Know what your AI agent really does.
AIA Guard is developing an assessment and certification programme for AI agents in Swiss healthcare and finance. Clear scope, tested controls, honest limits — and evidence an insurer can one day use.
The agent may prepare an action. Final submission requires the designated person’s approval.
An agent does more than answer.
It may access personal data, call tools and prepare consequential actions. The people who adopt it need to understand where its authority begins and ends.
An assessment makes the basis of a decision readable: what was tested, which controls applied and where the conclusion ends.
How to read a dossierWhat the method examines.
Eight domains connect technical controls to the evidence an independent reviewer needs.
Ownership, intended use, exclusions and a versioned inventory of the assessed agent.
02Data and confidentialityData flows, minimisation, access, retention and protection against unintended disclosure.
03Integration securityUntrusted input, tool permissions, secrets, dependencies and boundaries between tenants.
04ReliabilityRepresentative cases, unsupported output and regression evidence for the declared use.
05Human oversightApproval, escalation and stop mechanisms for consequential actions.
06Transparency and impactAI disclosure, understandable decisions and impacts on affected people.
07OperationsTamper-evident traces, incident response and controlled change after release.
08EvidenceProvenance, reproducibility, assessment boundaries and residual risk.
A review leaves a record.
A bounded record that teams can inspect, challenge and maintain.
- 01
Define the agent and its boundaries
- 02
Collect evidence and run authorised tests
- 03
Independently review the documented record
- 04
Record the decision and its limitations
- 05
Reassess when the agent changes
Built for builders and decision-makers.
For builders
Prepare the agent’s boundaries, controls and evidence before a buyer asks for them.
Prepare for reviewFor enterprises
Ask suppliers for a specific account of authority, evidence and residual risk.
Evaluate a supplierEvidence can support better future decisions.
Every assessment follows our insurability track and produces a policy-ready dossier — evidence shaped for a future insurer’s own review. Cover itself always comes from an insurer’s policy; our job is to make your agent ready to be insured.
See the insurability programPut the method to work.
Practical reading for the people building, buying and reviewing AI agents.
Explore all resourcesHow to evaluate an AI agent before you buy it
A practical path from a polished product demo to a bounded, evidence-based adoption decision.
Voice agents: the questions buyers forget to ask
What to examine when an AI voice agent listens, interprets and acts in healthcare or finance.
Build an evidence pack buyers actually trust
A compact pack that answers buyer questions without exposing secrets or personal data.
Questions worth asking early.
What is assessed?
The proposed method examines a specified agent version, its tools, data, intended use and exclusions.
Does an assessment guarantee safe use?
A result applies to its reviewed scope. Operational responsibility and residual risk remain.
What happens when the agent changes?
Material changes trigger a review of the scope and supporting evidence. The draft also proposes quarterly reassessment.
Is a reviewed agent automatically insured?
No — and that is by design. The review produces a policy-ready dossier for an insurer’s own decision. Cover comes only from an insurer’s policy, never from the assessment.
Bring your actual agent into view.
Tell us what your agent does, what it may touch, and what you need to prove. We reply with honest next steps.
Discuss your agent