Many organisations introduce AI agents and then place a human approval button after every step. It feels responsible, but it often creates a new bottleneck. The reviewer becomes fatigued, approvals become automatic, and the system gains the appearance of oversight without the substance.
The best AI leader does not approve everything. The best AI leader designs escalation rules so routine, reversible work can move while consequential decisions reach the right human with the right evidence.
Approval is not the same as accountability
A person clicking “approve” may not know what data the agent used, what alternatives it rejected or what permission it exercised. Accountability requires more than a final click. It requires clear ownership, traceable reasoning, bounded authority and a route to challenge the decision.
NIST’s AI Risk Management Framework emphasises defined roles and responsibilities for human–AI configurations. In practical terms, every automated workflow should answer: who owns the outcome, who monitors the system, who can stop it and who handles exceptions?
A four-zone escalation map
Green: act automatically
Use this zone for low-impact, reversible tasks within narrow permissions: classifying internal notes, creating a draft, formatting data or proposing meeting times. Actions should remain logged and reviewable.
Amber: act and notify
The agent may proceed, but a human receives a summary because the action affects customers, budget or public output. Examples include scheduling approved content, adjusting a campaign inside a small limit or answering a routine request from an authorised knowledge base.
Red: human approval before action
Require explicit review for financial commitments, publication of sensitive claims, legal or medical implications, use of personal data, unusual account activity, contractual changes and actions that are difficult to reverse.
Black: prohibited
Some actions should remain outside the agent’s authority: disabling audit logs, bypassing security, creating undisclosed impersonations, spending outside the task budget or deleting critical records without a recoverable process.
Escalate on signals, not only task names
The same task can change risk depending on context. A £20 advertising test is not the same as a £20,000 campaign. A draft email is not the same as a message sent to every customer. Escalation should therefore respond to signals:
- value exceeds a financial or volume threshold;
- confidence falls below an agreed level;
- the request conflicts with policy or prior instructions;
- new personal, confidential or regulated data appears;
- the action crosses from internal draft to external commitment;
- the agent attempts to expand its own permissions.
Design the escalation packet
Do not send a reviewer a vague alert saying “approval needed.” The packet should contain the requested action, purpose, evidence, uncertainty, cost, affected people, permissions used, alternatives and the consequence of delay. The human should be able to approve, modify, reject or route the case elsewhere.
Protect the human from approval fatigue
If every action is marked urgent, the system trains people to click. Track how many escalations are accepted unchanged, how many reveal genuine risk and how long reviews take. Repeated safe cases may move downward; repeated surprises should move upward or be prohibited.
Leadership becomes architecture
In an AI-enabled organisation, leadership is partly the design of decision rights. The question is not “Can the agent do this?” It is “Under which conditions may it do this, how will we know, and who remains answerable?”
Autonomy without boundaries becomes permission creep. Oversight without prioritisation becomes theatre. Escalation rules create the space between them.
Further reading: NIST AI Risk Management Framework.
Discover more from Marychuks.com AI, Psychology, Business & CreativeVerse
Subscribe to get the latest posts sent to your email.
