AI Transparency Notice

Purpose of this notice

Mantle uses artificial intelligence to help Australian financial advisers draft documents, extract information from files, and surface follow-ups. This notice explains, in plain language, what we use AI for, what data we send to AI providers, and the boundaries within which those systems operate.

This notice is voluntarily aligned to the ten guardrails in the Australian Government's Voluntary AI Safety Standard (2024) and complements our Privacy Policy and Terms of Use.

What Mantle uses AI for

Which AI providers we use

Large-language-model inference is provided by Anthropic, PBC via the Claude family of models. Depending on the model invoked, inference is routed either through:

• Anthropic's direct API (United States); or
• AWS Bedrock in the Sydney region (ap-southeast-2, Australia), where the model is available on Bedrock. Bedrock lets us keep inference on Australian infrastructure while still using Anthropic's models.

We use Sentry (United States) for application error monitoring — Sentry does not run generative AI on customer content.

See our Privacy Policy — Overseas disclosure for the full subprocessor table.

No training on your data

Under our commercial arrangements with Anthropic, personal information sent to Anthropic's API for inference is not used to train Anthropic's models. A short abuse-monitoring retention window applies at Anthropic's end, after which inference data is deleted.

Mantle itself does not train models on customer data.

Human in the loop

Every AI-generated output is a draft. No draft is sent to a client, filed as an advice record, or otherwise released to the outside world until a qualified adviser reviews and approves it. The Platform makes this boundary explicit — drafts are visibly marked as unapproved, and the "send" action is always in the adviser's hands.

Mantle's compliance detectors are diagnostic aids. A "pass" from a detector is not confirmation of legal or regulatory compliance, and does not shift responsibility from the adviser to Mantle.

Known limitations

Large language models can produce content that is fluent but wrong. In our context, the specific failure modes we have observed and designed against include:

• Hallucinated figures — dollar amounts, sums insured or premiums that do not appear in any source document. Mitigated by source-attribution detectors and required source citations for financial figures.
• Past-date framing — treating a date that has already passed as an upcoming deadline. Mitigated by an automatic date-sanity detector that regenerates on trip.
• Inferred intent or cover status — inferring that a client's cover has lapsed, or that they have accepted a recommendation, when the source material does not say so. Mitigated by dedicated detectors and prompt rules.
• Insurer naming in draft client emails — naming a proposed new insurer in the body of a client email rather than referring to "an alternative insurer". Mitigated at the prompt layer and by a pre-send detector.
• Stale references — using tax or super settings from a prior financial year. Mitigated by a single source of truth for regulatory figures, updated each financial year.

These are not the only failure modes. Advisers must review every draft as if the AI made no such checks.

Governance and accountability

Voluntary AI Safety Standard alignment:

• Accountability. The founder of Mantle is accountable for the safe use of AI in the Platform. Named contact: [email protected].
• Risk process. We maintain a rule catalogue of known compliance and quality risks, with detectors that run at generation-time and pre-send.
• Data governance. Client data is stored on Australian servers; AI inference passes through Anthropic in the United States under a no-training DPA (see Privacy Policy).
• Testing and monitoring. Structural and content detectors run on every generated output. Errors and prompt regressions are tracked via Sentry.
• Human oversight. Every AI-generated output is reviewed by a human adviser before release. Nothing is sent to a client automatically.
• Transparency to end users. Advisers see which model produced an output and which detectors ran. Clients who receive advice see the licensee's disclosure of AI use as part of the licensee's own Financial Services Guide.
• Contestability. Individuals can request access, correction, or complaint handling via [email protected].
• Supply-chain transparency. Subprocessor list published in the Privacy Policy.
• Records. We keep records of AI use consistent with the record-keeping obligations that customer firms owe under Chapter 7 of the Corporations Act 2001.
• Engagement. Feedback welcome at the address above.

APP 1.7 automated decision-making

From 10 December 2026, APP 1.7 will require our Privacy Policy to describe the kinds of personal information used by automated systems that make, or do something substantially and directly related to making, a decision that significantly affects an individual.

Mantle's automated systems do not make final advice decisions. Every recommendation, flag, or gate is reviewed by a human adviser before any client-facing action. This notice is our initial transparency statement; the Privacy Policy will be updated ahead of the December 2026 commencement to reflect any changes in scope.

Questions and feedback

Questions, feedback, or complaints about our use of AI: [email protected].