Policies and guardrails

Path: Policies.

CEO common guardrails apply to all of your AI employees and to workflow Brain nodes. Use them for tone, forbidden actions, approval rules, and “never invent customers / prices”.

After you change policies, Resync org docs if the product asks you to, and start a new chat so employees pick up the latest text.

Action control (same page): three states per action family — Autonomous, Approval required, Prohibited. External messages default to approval; deletes default to prohibited. Approval uses a short-lived, owner-scoped grant; an agent cannot self-approve with request flags. A grant may be restricted to one tool, recipient, campaign, amount, and use. Prohibited always wins. Enforcement applies to read and write tool invokes for your company.

Use Scoped overrides & recurring grants for one goal, workflow, employee, or tool. Resolution is goal → workflow → employee → tool → company. A recurring email or publishing grant can be Autonomous while restricted to permitted email IDs or website domains, an expiry, and a maximum number of uses. Non-matching constraints fail closed; expired or exhausted overrides return to the next applicable policy.

The searchable target picker supports multi-select, so one save can apply the same bounded rule to several targets. A scheduled-goal override remains effective across that goal's future scheduled fires.

Policies are not a substitute for tool grants, Maker/Checker, or IP allowlists. Stack them: policy says what should happen; Action control, tools and Checker enforce what can happen.

See Example stress test run for a scored outcome (verified CRM, spend cap, approval-required send).