Autonomous back office agents with audit (L5)
Routine internal work runs without a human in the loop: vendor renewal reminders, access reviews, expense policy checks, document filing. Every action is logged, bounded, and reversible, and exceptions route to a named human.
The steps
- 01
Choose reversible work only
Tool: Manual
Automate only tasks where a mistake costs minutes, not money or compliance standing. Payments, contract signature, and access grants stay human. Owner: ops lead plus security. DoD: allowed and forbidden task lists published.
- 02
Bound the tools and the budget
Tool: Agentforce
Explicit tool allowlist, rate limits, and a spend cap per agent. An unbounded agent is an incident waiting for a trigger. Pitfall: generous permissions granted during a pilot and never tightened. DoD: limits configured and tested by attempting a blocked action.
- 03
Log every action to one place
Tool: Slack
Single audit channel or table with actor, trigger, action, and outcome. If you cannot answer what did the agent do last Tuesday in under a minute, you are not ready for L5. Owner: ops lead. DoD: audit query demonstrated live.
- 04
Run a monthly kill review
Tool: Manual
Review every agent monthly against value delivered and incidents. Default action is retire unless someone defends it. Owner: ops lead. DoD: monthly review with at least one retire or renew decision recorded.
Tools in this playbook
- Manual
- Agentforce
- Slack
Next playbooks
Unfamiliar terms are defined in the AI and Revenue Dictionary. Related frameworks live in the framework library.