AIRS v1.1 · Open Standard Methodology

How AIRS measures AI insurance readiness

AIRS v1.1 · May 2026

The methodology behind the AI Insurance Readiness Score: domain architecture, the five-point maturity scale, the evidence model that underpins defensibility, the scoring approach, and the framework lineage that situates AIRS alongside NIST AI RMF, ISO/IEC 42001, the EU AI Act, and the NAIC Model Bulletin.

From the Editors
Methodology Preamble

AIRS v1.1 was constructed to satisfy a single institutional requirement: produce, from observable evidence, a score that an underwriter can defend, a regulator can cite, and a counterparty can replicate. This page explains how each design choice in the specification serves that requirement.

Where the specification is the canonical reference, this methodology page is the institutional explanation: why five domains rather than three or seven, why a five-point maturity scale rather than a binary checklist, why the evidence model insists on artifacts rather than attestations, and how the composite score and tier mapping translate maturity into underwriting signal.

01

Domain Architecture

AIRS evaluates five thematic areas of AI security risk, each composed of five discrete factors. The domain set was chosen for disjoint coverage of the failure modes that drive insured loss in AI-dependent systems: model behavior, output liability, dependency exposure, regulatory posture, and continuity.

Domain 1 Model Integrity Weight: 25% · 5 factors

Whether the model itself behaves as the operator believes. Anchors the framework because every downstream domain inherits its assumptions from how the model was trained, versioned, and monitored.

  1. 1.1 Training data provenance
  2. 1.2 Adversarial robustness testing
  3. 1.3 Model versioning & rollback
  4. 1.4 Poisoning detection
  5. 1.5 Drift monitoring
Domain 2 Output Liability Weight: 20% · 5 factors

Whether the operator can stand behind what the model produces. Maps directly to the loss vectors most frequently cited in early AI insurance claims: hallucination, bias, unauditable decisions, and unattributed content.

  1. 2.1 Human-in-the-loop oversight
  2. 2.2 Hallucination controls
  3. 2.3 Bias & fairness testing
  4. 2.4 Output auditability
  5. 2.5 Content provenance
Domain 3 Supply Chain Security Weight: 20% · 5 factors

Whether the operator understands and controls the dependencies the model relies on. AI systems inherit risk from foundation-model providers, API surfaces, and sub-processors at a depth not contemplated by traditional third-party risk programs.

  1. 3.1 Model card & vendor documentation
  2. 3.2 Third-party AI risk
  3. 3.3 API key governance
  4. 3.4 Foundation-model dependency mapping
  5. 3.5 Sub-processor controls
Domain 4 Regulatory Compliance Weight: 20% · 5 factors

Whether the operator's posture aligns with the AI governance frameworks that carriers, reinsurers, and regulators are increasingly treating as the baseline of insurability. Designed to crosswalk cleanly to NIST AI RMF, ISO/IEC 42001, the EU AI Act, and emerging state AI laws.

  1. 4.1 NIST AI RMF alignment
  2. 4.2 ISO/IEC 42001 conformance
  3. 4.3 EU AI Act readiness
  4. 4.4 State AI law monitoring
  5. 4.5 GDPR/CCPA AI-specific controls
Domain 5 Systemic Resilience Weight: 15% · 5 factors

Whether the operator has planned for the days when AI fails. Recovery architecture, drilled response plans, and continuity of AI-dependent workflows are the difference between an incident and a market-moving loss event.

  1. 5.1 AI incident response plan
  2. 5.2 Recovery time objectives
  3. 5.3 AI-specific tabletop exercises
  4. 5.4 Inter-system failure modes
  5. 5.5 Business continuity for AI-dependent workflows

Twenty-five factors total. Each evaluated on the five-point maturity scale defined in the next section.

02

The Five-Point Maturity Scale

Each of the twenty-five factors is rated on a five-point maturity scale rather than a binary checklist. Maturity scoring rewards the operational reality that a control's value scales with the discipline of its implementation: a documented policy is not the same as a tested process, and a tested process is not the same as a continuously monitored one.

1
Initial

No formal control. Activity is ad-hoc, informal, or absent.

2
Defined

Policy exists. Documented intent without consistent execution.

3
Implemented

Control is in place and operating, with evidence of execution.

4
Measured

Control is tested, instrumented, and produces audit-grade output.

5
Optimized

Control is continuously monitored and improved; failures are surfaced and remediated.

Maturity levels translate to factor scores of 1.0 through 5.0. Domain scores are the unweighted mean of their five factors.

03

The Evidence Model

The defensibility of an AIRS score rests on the artifacts that support it. Self-attestation alone is not sufficient evidence at any maturity level above two. The evidence model recognizes three categories of admissible artifact, and assessments must cite at least one from each category to claim a maturity level of three or higher.

Category I

Documentation

Written policies, standards, model cards, data sheets, vendor contracts, and governance charters that specify the intent and design of a control. Documentation establishes that a control was contemplated and approved, but does not establish that it operates.

  • Approved AI risk policy
  • Model card with intended-use statement
  • Sub-processor inventory with DPAs
Category II

Configuration

Technical artifacts that show the control is wired into the operating environment: code-level guardrails, infrastructure-as-code definitions, monitoring rules, IAM policies, and key-management configurations. Configuration evidence proves intent has been translated into the running system.

  • Hallucination guardrail rules in CI
  • API key rotation policy in IaC
  • Drift-monitoring alert thresholds
Category III

Observed Behavior

Records that demonstrate the control has produced outcomes under real conditions: incident logs, tabletop exercise reports, audit findings, red-team results, drift alerts that fired and were resolved, and remediation tickets that closed against SLA. Observed behavior is the highest grade of evidence.

  • Quarterly tabletop exercise after-action report
  • Adversarial red-team findings & closures
  • Incident response post-mortems with metrics

The evidence model is the single most consequential design choice in AIRS. It is what makes a score citable in an underwriting file and admissible in a regulatory review.

04

Scoring Approach

AIRS produces a single composite score on a 0–100 scale. The scoring formula is intentionally legible: a regulator, auditor, or counterparty can reproduce it from the spec without proprietary tooling.

Composite Formula

Composite = Σ (Domaini × Weighti) × 20

Where Domaini is the unweighted mean of its five factor scores (1.0–5.0), and weights sum to 1.00 across the five domains.

Worked example

An entity scoring an average of 4.0 across all five domains produces a domain-weighted mean of 4.0. Multiplied by 20, the composite score is 80 — the threshold for Tier 1, AI Insurance Ready. An entity averaging 3.25 across all domains produces a composite of 65 — the threshold for Tier 2, Conditionally Insurable. Scores below 30 fall into Tier 5, Uninsurable, where the entity is not eligible for coverage at the current assessment and is provided a roadmap to reach a higher tier. The minimum theoretical score is 20, achieved when every factor sits at maturity level 1.

See Section 7 of the specification for the complete formula, edge-case treatment, and the conformance criteria that govern how factor scores are derived from evidence.

05

Tier Mapping

Composite scores map to five rating tiers. The tier structure is the methodology's translation layer between maturity assessment and underwriting decision — designed to align with the decision architecture of institutional carriers and reinsurers.

Score Range Tier Classification Underwriting Signal
80–100 Tier 1 AI Insurance Ready Standard underwriting; preferred terms
65–79.99 Tier 2 Conditionally Insurable Coverage available with conditions or sub-limits
50–64.99 Tier 3 Elevated Risk Coverage requires remediation milestones or premium loading
30–49.99 Tier 4 Remediation-Track Conditional binder pending defined remediation
0–29.99 Tier 5 Uninsurable Not eligible for coverage at this assessment

The theoretical minimum AIRS score is 20 (all factors at maturity level 1). Tier thresholds are fixed in the specification and do not vary by industry, jurisdiction, or assessor. Domain floor rule (v1.1): T1 requires a minimum domain score of 3.5; T2, 2.5; T3, 1.5. A subject entity cannot reach a tier whose domain floor is unmet on any single domain, regardless of composite. Boundary scores resolve to the higher tier.

06

Defensibility & Framework Lineage

AIRS is intentionally derivative. It does not propose a new model of AI risk; it operationalizes the consensus that the established AI governance frameworks have already produced, and translates that consensus into an underwriting-grade signal. The framework lineage is the methodology's defensibility argument.

Source Framework

NIST AI Risk Management Framework

AIRS Domain 4 (Regulatory Compliance) maps directly to NIST AI RMF. Multiple domain-1 and domain-2 factors operationalize specific RMF subcategories at the control level.

Source Framework

ISO/IEC 42001 — AI Management Systems

AIRS factors 4.2, 5.1–5.5, and the maturity scale itself draw on ISO/IEC 42001's AI management-system controls and continuous-improvement structure.

Source Framework

EU AI Act

AIRS readiness factors crosswalk to the conformity assessment, transparency, and risk-management obligations imposed on high-risk AI systems under the Act.

Source Framework

NAIC Model Bulletin on AI

The NAIC's expectations for AI use within insurer operations are directly reflected in AIRS factors 2.1–2.4 and the governance architecture of Domain 4.

View the Full Regulatory Crosswalks

07

Conformance & Assessment

AIRS distinguishes between three levels of assessment, each appropriate to a different institutional purpose. The specification defines the conformance criteria that govern when an AIRS score may be cited externally and what artifacts must be retained.

Level A

Self-Assessment

The free public calculator. Suitable for internal benchmarking, gap analysis, and remediation planning. Self-assessed scores must not be cited externally as conformant AIRS results.

Level B

Independent Assessment

A score produced by a qualified third-party assessor following the specification's evidence requirements. Suitable for underwriting submission and counterparty review.

Level C

Conformant Citation

A score that may be cited externally as an AIRS result. Requires Level-B assessment plus retention of evidence artifacts for the specification's audit period.

Section 9 of the specification defines the full conformance criteria, including assessor qualifications, evidence retention requirements, and the disclosure language required when citing an AIRS score externally.

08

Suggested Citation

AIRS is published as an open standard. Organizations may freely reference, implement, and build upon the methodology for internal risk assessment, underwriting, regulatory compliance, and academic research, provided that attribution is given and the standard version is cited.

Suggested Citation Format

AIRS Standards Body. (2026). AI Insurance Readiness Score Open Standard Specification (AIRS v1.1, Document Version 1.1). Editorial stewardship and methodology administered by AI Security Intelligence LLC. https://www.aisecurityintelligence.com/airs-v1-specification

In-Text Citation

(AIRS Standards Body, 2026, AIRS v1.1)

Reproduction or redistribution of the specification document in its entirety requires prior written permission from AI Security Intelligence LLC.

Receiving Line

For citation, conformance, or institutional inquiries

The methodology is published in full and free to reference, cite, and implement. The Standards Body receives correspondence from carriers, reinsurers, regulators, brokers, and peer bodies directly.

standards@aisecurityintelligence.com