Responsible AI

What the model does, and what it doesn't

Aicura runs on Anthropic's Claude models through Amazon Bedrock in Australia, for every customer wherever the customer is based. Model calls use Bedrock's Australian cross-region inference, so each one runs in Sydney (ap-southeast-2) or Melbourne (ap-southeast-4), whichever has capacity. Embeddings run in Sydney only. Claude is used for document analysis (reviewing privacy and governance documents), drafting disclosures and policy text from structured register data, pre-populating risk assessment fields and reading the helper documents customers upload.

The platform does not use Claude, or any model, to make compliance determinations, give legal advice or produce final outputs without human review. The model is one component of a layered architecture, in which Aicura's interpretive content (assessment briefings, framework mappings, regulatory criteria) is curated by Aicura, persisted in the platform and applied via prompts. The model executes interpretation against that content and does not constitute the interpretation itself.

The AI models behind Aicura

Aicura reaches its models through Amazon Bedrock in Australia. Requests go to AWS, not to the model provider, and run under Aicura's agreement with AWS.

• Where the models run. Model calls use Bedrock's Australian cross-region inference, so each one runs in Sydney or Melbourne. Embeddings, the numeric representations Aicura uses to search and match text, are generated in Sydney.
• No training on your content. Your inputs and outputs are not used to train or improve any foundation model.
• No access for model providers. The model providers, including Anthropic, cannot access your prompts or completions.
• How long AI logs are kept. Aicura keeps a log of each AI call (prompt and output) for up to 90 days for debugging and quality assurance, then deletes it automatically.
• Changing models. Aicura calls Bedrock through its Converse API, which uses one request format across model families. Moving to a different Bedrock-hosted model is a configuration change, not a rebuild of the product.

Human oversight

Every AI-generated output is a draft requiring human review before reliance, sharing or publication. This is made explicit in the product UI at the point of generation rather than buried in terms, and AI-generated content is labelled. Aicura does not publish anything on a customer's behalf.

Data handling

• AI processing runs in Amazon Bedrock in Sydney and Melbourne, so customer data does not leave Australia for AI processing.
• Requests go through Amazon Bedrock, not directly to Anthropic. Under Aicura's agreement with AWS, customer inputs and outputs are not used to train or improve foundation models, and model providers (including Anthropic) cannot access prompts or completions.
• AI processing logs (prompts and output) are retained up to 90 days for debugging and quality assurance, then automatically deleted, with the canonical record sitting in the application database. AWS-side full-text invocation logging is intentionally not enabled, to avoid duplicate at-rest copies of customer content.
• Multi-tenant isolation and customer-managed-key encryption apply to AI-related data as to everything else.

Guardrails, and a plain note about them

A model guardrail capability exists but is deliberately not applied as a generic filter. Generic content and topic filters false-positive on regulatory text, since anti-discrimination law reads as "hate" and "legal/financial advice" topic blocks hit exactly Aicura's domain, and PII anonymisation would strip the contact details that generated governance and privacy policies are supposed to contain. Critically, there is no free-form prompt surface, because every input is structured data Aicura constructs or a document the customer chose to upload, so a generic guardrail has nothing to protect that it wouldn't corrupt. This holds even for customers running Aicura's guardrail enforcement, because the AI gateway and its scanners run in the customer's own environment and their AI traffic never reaches Aicura. The relevant control for untrusted text inside uploaded documents is handled at the application layer (treating document content as data, not instructions), not by a blanket model filter.

Frontier-model dependency & concentration risk

For regulated customers, such as APRA-regulated entities in Australia or firms under operational resilience rules elsewhere, adopting Aicura does not materially increase concentration on a single AI vendor.

• The platform abstracts the model. The interpretive content, register, assessments, generated documents and workflow are Aicura's, not the model provider's. On your AI risk register, the appropriate entry is "Aicura AI governance platform (interpretive guidance, runs on Amazon Bedrock in Australia)" rather than the underlying model.
• Substitution is a configuration change, not a rebuild. The Bedrock Converse API provides a single wire format across model families, so a model swap is a configuration change plus a smoke test. Prompts already run against non-Claude models in development, confirming portability.
• How we would respond to a disclosed model concern. Assess impact on Aicura's specific use, meaning document analysis and structured generation rather than code execution, autonomous action or tool use, notify affected customers and switch to an alternative Bedrock-hosted model if the concern materially affects our use.

Where this is disclosed

Our use of large language models, the no-training commitment and the 90-day log retention are disclosed in our Privacy Policy. The probabilistic, draft nature of AI output and the requirement for human review are in our Terms of Service. This page is the standing public statement.

Formal AI management system certification

Aicura is evaluating ISO/IEC 42001, the AI management system standard, as the formal framework for its own AI governance. Aicura is not certified to it, and we don't claim it.