A guide for insurtech and fintech founders building AI-driven chat, recommendation, or comparison features

The feature that’s easy to ship, difficult to unwind

Most insurtech teams building an AI-driven chat feature or a “personalised” comparison flow are optimising for conversion, not compliance. That’s the right instinct for product velocity, and the wrong one to leave unchecked, because the FCA’s test for regulated financial advice doesn’t care what you called the feature in your roadmap.

Under the Financial Services and Markets Act 2000 (FSMA), the line between “guidance” and a “personal recommendation” is simple to state and easy to cross without noticing: guidance is general information; a personal recommendation takes someone’s individual circumstances and steers them toward a specific product. Only the second is likely to amount to a regulated activity requiring FCA authorisation, unless an exemption applies.

A chatbot that asks a handful of onboarding questions and then says “based on this, here’s your best option” may amount to a personal recommendation, depending on how the interaction is designed.

Why this hits differently for a startup than for an established broker

Established brokers usually have a compliance function reviewing content after the fact. Early-stage insurtechs often don’t, the product and the “advice surface” are the same thing, built by the same small team, shipped on the same release cycle as everything else. That collapses the normal separation between “what we say” and “what we build,” and it means the regulatory question has to be asked at the design stage, not caught in a later review.

It also means the risk compounds with the thing founders are usually proudest of: traction. A single ambiguous recommendation is a product decision. A thousand users a month going through the same flow is a pattern, and patterns are what supervisory attention finds.

What’s actually at stake

This isn’t a theoretical compliance nuance. The consequences sit in FSMA itself:

• Criminal liability. Carrying out a regulated activity, including giving personal recommendations, without authorisation is a criminal offence under sections 19 and 23 unless appropriately authorised or exempt.
• Unenforceable agreements. Sections 26 and 27 mean agreements made through unauthorised regulated activity can be unenforceable against the customer, meaning a user could walk away from a contract while your business still carries the cost.
• Restitution and disgorgement. Sections 382 to 384 give the courts and the FCA respectively the power to require repayment of profits or compensation for customer losses connected to unauthorised business.

None of this requires intent. A well-designed, well-meaning recommendation feature can cross the line as easily as a careless one, arguably more easily, because good UX is often what makes a comparison feel personal.

Where Consumer Duty adds a second layer

Even a feature that stays clearly on the “guidance” side of the FSMA line isn’t automatically in the clear. Consumer Duty requires firms to act to deliver good outcomes for retail customers and be able to demonstrate how their products, communications and customer journeys support those outcomes. A feature that passes the personal-recommendation test can still fail a Consumer Duty review if nobody checked what the recommendation logic is actually optimising for.

What to check before you ship

A few practical questions worth running through your team before (not after) a feature goes live:

• Does the feature use individual inputs to narrow toward one answer? If yes, you’re closer to a recommendation than guidance, regardless of the UI framing.
• Who signed off on the design, not just the output? Where the Senior Managers & Certification Regime (SM&CR) applies, firms should ensure there is clear accountability for the governance and oversight of AI-enabled customer journeys.
• Can you show what a specific user was shown and why? Consumer Duty and complaint handling both require decision-level records, not just a policy document describing the system in general.
• Has this been reviewed at prototype stage, or only after it scaled? The cheapest time to answer this question is before the feature has users; the most expensive time is after a complaint or a supervisory request surfaces it for you.

Quick answers

Does my chatbot need FCA authorisation? 

Only if it gives personal recommendations, content that uses someone’s individual circumstances to steer them toward a specific product. General information and comparison tools that don’t do this can typically operate as guidance without authorisation, but the design details matter more than the labelling.

Does it matter that an AI model generated the recommendation, not a person? 

No. FSMA’s test looks at the effect of the content, not who or what produced it.

What’s the single highest-risk feature type for early-stage insurtechs?

Personalised comparison or “best fit” recommendation flows, precisely because they’re often built to feel helpful and specific, which is what makes them recommendations in the regulatory sens