Insurance companies process thousands of claims daily — property damage, medical bills, liability disputes. AI can accelerate adjudication, flag fraud, and reduce cycle times. The challenge: policyholder data includes medical records, financial details, and personally identifiable information that must never leak.

AKIOS solves this with the Security Cage: an ephemeral, sandboxed runtime where AI processes claims under strict, code-defined policies.

The Problem

Claims processing is a bottleneck. Manual adjudication is slow, inconsistent, and expensive. AI can evaluate a claim in seconds — but connecting a model to policyholder data creates a new risk vector. Medical records, SSNs, policy numbers, and financial details are all at stake. A data breach doesn't just mean regulatory fines — it means class-action lawsuits and lost customer trust.

AKIOS lets you deploy AI on claims data without exposing any policyholder identifiers.

The Regulatory Landscape

US insurance companies face a unique regulatory stack:

RegulationScopeHow AKIOS Enforces It
State Insurance Regulations Each state has its own data protection and claims handling requirements (NAIC Model Laws as baseline) Policy templates per state ensure the cage enforces jurisdiction-specific rules.
HIPAA Health insurance claims involve PHI — same protections as core healthcare systems Medical record identifiers redacted before AI processing. Zero PHI exposure.
NAIC Model Audit Rule Insurers must maintain auditable records of all automated decision-making affecting policyholders Merkle-chained audit trail — every classification cryptographically signed and immutable.
FCRA When AI uses credit-related data for underwriting, adverse action requirements apply Complete decision provenance exported per claim. Adverse action explanations included in output.
CCPA / State Privacy Consumer data rights — access, deletion, opt-out of automated decision-making All PII redacted before AI processing. Audit trail proves no personal data retained by model.

AKIOS enforces these at the runtime level — the AI agent never operates outside the compliance boundary.

The Concept: Policy as Code

AKIOS introduces the concept of a "Security Cage" — an ephemeral, sandboxed runtime environment where data is processed under strict, code-defined policies. For insurance, the key feature is per-state policy templates: compliance rules can be configured per jurisdiction, ensuring the AI respects state-specific claims handling requirements.

The Workflow: Automated Claims Adjudication

StepWhat HappensSecurity Control
1. Ingestion Batch of new claims loaded into the Security Cage SSNs, policy numbers, and medical IDs redacted before AI sees them.
2. Redaction Policyholder identifiers stripped and replaced with tokens 50+ PII patterns including medical codes, financial accounts, addresses replaced.
3. AI Analysis LLM evaluates claims against policy terms — damage assessment, coverage verification, medical necessity Budget capped ($0.30/claim), network isolated, no persistent storage.
4. Classification Claims classified as approve, deny, or escalate with confidence scores and rule citations AI cannot authorize payments. All approvals require human confirmation.
5. Audit Every inference, data access, and output cryptographically signed State regulators can verify the complete decision path for any claim.

Architecture

graph LR
    CMS["Claims Mgmt\nSystem"] -->|"claim files\n(encrypted)"| FS["filesystem agent\nbatch ingest"]

    subgraph CAGE["AKIOS Security Cage"]
        FS --> PII["Redaction Engine\n«SSN» «POLICY» «MEDICAL» «CC»"]
        PII --> LLM["llm agent\nclaims triage"]
        LLM --> TE["tool_executor\nadjudication rules"]
        TE --> VALID["Output Validation\nno payment auth"]
        VALID --> MERKLE["Merkle Chain\nSHA-256 signed"]
        MERKLE --> COST["Cost Kill-Switch\n$0.30 / claim"]
    end

    COST -->|"classifications\n(redacted)"| Queue["Decision Queue\nAdjuster Dashboard"]
    Queue --> Adjuster["Claims\nAdjuster"]
    MERKLE -->|"audit export\n(immutable)"| Compliance["NAIC Audit Trail"]
    Compliance --> DOI["State Dept of\nInsurance"]

Policy Configuration

The entire compliance posture is defined in a single YAML file:

# insurance-claims-policy.yml
security:
  sandbox: strict
  network: isolated
  allowed_endpoints: []  # zero network access
  pii_redaction:
    enabled: true
    patterns: [ssn, policy_number, medical_record, credit_card, address, dob]
    mode: aggressive
  budget:
    max_cost_per_run: 0.30
    currency: USD
  audit:
    merkle_chain: true
    export_format: jsonl
    retention_days: 2555  # 7 years — NAIC record retention
  constraints:
    payment_authorization: disabled  # AI cannot approve payments
    human_review: required_for_denials

What the Claims Adjuster Sees

At the end of the workflow, the claims adjuster receives a structured report:

FieldValue
Claim IDCLM-2026-0210-****5847
Classification⚠️ Escalate — Coverage ambiguity detected
AnalysisWater damage claim cites "sudden and accidental" but description indicates gradual seepage over 3 months
Policy ReferenceSection 4.2.1 — Exclusion for gradual deterioration
Fraud IndicatorsNone detected (0.12 risk score)
Confidence89%
Audit Hashd8f2a1...c47b
Policyholder Data Exposed❌ None — all PII redacted before analysis

No SSNs. No policy numbers. No medical details. Just adjudication intelligence with a cryptographic proof chain.

Why It Matters

  • Zero Data Exposure: SSNs, medical records, and policy details are redacted before AI processing. Even if the model is compromised, there's nothing to leak.
  • State-Level Compliance: Policy templates can be configured per state, ensuring the AI respects jurisdiction-specific claims handling rules.
  • Fraud Detection: The sandbox can cross-reference claims patterns without exposing individual policyholder data across claims.
  • Auditable Decisions: Every classification includes a cryptographic proof chain. State insurance commissioners can trace exactly how a decision was made.
  • No Unauthorized Payments: Hard-coded constraints prevent the AI from authorizing any payments. All financial actions require human confirmation.

Try It Yourself

pip install akios
akios init my-project
akios run templates/batch_processing.yml

Secure your AI. Build with AKIOS.