AI Safety & Guardrails

Boundaries that can't be crossed.

Every AI output is bounded by ASIC-compliant guardrails enforced at the API level. These rules can't be overridden by users, prompts, or edge cases.

Hard boundaries

What the AI will never do. What it always does.

These aren't preferences. They're hard constraints enforced in every API call, on every analysis, for every user.

Never

Provides personal financial advice

The AI will never recommend a specific product, tell a client what to buy, or suggest one insurer over another. All outputs are general information only.

Always

Appends the General Advice Warning

Every AI output automatically includes the General Advice Warning. Enforced at the API level — it cannot be removed, bypassed, or suppressed.

Never

Fabricates information

If a clause isn't in the document, the AI says "not addressed." If ambiguous, it flags for human review with a confidence score. Refuses rather than guesses.

Always

Cites the source clause

Every finding includes a precise citation — section number, page, clause heading. No citation, no claim.

Never

Answers from training data

Ask IA responses are grounded exclusively in your uploaded documents. No internet, no generic insurance knowledge, no other clients' policies.

Always

Flags items for human review

Medium and Low confidence findings are explicitly flagged. The AI identifies what it's uncertain about and routes those items to your review queue.

Never

Accesses other tenants' data

Every firm's documents are completely isolated at the database level. Tenant-scoped queries enforced on every API request.

Always

Logs every interaction

Every analysis, question, and AI response is recorded in an immutable audit trail with timestamps. 7-year retention for ASIC compliance.

General Advice Warning

Appended to every output.

Under Australian financial services law, any general advice must be accompanied by a General Advice Warning. Insurance Advice appends this automatically — enforced at the backend.

  • Appended by the Xano backend — not frontend code that could be bypassed
  • Cannot be removed, hidden, or modified by any user
  • Included in exports, client summaries, and PDF downloads
  • Wording aligned with ASIC regulatory guidance
AI Analysis Output
Property limit: $5,000,000. BI cover: 12 months indemnity period. Flood excluded under GE.12 with no carve-back. Machinery breakdown sub-limit $50,000...
General Advice Warning

This information is general in nature and has been prepared without taking into account any person's individual objectives, financial situation, or needs. Before acting on this information, you should consider its appropriateness and seek independent professional advice.

Enforced server-side · Cannot be removed
Review workflow
AI generates analysis
Structured output with citations
Draft
Broker reviews findings
Verifies citations, checks flagged items
Review
Broker approves output
Watermark removed, logged in audit trail
Approved
Shared with client
With GAW, broker attribution, citations
Sent
Human in the Loop

AI assists. Professionals decide.

Every output is marked as AI-assisted and requires human review before reaching a client. The licensed professional retains full decision-making authority.

  • All outputs watermarked "DRAFT" until broker approves
  • Medium/Low confidence items require explicit sign-off
  • No AI output shared with clients without broker review
  • Approval actions logged in audit trail
Regulatory Alignment

Built around the rules you operate under.

Every guardrail maps to a specific Australian regulatory requirement. Enforced in code, not theory.

AFSL Perimeter

Positioned as a professional analysis tool — not a consumer advice engine. All outputs stay within the general advice boundary.

Best Interests Duty

s961B evidence generated automatically — products investigated, gaps identified, rationale documented with every comparison.

ASIC RG 255

Algorithm version control and input/output logging meet ASIC's expectations for digital financial advice. Every version recorded.

ICA GICOP 2024

Claims notification timeframes, renewal disclosure, and complaints handling from the General Insurance Code embedded in analysis flags.

Privacy Act / APPs

APP 1, 5, and 8.2(b) compliance — privacy policy names Anthropic as subprocessor, cross-border disclosure documented.

Dec 2026 AI Amendments

Privacy Act automated decision-making transparency amendments. Designed from day one with explainability and human-in-the-loop review.

Explainability

Every output is explainable.

Every AI output includes a reasoning trace, confidence rationale, and full citation chain. Built to exceed ASIC REP 798 expectations.

  • Reasoning trace stored alongside every output
  • Per-field confidence rationale — why the AI is or isn't confident
  • Full citation chain from output → clause → page → PDF
  • Model + prompt version logged for reproducibility
Explainability record
Model
claude-sonnet-4-20250514
Prompt
pds_analysis_v2.4 · 7-step extraction
Thinking
Adaptive · effort: high · 142 clauses
Output
Structured JSON · 28 exclusions · 3 gaps
Review
2 items flagged Medium · 1 flagged Low
Contact

AI-powered insurance.
Coming soon.

We're here to answer your questions. We usually reply within one business day.

Prefer email? Write to hello@insuranceadvice.app