# AI processing and human review How Aiviance selects an AI route, limits what models may receive and keeps people responsible for consequential actions. AI in Aiviance prepares work; it does not become the accountable decision-maker. The processing path is designed around region, data sensitivity, permission and confirmation. ## A policy selects the processing route Each AI request declares the capability it needs, the customer's governance region and the classification of the data involved. The provider router considers only providers allowed for that combination. For Australian customers, supported model workloads use AWS Bedrock in Australia. The highest-sensitivity class is excluded from generative AI and large-language-model providers. The current governed document-understanding register covers P1 and P2; the AI Use Disclosure permits an approved OCR processor to handle P0 solely for extraction where that flow is enabled, never as generative-model context. Where no permitted provider supports a capability, the platform should stop or degrade the feature rather than silently use an unapproved route. ## Sensitive values are limited before dispatch Aiviance classifies information before model dispatch and applies privacy controls to model inputs and application logs. Clear-text credentials and the highest-sensitivity identifiers are not general model context. Masking and redaction reduce exposure; they do not make financial information anonymous. Names, transaction context or other details may still be personal information when they are necessary to perform the requested task. ## AI output is a proposal, not an approval Read-only tools may retrieve information only within the active entity scope. A material write must pass the same permission and policy checks as the corresponding product action. Sensitive operations can require a fresh step-up authentication, and higher-risk business writes are routed through proposal and approval controls. A user can review, reject or confirm a proposed action. Confirmation is not inferred from opening a conversation or asking a question. ## Processing leaves an operational record The platform records the provider, processing region, data classification and policy outcome for model calls. It also records tool activity used to prepare an answer. The tool-level provenance record is currently best-effort and is not yet presented as a complete user-facing evidence view. Aiviance therefore distinguishes an inspectable proposal from a claim that a model can explain its own reasoning. ## Model training and provider terms Aiviance does not use customer prompts or completions to train its own models. Provider-side handling depends on the agreement and route recorded for that provider. Current provider and location information belongs in the public sub-processor register and AI Use Disclosure, which should be read for the applicable arrangement. ## Related policies - [AI Use Disclosure](/legal/ai-use) - [Privacy Policy](/privacy) - [Sub-processor register](/legal/sub-processors)