solutions · insurance

Wytness for insurance

Underwriting, claims, fraud. Every AI decision logged in a form your regulator can read.

YOUR RUNTIMEWYTNESS · OPERATORYour AI agentsees claim identifierin clearWytness PII layermint pseudonym · sign event<PII:CLAIM_*>keys never leave yousignedOperator surfaceholds pseudonym onlywytness_envelope_sigagt_merkle_entry_hashSupervising adjusterapproves where policyrequires itState DOI, APRA, EU AI Act exportsigned export · pseudonyms only · tamper-evident

the question every regulator asks

Can you produce the evidence?

The NAIC Model Bulletin (adopted by 20+ state DOIs), the Colorado AI Act, NY DFS Circular Letter 7, APRA CPS 230/234, and the EU AI Act (life and health pricing is Annex III §5(c) high-risk) all ask the same thing of an AI insurance decision: which model version, against which inputs, with which guardrails active, produced this adverse-action notice? The pricing engine has logs, the agent harness has traces, the claims platform has an audit table — none of them, alone, answers that. And shipping any of it to an observability SaaS is itself a disclosure and vendor-risk event.

APRA CPS 230 §47APRA CPS 234 §35FINMA Circular 2023/1NAIC Model Bulletin on AIColorado AI Act (SB21-169)NY DFS Circular No. 7 (2024)EU AI Act Annex III §5(c)GDPR Article 22FATF Recommendation 10

a scenario

A syndicate audit, with and without

A Melbourne insurtech writes home cover through a Lloyd's binder and triages two thousand first-notice-of-loss claims a month with an agent — reading statements, policy schedules, and weather data, auto-progressing simple claims. The syndicate's audit lead asks: for last quarter, show every auto-progressed claim under $5,000, its exact tool calls, the policy version in force, and proof no claimant health information reached your model provider.

Without Wytness that's CloudWatch, an OpenAI record they can't export cleanly, and a claims warehouse of raw statements — no shared identity, no signatures. With it, one signed export: claimant tokens, tool calls, model and policy version fingerprints, auto-progress decision, adjuster flag, chain links — mapped to CPS 230 §47 and EU AI Act Article 12. The audit closes in one cycle, and the same export answers the ACCC a month later.

event.json
Insurance audit record

// Event identity:

id:"evt_01HX9F8K4M6Q2T7P3X9N5R"
time:"2026-05-22T14:02:33.891Z"
agent_id:"agt_claims_triage_v2"
source:"agt://claims-triage-agent/v2"

// What the agent did:

tool_name:"claim.approve"
decision:"allow"
source_format:"wytness_agt_py"Python wrapper
supervising_adjuster_id:"adj_4a91"

// Parameters (PII tokenised before this row was sealed):

data.claim_id:"<PII:CLAIM_a3f9b1c4>"pseudonymised
data.policy_id:"<PII:POLICY_TOKEN_8e2d>"pseudonymised
data.peril:"storm"
data.amount_aud:2480
data.policy_version:"hc-2026-Q2.v3"
data.auto_progress:true
data.adjuster_approved:true

// Integrity proof (3-layer):

agt_merkle_prev_hash:"sha256:91c4e72d..."chain link
agt_merkle_entry_hash:"sha256:c4e7a91f..."SHA-256
wytness_envelope_sig:"MEUCIQDPx8m2..."Ed25519
wytness_envelope_key_id:"key_4b18"

one auto-approved storm claim from the scenario — claimant identity arrives as tokens

the deliverable

What you can prove

  • Which agent produced which underwriting decision: input fields seen, model version, rating-engine output, supervising-actuary flag.
  • That a claims-triage agent stayed inside the perils and policy terms it was scoped to, and never paid out on an exclusion.
  • That a pricing agent's behaviour after a model upgrade matched pre-upgrade behaviour on a held-out sample your appointed actuary signs off on.
  • That no personal data — Medicare numbers, health conditions, claim photographs — left your environment in clear; only tokens reached Wytness.
  • That a state DOI examination asking 'show every adverse-action notice generated by an AI agent in Q3' can be answered in days, not as a discovery sprint.

how the surfaces map

Three surfaces, one product.

/product/ledger

Ledger

Every underwriting, claims, and broker-facing decision signed per event; personal information pseudonymised in your runtime. BYOS keeps the raw log in your tenancy; Evidence Packs map to SOC 2 CC7.2, EU AI Act Articles 12–13, ISO 27001 A.8.15.

/product/registry

Registry

Inventory of pricing models, claims agents, fraud agents, and broker assistants with model-version history and risk classification. Anomaly rules flag scope creep, off-hours adjudications, and decision-distribution shifts after an upgrade that need actuarial re-review.

/product/shield

Shield

Approval gates on declinations, large payouts, and rating-factor overrides, routed to the ServiceNow GRC queue your operational risk team already works.

for your auditor

Frameworks that matter here

SOC 2 Type II

Trust services criteria your reinsurer's vendor team already evaluates.

ISO 27001:2022

Information security baseline expected by every major carrier.

EU AI Act

Life and health pricing is high-risk under Annex III §5(c).

GDPR

Article 22 covers solely-automated underwriting and claims decisions.

APRA CPS 230 / 234

Operational risk and information security — control matrix in progress.

NIST AI RMF 1.0

MEASURE and MANAGE support via the signed trail — in progress.

NAIC Model Bulletin on AI

Framework alignment available on request.

FINMA Circular 2023/1

Framework alignment available on request.

the boundary

What Wytness does not do

  • We are not a policy administration system. Wytness records what your agents did against Guidewire, Duck Creek, or your own PAS; it does not replace the PAS.
  • We are not an actuarial model risk framework. We give your appointed actuary and MRM committee the evidence to apply SR 11-7 or the NAIC Model Bulletin to AI agents; the policy work is theirs.
  • We do not adjudicate fairness. Wytness logs the inputs, outputs, and decisions; whether a pricing pattern constitutes unfair discrimination under your state's law is a legal and actuarial judgement, not a vendor's call.

Questions about your AI risk obligations? Ask us.

We set no cookies. Sign-in and preferences use essential first-party browser storage only — no tracking, advertising, or third-party analytics. Privacy Policy