AstérismeApplied AI consultancyHumans lead. AI assists.
Insurance

Four insurance processes where the AI prepares and the handler decides.

Each sheet states what the AI does, what stays with your teams, how it is tested on your own files before being used, and what makes it stop. These are typical examples, not client results.

The principle

Settlement stays with the handler.

The AI checks claim notifications and prepares replies. It decides no settlement and prices no policy. Every process has an error threshold written down before the test. If it is not met, the process stops.

FP-10Customer relations

Insurance claims

Claim formAttachmentsAIChecks and summarisesClaims handler decidesClaim forwarded
The AI
Checks and summarises.
Check
Claims handler decides.
Error to avoid
An incomplete claim forwarded.
Test before use
Past claims: completeness judged by the AI compared with the handler's judgement.
Excluded
The decision to pay out.
Where
Dedicated server, in Belgium
What we measure
Time per notification. Files passed on incomplete.
Threshold, set before the test
Every file judged incomplete by the handler is found in the test sample.
Stop criterion
One file passed on incomplete in service: back to testing.
Framework
The settlement decision stays with the handler. Health data never on a public AI.

Insurance

FP-05Legal and compliance

Contract review

Contract receivedChecklistAIFlags clauses to checkLawyer decidesAnnotated contract
The AI
Flags clauses to check.
Check
Lawyer decides.
Error to avoid
A risky clause missed.
Test before use
Contracts already reviewed: clauses flagged by the lawyer compared with the AI's.
Excluded
The decision to sign.
Where
On your premises, local
What we measure
First-reading time per policy. Clauses recorded.
Threshold, set before the test
Every clause flagged by the legal officer in the test policies is found.
Stop criterion
A risk clause missed in service.
Framework
The decision to sign stays with the insurer.

Insurance

FP-11Customer relations

Customer replies

Client questionYour documentsAIDrafts a sourced replyStaff member validatesReply sent
The AI
Drafts a sourced reply.
Check
Staff member validates.
Error to avoid
An invented answer, not based on your documents.
Test before use
Real questions: each reply checked against the source it cites.
Excluded
Complaints and disputes.
Where
Dedicated server, in Belgium
What we measure
Response time. Replies sent without correction.
Threshold, set before the test
Every reply cites the clause or document it comes from.
Stop criterion
An invented reply, outside the general conditions.
Framework
Complaints excluded: they follow your complaints procedure.

Insurance

FP-14Legal and compliance

AI tools register

Tools in useAIPre-fills the registerCompliance officerRegister up to date
The AI
Pre-fills the register.
Check
Compliance officer.
Error to avoid
A tool used outside the register.
Test before use
Inventory compared with the tools actually installed and used.
Excluded
Choice of suppliers.
Where
Dedicated server, in Belgium
What we measure
Tools recorded. Tools found outside the register.
Threshold, set before the test
The register covers every tool actually installed on the date of the test.
Stop criterion
A tool used outside the register.
Framework
AI Act: register of uses and documented training under Article 4.

Insurance

Step validated by a person. The large star is the step entrusted to the AI.

Hosting

Policyholder data stays with you.

Claim notifications contain personal data, sometimes health data. It stays on a dedicated server in Belgium or on your premises, never on a public AI without your written agreement.

Framework

High risk stays out of scope.

The AI Act classifies as high risk the systems used for risk assessment and pricing of natural persons in life and health insurance (Annex III, point 5(c)). Our sheets price nothing and leave every decision to the handler. For Article 4 of the AI Act, training your teams is a documented measure, with a named certificate.

Course

Eight weeks, three to five processes.

A free thirty-minute diagnostic. Then two weeks of context and training, two weeks reviewing processes with the people who run them, and four weeks putting three to five processes into service, one at a time. A month of adjustments follows. Price on quotation, after the diagnostic.

Thirty minutes on one of your processes.

By video call or at your premises. We say plainly whether the AI gains anything there.

Book a free diagnostic