Our AI principles

Anything that becomes an official record, sends an external message or changes a policy waits in the Agent Inbox for a person with the right role. Only low-risk actions such as reminders can be automated, and only when the institution chooses.

Agents must cite the source of every claim. Output without sources is rejected before it reaches a person. Minutes link every sentence to the moment in the transcript it came from, and write [CHECK] instead of guessing.

Each agent acts on behalf of a named person and can never see or do more than that person. Confidential items are filtered out before an agent searches.

Every AI output carries an “AI draft” label until a person adopts it, and adoption is recorded with their name and the time.

In assessment, declarations and the writing process come first. AI-detection scores are only ever a weak signal and can never, on their own, open an integrity case.

Requirements come only from cited source documents. Placeholders are clearly marked DEMO until the official text is loaded.

Every agent run records its inputs, tools, sources, model, prompt version, cost and approver. Agents are evaluated against gold-standard examples before release.

We don’t train models on customer data, we use zero-retention AI settings where available, and prompts and outputs stay in your institution’s tenant.