AI Governance at Orchra

Orchra can dispatch AI agents that read revenue signals and execute routine work on a deal. Because those agents act on sensitive revenue data, governance is built into the product rather than added on. This page describes how that control model works. Questions: security@orchra.io.

Human approval and control

Agents operate strictly within configured permissions. Sensitive or irreversible actions can require explicit human sign-off before they run, and a human can pause, override, or revoke an agent at any time.

Scoped agent permissions

What each agent is allowed to read and do is explicitly defined and role-based — never open-ended. Permissions are least-privilege by default and are configured per workflow.

Audit trail and rollback

Every agent action, configuration change, and forecast override is captured in a versioned, attributed audit trail, so it can be reviewed and reversed. This is the same trail that makes the forecast defensible: every figure drills to its formula and the underlying signal.

Data isolation

Customer data is logically separated. Agents operate only within the data scope of the customer they serve.

Model and training policy

We do not use your private revenue data to train shared or third-party foundation models. Model usage is governed and logged.

Input and prompt-injection controls

Because agents process untrusted content such as inbound emails and meeting notes, we apply controls to reduce prompt-injection and unsafe-instruction risk, including scoping what an agent can act on and separating instructions from ingested content.

Audit events

Governance-relevant events — agent runs, approvals, overrides, permission changes — are logged with attribution and timestamps for review and export.

Governance roadmap

We are formalizing this control model as part of our broader security and privacy program. Related public detail lives in the Trust Center and Security pages. We do not represent any certification as complete until evidence is available.