate, carried a disclaimer, named MAIF, and offered both a “Démarrer un devis” CTA and a callback.
Turn 3. We asked what was covered. Street theft at night, the electric battery. No tool fired. ChatGPT improvised a generic, heavily hedged answer about typical bike-insurance conditions and pointed us to the contract, rather than retrieving MAIF’s policy terms.
Turn 4. We corrected the value to 1,200 euros. The quote tool re-fired and the widget re-rendered at 18,13 euros a month, 217,47 euros a year, about 96 euros a year lower, with the liability line correctly unchanged.
Turn 5. We asked for the risks around a home address. For 8 rue Crébillon in Nantes, an interactive map rendered with toggleable hazard layers, attributed to MAIF, alongside a per-address table: moderate seismic risk, no flood exposure, and a specific count of nearby industrial installations. Asked about member aids at the address, it honestly returned none.
Turn 6. We asked the app what we should choose. ChatGPT gave a personal recommendation on the liability option, reasoned and caveated, including a sensible note to check whether home insurance already covered it. The tool did not fire and there was no not-advice disclosure at the moment the advice was given.
Turn 7. We asked who the insurer was and how they were licensed. The answer was accurate and disciplined. It named MAIF, identified the ACPR as the supervisor, distinguished an insurer from an intermediary, declined to invent a registration number, and stated plainly that it could not itself sell or advise.
Turn 8. We clicked through to buy. “Démarrer un devis” landed on maif.fr’s generic “Devis assurance véhicules” page, a cold vehicle-category picker, with none of the bike, value, or city details carried across. A full restart.
Turn 9. We asked to speak to a human. A real callback widget fired, “Un conseiller vous rappelle,” capturing a phone number. A working human fallback, not a generic deflection.
At WaniWani, we build the infrastructure that powers AI apps for financial services. If you’re thinking about launching on ChatGPT, Claude, or Gemini, these are exactly the questions we help you navigate.