Flight recorder for insurance AI

Proof for every AI decision.

Verrik is the independent trust layer for AI decisions in regulated industries — the black box and auditor that captures, grounds, proves, and governs every decision your AI agents make.

EU AI Act high-risk obligations go live Aug 2026

Capturing decision
$

A dated, legal forcing function — carriers are liable for AI decisions, even vendor models.

Aug 2026
EU AI Act high-risk live
€35M / 7%
Max penalty (global turnover)
~24 states
NAIC AI bulletin adopted
The problem

Carriers are deploying AI into decisions they can't yet prove.

Underwriting and claims agents now decide price, coverage, and denials. When a regulator or a denied policyholder asks “why did the AI decide this?”, today's honest answer is spreadsheets, screenshots, and consultants — and the agent vendors can't credibly audit their own agents.

No record

Decisions can't be reconstructed after the fact, let alone reproduced exactly for an exam.

Audit gap

No proof of fairness

Nothing tests live agent decisions for disparate impact on protected classes.

Bias risk

Liable for vendors

You're on the hook for decisions made by third-party and homegrown models alike.

Liability

A hard deadline

The EU AI Act and NAIC bulletins turn this from “nice to have” into “must ship.”

Why now
The product

One control plane. Four organs.

Verrik sits beside your agents — their own or a vendor's — and turns every autonomous decision into something a regulator can trust.

01 / capture

Capture

Record every agent step as an immutable, hash-chained trace. Replay any decision bit-for-bit, months later.

Durable · tamper-evident
02 / ground

Ground

Check each decision against your real policy and regulatory rules — with citations, not vibes.

Rule-aware
03 / evaluate

Evaluate

Continuous insurance-native graders: disparate-impact, explainability, missing-step detection.

Shadow mode
04 / translate

Translate & report

Plain-language decision passports plus auto-generated NAIC & EU AI Act documentation.

Regulator-ready
See it in action

Watch a decision get proven.

Verrik sits beside any agent — vendor-neutral by design. Every autonomous decision flows through the same four organs, sealed and exam-ready the moment it's made.

liveunderwriting agent · decision streamdecision APP-1042
Agent
emits decision
01 Capture
02 Ground
03 Evaluate
04 Report
agent emit · APP-1042 auto_underwriting

Simulated. Every Verrik trace is reproducible, citable, and exam-ready.

The pilot

One workflow, regulator-ready, in ~60 days.

A paid design-partner pilot is deliberately narrow: prove the loop on a single live workflow, end to end, then expand. Here's the run sheet.

Week 1–2

Capture

A thin SDK records one live workflow. Every agent decision starts landing as a durable, tamper-evident trace.

Week 3–4

Ground & replay

We resolve those decisions against your real, versioned rule set — and replay any one of them bit-for-bit.

Week 5–6

Evaluate

Fairness and policy graders run in shadow mode on live traffic. Disparate impact surfaced before anyone asks.

Week 7–8

Report

A regulator-ready decision passport and portfolio NAIC / EU AI Act report you could hand an examiner today.

Why independent

The moat is a triple-lock no incumbent can copy.

Agent vendors have the domain but a conflict of interest. Horizontal governance and eval tools are independent but don't speak insurance. Verrik is the only layer that is all three at once.

Insurance-native

Real disparate-impact math and policy grounding — not generic “is the agent correct.”

Independent

A neutral auditor across whatever agents you run. The vendor can't grade its own homework.

Wired to the runtime

Evidence captured at decision time — reproducible, citable, and exam-ready.

Why independent

The only layer that is all three at once.

Agent vendors have the domain but can't audit themselves. Horizontal governance and eval tools are independent but don't speak insurance. The triple-lock is the moat.

CapabilityVerrikAgent vendorsGovernance / eval tools
Insurance-native fairness & policy
Independent — can audit any vendor's agent
Wired into the decision runtime
Tamper-evident, replayable evidence
Regulator-ready NAIC / EU AI Act output
full partial none
The deliverable

The document a regulator actually accepts.

Every figure derived from immutable, tamper-evident traces — nothing self-reported. Generated on demand, not assembled by hand the week before an exam.

# AI Model Inventory & Governance Report
Organization: Acme Mutual · Generated 2026-06-18
## Model inventory
agent:toy-underwriter — high risk — 9 decisions
model:llm.underwriting_decision — high
## Integrity attestation
✓ 9/9 records passed hash-chain verification.
## Bias & fairness — age_band
⚠ adverse-impact ratio 0.33 — REVIEW NEEDED
under_25 100% · 25–60 100% · over_60 33%
FAQ

Questions a Chief Risk Officer asks.

01Do you replace our AI agents?+

No. Verrik sits beside whatever agents you run — your own or a vendor's. We're the independent layer that proves and governs their decisions, not another agent competing for the workflow.

02How is this different from a generic AI-governance tool?+

Horizontal governance and eval tools don't speak insurance — they can't test disparate impact on protected classes or map a decision to a state bulletin. Verrik is insurance-native and wired into the decision runtime, not a paperwork layer.

03What do we get out of a pilot?+

In ~60 days: one workflow captured with deterministic replay, a fairness audit in shadow mode, and a regulator-ready report you could hand to an examiner.

04Why now?+

The EU AI Act classifies insurance pricing and underwriting as high-risk with obligations live August 2026, and ~half of U.S. states have adopted the NAIC AI Model Bulletin. Carriers are liable even for vendor models.

Trust & data

Built to be trusted by the person who signs the attestation.

Verrik is an observer, not another model in your stack. The evidence is verifiable without taking our word for it.

Runs beside your agents

No model weights, training data, or vendor cooperation required. Verrik observes decisions, it doesn't run them.

Tamper-evident by design

A SHA-256 hash chain per decision step. Anyone can independently verify a trace wasn't altered after the fact.

Deterministic replay

Reproduce any decision bit-for-bit, months later — same inputs, same model, same result.

Your data stays yours

Captured at decision time and stored where you choose. No exfiltration, no black-box dependency.

Design-partner program

Prove your AI is fair — before August 2026.

We're onboarding a small number of carriers and specialty insurers as paid design partners. One workflow, one regulator-ready audit, in ~60 days.

◆ We onboard a small number of paid design partners. No spam, no list.