Does insurance secrecy apply when an employee uses ChatGPT?

Yes. Art. 35 VAG places employees of supervised insurers under confidentiality obligations analogous to banking secrecy, and social health insurers are additionally bound by Art. 84 KVG. A prompt containing client or insured data sent to an external LLM provider is a potential breach. NativeAI Guard anonymizes client identifiers, diagnoses and policy numbers before the prompt leaves the organization.

Is insurance pricing AI high-risk under the EU AI Act?

From August 2026 the EU AI Act classifies AI used for risk assessment and pricing in life and health insurance as high-risk, with binding logging, human-oversight and governance duties. Employment-related AI decisions are also high-risk. NativeAI Guard contributes the logging and oversight gates these obligations require.

How can an insurer explain an AI content-blocking decision to a regulator?

Every NativeAI Guard decision is logged together with the policy and the data category that triggered it, and human overrides are recorded with an audit trail. Anonymized-storage mode keeps the full audit trail without exposing raw personal data, which is the balance social health insurers need under Art. 84 KVG.

Can AI claims agents leak insured data?

Yes, in both directions. Claims agents ingest submissions, medical reports and correspondence, which can carry hidden prompt-injection instructions, and their responses can expose another claimant's health data. NativeAI Guard inspects what the agent ingests and what it responds with, blocking injected instructions and cross-claimant data leaks.

Industries / Insurance · insurers & health insurers

AI protection for insurers, built for secrecy under Art. 35 VAG.

Art. 35 VAG places insurance employees under secrecy obligations analogous to banking, and the EU AI Act classifies risk assessment and pricing for life and health insurance as high-risk from August 2026. The same workflow triggers both.

DEADLINESNOW · Art. 35 VAG · DSG in forceJAN 2025 · DORA in forceAUG 2026 · EU AI ACT HIGH-RISK (risk assessment & pricing)

The AI governance framework for insurance

Six obligations. Here's our part, honestly.

We mark what NativeAI Guard solves for the channels we protect, where it contributes, and what stays with your processes.

Art. 35 VAGInsurance secrecy, confidentiality obligations for employees of supervised insurers.✓ We solve · DLP + anonymization

Art. 84 KVGConfidentiality of insured persons' data for social health insurers, distinct from and additional to Art. 35 VAG.✓ We solve · DLP + anonymization

Swiss DSG / GDPRHealth data is a special category of personal data: data minimization, records, cross-border limits, personal liability.✓ We solve · anonymization + RoPA

EU AI ActRisk assessment and pricing for life/health insurance are high-risk from Aug 2026, plus AI for employment decisions.◐ We contribute · logging + oversight

FINMA 2018/3External LLM use is an outsourcing relationship for insurers too, documentation and monitoring required.◐ We contribute · monitoring artifact

DORAApplies to EU-facing insurers, ICT third-party register and incident reporting.◐ We contribute · logging + RoI evidence

Seen at Swiss insurers

The claims summary

An underwriter pastes a customer's medical history into an LLM to summarize a claim. Health data, a special category of personal data, just left the perimeter under the terms of service of a consumer product.

NativeAI Guard: [PERSON], [DIAGNOSIS], [POLICY-NO], the summary still works, the health data stays in the company.

The explainability question

"How do we justify the decision of a fuzzy LLM classifier to the regulator?", the objection we hear from every insurance CISO. A block you cannot explain is a block you cannot defend.

NativeAI Guard: every decision is logged with the policy and data category that triggered it; human override with audit trail. Anonymized-storage mode keeps the full audit trail without exposing raw personal data, the control health insurers need under Art. 84 KVG.

The first claims agents

Insurers are already piloting AI agents that read claim submissions, medical reports and correspondence to triage or pre-assess a claim. Those very documents are untrusted content that can carry a hidden prompt injection, and the agent's response can just as easily carry another claimant's health data out.

NativeAI Guard: inspects what the claims agent ingests and what it returns, catches injected instructions and blocks data leaks across claimants. NativeAI Guard Agent DLP & Firewall →

Writing policies in natural language instead of regex, that is the key differentiator of NativeAI Guard.

SECURITY ENGINEER · SWISS HEALTH INSURER

Policies, managed

The rulebook for insurers is already written. You adapt it instead of inventing it.

A base rulebook is already written and active from day one: the rules from the Swiss DSG, GDPR and the EU AI Act, plus the requirements your industry is specifically subject to. It covers the standard cases without you writing a single line. Security and compliance practitioners review it, and we keep it current as the law changes. On top of it sit the rules that apply only to you: your own data structures, your internal requirements, your exceptions.

Insurance packArt. 35 VAGArt. 84 KVG · health insurersSwiss DSG · GDPREU AI Act, pricingFINMA 2018/3

During onboarding we tune the pack to your own data structures and internal rules, working from the compliance documents you already have.

Common questions

Frequently asked questions

Can the anonymized fields be restored in the model's answer, so the response is still usable?

Re-identification on the return path is in active development and does not ship today; it sits near the front of our development roadmap because several customers have asked for it. What works today: anonymization preserves context. Identifiers become typed placeholders, so the model still produces a useful, correctly structured answer that your people can complete internally.

How do we justify an AI-based blocking decision to a regulator? An LLM classifier looks fuzzy.

By making the decision reconstructable rather than by claiming the model is infallible. Every decision records which policy applied, which data category was detected, the content that triggered it, the action taken and the model version that made the call, so it can be replayed after the fact. Deterministic rules run alongside the semantic layer, so the obligations you must never miss are not left to a probabilistic judgment. And a human override with a logged justification means the final decision has an accountable owner. Regulators generally ask whether a control is documented, consistently applied and auditable, not whether it is deterministic.

We hold health data, banking data and insurance data under one roof. Can policies differ by business line?

Yes, and they should. Policies are scoped per group, business line or team, so Art. 35 VAG, Art. 84 KVG and the rules for health data can be enforced differently where they apply, rather than flattening everything to the strictest common denominator and blocking work that is perfectly legitimate.

Protect your employees' AI usage from Art. 35 VAG exposure, with a pilot phase that we run for you.

An insurance policy starter pack, plus a report for management on data leaks, shadow AI, AI usage and model costs. Anonymize mode lets underwriters and claims teams work at AI speed while customer and health data stay inside the company.