AI safety

Our stance

AI in Rivantir exists to do the busywork — categorize, match, draft, flag. It never holds the authority. The design goal is simple: nothing applies itself behind your back.

Propose, then approve

Every change an agent wants to make is filed as a proposal in your Approval Inbox: what it wants to do, why, and with what confidence. You approve or dismiss — one by one, or by allowing high-confidence matches through automatically. That autonomy dial is set per workflow step, by you, and can be turned back at any time.

Hard limits

Some boundaries are structural, not settings:

• bank access is read-only, through regulated PSD2 providers — no agent can move money, ever;
• every agent run is logged and auditable, with usage quotas per plan;
• applied actions are reversible, and the audit trail shows who or what changed anything;
• agents operate strictly inside one organization’s data.

Data handling

AI features send model providers only the excerpts needed for the task, under data-processing agreements that exclude training on your data. Monetary values are passed to models in a controlled format to prevent misreading, and outputs are validated before they ever become proposals.

Tell us if something feels wrong

If an agent proposes something unsafe, misleading or simply strange, use the feedback control on the proposal — or write to safety@rivantir.com. Reports like these directly shape our guardrails.