Teams ask for a policy template because the blank page is the hard part. Here is a starting set for the three AI employees most workspaces hire first, expressed as the four controls Kavanah exposes: capability scopes, an autonomy level, action policies by category, and guardrails on specific tools, plus the Negation every agent reads from the workspace charter. Treat it as a default to edit, not a standard to adopt.
1. The Planner
Job: triage new requests, plan sprints, assign work, chase blockers. Scopes: tasks, projects, people and availability; no mail, no finance, no deals. Autonomy: safe mode, which already strips sends, payments and deletes. Policies: none needed beyond safe mode. Guardrails: block archive_task and delete_task so a cleanup can never remove history. Schedule: a Monday-morning planning run and a daily triage. Reports to: the delivery lead. This employee can run at level four within a month because every action it takes is cheap and reversible.
2. The Client Desk
Job: draft client updates, answer routine client questions from the project record, log change requests. Scopes: tasks, clients, the shared inbox; no finance, no deals. Autonomy: full within policy, because it needs to send. Policies: external send is approve, so every outbound email parks in the approval queue with its draft and reasoning; finance is block. Guardrails: block any tool that changes a deadline, so the agent can report a date but never move one. Negation line in the charter: we never promise a delivery date in writing without a human approving it.
3. The Executive Agent
Job: the weekly state-of-the-business report, the revenue forecast, the escalation when a project is slipping. Scopes: read across everything, write to reports and tasks only. Autonomy: full within policy. Policies: external send is block for this one; it reports to a person who decides what leaves the building. Finance is block. Guardrails: block update_deal_stage and archive_deal so the forecast reflects reality rather than the agent's tidying. Reports to: the CEO, who gets the report and the escalations.
4. The Lines Every Agent Shares
A daily token ceiling, so a runaway run costs a known amount and stops. A credit budget per schedule. A standing rule that an agent never creates a task without a due date and an owner, which keeps the board honest. And a Negation in the workspace charter, written in plain language, that the agent must honour before it creates work. Keep the Negation short; five lines the whole company would agree with beat fifty nobody reads.
5. How to Know the Policy Is Working
Read the action ledger weekly and look for two things. Actions you had to undo, which mean a scope is too wide or a Negation is missing. And approvals you rubber-stamp every time without editing, which mean a category is ready to move from approve to allow. A policy that never changes is not being read.
Vendors increasingly describe their agents as lawful by design, and that is the right phrase. The design is the policy above. The law is whatever your team writes into it.
