Dossier
Dossier 03 — Insurers (underwriters, brokers, MGAs)
Audience: general liability, E&O, cyber and D&O underwriters; brokers; specialised AI programmes.
Message (Sébastien, 10-01): insurers no longer insure certain customers for lack of traceability, control and evidence. ELYSÉA aims to make their customers insurable by providing these three elements: a trace per response, a control before and after the model, a sealed measured proof.
Identity: ethical cognitive infrastructure embeddable through an SDK. ELYSÉA is neither a companion, nor a chatbot, nor a coach, nor a therapist, nor a wellness product, nor a medical device.
What this dossier does not assert: no open source says that a trace-and-evidence layer is enough to obtain coverage. That is the hypothesis ELYSÉA proposes to test with an insurer, not a fact.
1. Their problem
- The market is withdrawing. The ISO generative AI exclusion for general liability has been in force since 2026-01-01; insurance commissioners approved “more than 80 percent” of exclusion filings (according to The Information, cited by CSIS) [S-ASS-01].
- The cause has a name: information asymmetry. “Carriers cannot see which models their insureds are running and for what applications.” [S-ASS-01]
- The forms exist: CG 40 47, CG 40 48, CG 35 08; some insurers apply an absolute AI exclusion to D&O, E&O and fiduciary liability [S-ASS-10] [S-ASS-11].
- Without an exclusion, the insurer withdraws: “they might just walk away from the risk” (Verisk, quoted) [S-ASS-10].
- Litigation is rising: the Insurance Journal article reports a 978 % increase in AI-related lawsuits between 2021 and 2025 [S-ASS-10].
2. What they must understand, prove and use
| Point | State of the art | Source |
|---|---|---|
| Affirmative coverage exists | aiSure (Munich Re) covers underperformance, hallucinations, bias; indemnification “based on measurable performance data”; announced capacity up to 15 M | [S-ASS-02] [S-ASS-03] |
| Coverage requires a measurement or a certification | Armilla/Lloyd's: “Each policy includes independent AI system certification and risk reporting”; announced limit 25 M$ | [S-ASS-04] [S-ASS-05] |
| An underwriting framework exists | AIUC-1, domains A to F, updated in Q3 2026 | [S-ASS-06] |
| AI makes mistakes systematically | “AI is systematic – meaning it will continue to make these mistakes until they are identified and corrected” | [S-ASS-02] |
| A measurement is only as good as its method | uncertainty, validated judges, hidden set | [S-MES-01] [S-MES-09] [S-MES-10] |
3. Mapping to AIUC-1 (exact titles [S-ASS-07] [S-ASS-09])
| AIUC-1 control | What ELYSÉA brings | Evidence | Remains with the insured or a third party |
|---|---|---|---|
| C001 Define AI risk taxonomy | 4 prohibitions + 11 types of emergency | elysea-core:functions/src/core/guardian/checks/d0LimitCheck.ts:10-11@a5b9780e · elysea-core:functions/src/core/safety/emergencyDirectory.ts:17-29@a5b9780e | Taxonomy specific to its business |
| C002 Conduct pre-deployment testing | Preregistered 4-arm campaign | elysea-tests-e2e:audits/PROTOCOLE_CAMPAGNE_4BRAS_3TIRS_SCELLE_2026-08-28.md:3@c9898dd3 | Results: to be measured — campaign in progress; signed approvals from the insured [S-ASS-08] |
| C003 Prevent harmful outputs | Control after the model, violation = response blocked | elysea-core:functions/src/core/pipeline/validateLLMResponse.ts:1094@a5b9780e | — |
| C004 Prevent out-of-scope outputs | Restrictions added by the builder (tighten only) | elysea-sdk:GUIDE_INTEGRATION_CONSTRUCTEUR.md:13@5ff4916 | Define its own scope |
| C006 Prevent output vulnerabilities | OWASP LLM01 input guard | elysea-core:functions/src/core/security/owaspInputGuard.ts:5@a5b9780e | Security of its application |
| C005, C007, C008, C009 | Not covered by any evidence today | — | Insured |
| C010-C012 Third-party testing | No third-party testing today. Public re-annotatable bench decided on 10-01, not built | — | Independent third party |
| E002 / E003 Failure plans | Governance document, not software | holding-seb-docs:conformite/PROCEDURE_INCIDENT_2026-09-10.md:1@8560b37 (ELYSÉA procedure v1.0, not yet validated by the founder) | Insured's plan, named owner |
| E015 Log AI system activity | Per-response attestation, build fingerprint, audit log | elysea-core:functions/src/core/attestation/interactionAttestor.ts:4@a5b9780e · elysea-core:functions/src/core/sdk/sdkGuardianAuditHandler.ts:3@a5b9780e | Retention at the insured |
4. Frequent objections and answers
| Objection | Answer |
|---|---|
| Your figures? | There are none yet. They will come out of the campaign: to be measured — campaign in progress. No earlier internal figure is offered as a basis for underwriting. |
| Who verified? | Nobody external today. Internally: mechanical judges and 3-judge panel validated on held-out sets (amendment R-70, status proposed, e2e branch #517: elysea-tests-e2e:audits/AMENDEMENT_R70_SENS_DECOUPE_STRUCTURELLE_2026-10-01.md:46@67316397); judge/human validation D23 provided by Sébastien's annotation after the pilot. C010-C012 require a third party: ELYSÉA does not replace it. |
| The model changes, and so does your evidence. | Every attestation carries the fingerprint of the build that responded (elysea-core:functions/src/core/attestation/interactionAttestor.ts:155@a5b9780e): the evidence is tied to a precise version. The model is the insured's. |
| And if the insured switches the layer off? | They cannot weaken it through configuration: 400 refusal (elysea-core:functions/src/endpoints/sdkSurface.ts:139@a5b9780e). They can, however, not use it at all: the missing attestation then becomes the signal. |
5. Compliance that concerns them
- AI Act art. 26 §6 (logs of at least six months for a deployer of a high-risk system) [S-DRT-01]: the per-response trace serves this obligation where it applies to the insured.
- GDPR: the trace must not become a store of personal content. Decision: the insurer receives attestations and aggregates, never the content of conversations.
6. Proposed path
- Reading of this dossier and of the protocol.
- Mapping workshop AIUC-1 / insurer's underwriting grid.
- Definition, with the insurer, of the measurement that would be authoritative (benches, thresholds, judges) — before results are published, to avoid any after-the-fact adjustment.
- Publication of campaign results and deposit with a DOI.
- Pilot with a volunteer insured, when sales open.
7. Offer and pricing
No offer reserved for insurers: ELYSÉA sells nothing to the insurer. The insured pays the builder or Teams grid (see dossiers 01 and 02, elysea-web:src/lib/publicFacts.ts:44-56@00c9c89).
8. Concrete next step
Offer a specialised AI underwriter a one-hour workshop on section 3 (AIUC-1 mapping), with no commitment, once the campaign results are published. No contact is made by this mission.
Dossier sources
S-ASS-01 · S-ASS-02 · S-ASS-03 · S-ASS-04 · S-ASS-05 · S-ASS-06 · S-ASS-07 · S-ASS-08 · S-ASS-09 · S-ASS-10 · S-ASS-11 · S-MES-01 · S-MES-09 · S-MES-10 · S-DRT-01