16 minute read
Agentic AI does not mean removing humans from your operations. It means deciding, deliberately, which decisions belong to agents and which belong to people. Getting that boundary right is the most important design choice in any agentic deployment.
Most organizations start in one of two failure modes. The first is over-caution: requiring human approval for every agent action, which eliminates most of the efficiency gains and leaves staff doing the same volume of review they were doing before. The second is over-delegation: letting agents act freely across a wide scope, which produces fast outcomes until it produces a very bad one.
The goal is a principled middle path.
A Framework for Drawing the Line
We use three criteria when deciding whether a given action should be agent-autonomous or human-approved:
Reversibility. Can the action be undone if it turns out to be wrong? Sending a read receipt is reversible in the sense that it has low consequence. Issuing a refund or deleting a record is harder to walk back. Higher-consequence actions warrant higher human oversight, regardless of how confident the agent is.
Frequency and volume. Actions that happen hundreds of times per day — routine routing decisions, status updates, acknowledgment messages — are poor candidates for human review. The review bottleneck would negate the automation. Actions that happen rarely and have high stakes are better candidates for human checkpoints.
Variance in the input. Agents handle well-structured, predictable inputs reliably. As inputs become more unusual, ambiguous, or emotionally complex, human judgment adds proportionally more value. Build your escalation triggers around input variance, not just action type.
What This Looks Like Deployed
In practice, a well-designed agentic system has three zones: fully autonomous (agent acts without notification), notify-and-proceed (agent acts and informs a human), and hold-for-approval (agent queues the action and waits). The distribution across those zones should be tuned to your actual risk tolerance — not set once and forgotten.
The frequency criterion above is also an anti-theater rule: if a manager must approve 300 nearly identical actions each morning, you have not built governance — you have built a human bottleneck. Move that work into an autonomy zone (or raise the bar for what requires a gate), and reserve hold-for-approval for actions that actually need judgment.
What a Meaningful Approval Shows
An Approve button without decision context is theater. Every hold-for-approval action should surface an approver packet:
- Proposed action
- Supporting evidence
- Agent reasoning summary
- Confidence or uncertainty
- Policy triggered
- Potential downside
- Reversibility
- Recommended decision
- Similar prior outcomes
Approvals are recorded with who approved, when, and which packet they reviewed.
How You Know the Gate Isn’t Theater
Track oversight health — not just that gates exist:
- Approval volume
- Average review time
- Override rate
- Rubber-stamp rate (approvals with negligible review time or no packet engagement)
- Reviewer disagreement
- Actions automatically approved after repeated success (progressive autonomy with evidence)
Sampling and autonomy should step up only when these metrics show real review — not because a calendar says “30 days.” The full delivery standard is in the Operator’s AI Governance Checklist.
The right human-in-the-loop design is specific to your workflows and risk profile. We design that calibration on every engagement.