Why ChatGPT gets it wrong when asked to select the best travel insurance

Ask ChatGPT which travel insurance to buy for a two-week family holiday in Florida and an answer arrives in seconds. A shortlist of brands, a few standout features, perhaps a neat comparison table. It reads like the work of someone who has done their homework.

In most cases, nobody has.

Confident by design

General-purpose AI assistants are built to be helpful, so they answer. The trouble is that insurance questions slide very easily from information into something that looks like a personal recommendation, and the assistant rarely signals where that line sits. Choosing cover properly depends on understanding circumstances and needs. A general chatbot has no structured process for doing that, yet it will readily suggest what to buy.

Thin foundations

Look at what those suggestions are built on and the picture changes. The assistant draws on whatever public information is easiest to find: insurer product pages, marketing copy and the star ratings on policy grading websites.

That material is designed to highlight headline limits and selling points. It says nothing about the exclusion buried in the general conditions, whether an excess applies per person or per claim, or the condition requiring the emergency assistance service to be contacted before treatment. The suggestion sounds informed, but it is marketing, rearranged.

It folds when challenged

Push back and the weakness shows. Question whether the suggested policy excludes a particular situation, or share a more detailed reading of the wording, and the assistant will usually agree, apologise and change its answer.

This is not open-mindedness. The original answer had no fixed basis, so there was nothing to defend. Guidance that reverses under the mildest pressure was never really guidance.

Asking it to do the analysis does not fix it

The natural next step is to upload a policy document and ask for a thorough review. The result will look thorough, with headings, bullet points and a confident summary.

But policy wordings are long and complex. They rely on defined terms, cross-references, schedules and endorsements, and the meaning of one clause often depends on another several pages away. Without a framework for reading them, the assistant picks out whatever seems important, and its sense of importance is shaped by the same marketing and rating sites as before.

What comes back is a patchwork. Some points are relevant, some are not. Some are correct, some misread a defined term or miss a condition entirely. All of it is delivered in the same assured tone and packaged as a detailed, accurate analysis, with no way of telling which parts can be trusted.

No understanding of the circumstances

The right cover depends on the circumstances of the trip: destination, duration, who is travelling, the cost of the trip, and what protection may already exist elsewhere. A general assistant has no consistent way to establish these. It either asks nothing, asks different questions each time, or quietly fills the gaps with assumptions.

How Prompted does it differently

Prompted works inside ChatGPT too, but its analysis does not come from the model's general knowledge. It comes from a methodology built specifically for insurance, in which every trip is assessed as a scenario. There are four foundations.

Every policy is read the same way. Policy wordings are read into a consistent structure, so the same questions are asked of every policy. Exclusions, conditions and limits are captured alongside the headline benefits, which means policies can be compared like for like rather than on whatever each insurer chooses to promote.

Needs are extrapolated. The characteristics of each trip determine which situations are genuinely relevant. A family holiday in the United States, for example, puts medical expense limits and cancellation cover front and centre in a way that a weekend in Paris does not.

Policy analysis and comparison isn't always informative. Rather than starting with what a policy says about itself, Prompted starts with how things actually go wrong on a trip and how claims arise, then maps those situations to the parts of each policy that would respond. The analysis is anchored in outcomes that matter.

The result is distilled. The cover items relevant to a trip are brought together into clear, focused analysis: what is covered, what is not, and what conditions apply, with each point traceable back to the policy wording.

The outcome is structured, consistent and effective. The same scenario produces the same analysis every time. It does not shift under pushback, and it is not padded with points that have nothing to do with the scenario. The result is a clear view of how policies compare on the things that matter, as a basis for an informed decision.

The bottom line

General-purpose AI is remarkable, but insurance is not a subject where a confident, fluent answer is the same as a correct one. What matters is structure: a consistent way of reading policies, a proper understanding of each scenario, and analysis grounded in how claims really work. That is what Prompted is built to provide.

Prompted Limited (company number 16648304), registered office 27 Old Gloucester Street, London, England, WC1N 3AX, is an Appointed Representative of Innovative Risk Labs Limited, which is authorised and regulated by the Financial Conduct Authority (FRN 609155).