Guide: What Is The AI Agent Actually Allowed to Do?
A lightweight check for defining what one agent may recommend, prepare, approve or execute and where a human must take over. Pick one agent and one workflow, try not to design the enterprise autonomy framework today.
What Is This Agent Actually Allowed to Do?
A lightweight check for defining what one agent may recommend, prepare, approve or execute — and where a human must take over.
The agent’s capability should not be seen as authorisation.
What this Guide helps with
Use it when teams are moving from AI assistants that suggest work toward AI agents that can act in systems, send communications or make bounded decisions.
Keep it small
Pick one agent and one workflow. Do not design the enterprise autonomy framework today.
The Pattern behind this Guide
Agent Delegation Limits: this public Pattern quiz leads into this Guide.
The check
Break the workflow into actual verbs: read, draft, recommend, create, approve, send, change, pay, delete.
Recommend, Prepare, Approve or Execute for each action.
Value, risk, reversibility, customer impact, data sensitivity or another limit.
Who reviews, approves, overrides or receives escalation?
What happens when the agent is uncertain, blocked or outside limits?
Logs, approvals, action history and ownership that show what the agent did.
Use this small log
| Action | Delegation level | Boundary | Human control |
|---|---|---|---|
| Draft supplier response | Prepare | No external send | Buyer reviews |
| Approve low-value refund | Approve | Within agreed threshold | Escalate above threshold |
| Update CRM record | Execute | Defined fields only | Audit log + exception queue |
Where this commonly hides
Quick examples
| What you see | Question to ask |
|---|---|
| Agent can draft and send | Which verb is actually authorised? |
| Agent approves under £X | Who set the threshold and who owns exceptions? |
| Agent acts but logs are weak | Could a human reconstruct what happened? |
| Agent stops on ambiguity | Who receives the escalation and context? |
The Satire
The agent was authorised to 'help', unfortunately the 'help' still needs role and access definition.
Pattern quiz
Would You Let This Agent Recommend, Prepare, Approve or Execute?
Take the quick quiz that leads into this Guide.
More Patterns
Keep spotting what returns.
Explore more quiz-led workplace Patterns.
More Guides
Follow the work.
Use another lightweight Guide on one real workflow.
Want the deeper version when it lands?
The deeper Agent Delegation Limits Insight and future toolkit are coming later. Subscribe for the next Pattern drop and we will let you know when they land.
Lightweight self-check, not an organisational diagnosis or instruction to remove required controls or support.
