Worked example: three ways to decline a claim

An insurer wants to triage incoming claims so that straightforward ones settle quickly. Three teams propose three approaches. All three produce a recommendation to fast-track, review, or decline. Only two are AI systems, and the difference changes what has to exist on paper.

Option A: the rules engine

A business analyst writes rules. Claims under CHF 2,000 with a complete document set and no claim in the previous twelve months are fast-tracked. Anything involving a third party goes to review.

Is it an AI system? No. It applies rules a person wrote and can state. Nothing is inferred.

What still applies? A great deal. It is a solely automated decision if no person meaningfully reviews the outcome, so the data-protection duties attach regardless: under Article 21 of the revised Swiss FADP the insurer must tell the person the decision was automated, allow them to state their position, and let them ask for human review. Not being AI is not the same as not being regulated.

Option B: the classifier

A data scientist trains a model on five years of settled claims to predict which will close without dispute. It outputs a probability, and claims above a threshold are fast-tracked.

Is it an AI system? Yes. It infers from data; nobody wrote the rule it applies. That it is a fifteen-year-old technique and never updates in production is irrelevant. Adaptiveness is optional in the definition; inference is not.

What changes? The rule is now a property of the training data. If historical settlements disfavoured a group, the model learns that and the analyst cannot inspect a clause to find it. The organisation needs a record of what data it was trained on, what it was tested against, and how it performs for different groups, because none of that can be reconstructed from the code.

Option C: the language model

A third team feeds the claim narrative and attached documents to a commercial large language model with a prompt asking for a recommendation and a one-paragraph justification.

Is it an AI system? Yes, and it is built on a general-purpose AI model the insurer did not make.

What changes again? Three things. The justification paragraph is produced by the same process as the recommendation, so it reads as a reason but is not the reason; it is a plausible-sounding text generated alongside the answer. The claim narrative, which is personal data and frequently health data, has left the building. And because the insurer chose the purpose, fine-tuned the prompt and put the output in front of customers, the question of whether it has become a provider under Article 25 is live rather than theoretical.

What the comparison shows

The three options are roughly equal in business value and wildly unequal in what they require. Option A needs a documented rule and an appeal route. Option B needs everything in A plus data provenance, fairness testing against thresholds set in advance, and monitoring for drift. Option C needs everything in B plus a privacy assessment of what leaves the organisation, a decision on the vendor relationship, and a policy on whether a generated justification may ever be shown to a customer as a reason.

Note what the risk tracks. Not sophistication. Not cost. It tracks how far the decision has moved from a rule a person can state, and how much of a real person's life rides on it.

The artefact this produces

All three belong in the AI system register, including Option A, which is not AI. A register that only holds AI systems cannot answer the question an auditor actually asks, which is how consequential decisions get made.

As of 24 September 2026.