Approval is different from review

A person can review something without having real authority over what happens next.

Imagine a workflow that sends a report to a manager but automatically publishes the report two seconds later. The manager technically "saw" the output, but there was no meaningful approval gate.

A real approval gate has three features:

  1. the workflow pauses;
  2. the approver has enough evidence to make a decision;
  3. the approver can allow, reject, revise, or escalate the work.

Without those features, "human in the loop" can become a label rather than genuine governance.

Put approval where the decision matters

Not every step needs a formal approval. That would make simple work unnecessarily slow.

Approval gates are most useful before actions that are difficult to reverse, externally visible, financially meaningful, security-sensitive, ethically significant, or otherwise consequential.

Examples might include the point before:

  • a public document is released;
  • a customer or employee is affected by a decision;
  • money is committed;
  • a system change is executed;
  • sensitive information is shared;
  • a final professional recommendation is adopted;
  • a ministry or family action is taken based on generated advice.

The specific gate depends on the work. The principle is to place human authority before the consequence—not after it.

Give the approver evidence, not just output

An approval gate is weak if the person sees only a polished final answer.

The approver may need:

  • the original goal;
  • the requirements or constraints;
  • the sources or evidence used;
  • unresolved questions;
  • important assumptions;
  • verification results;
  • known risks;
  • changes made during review.

The goal is not to overwhelm the approver. It is to give the person enough context to exercise judgment rather than simply click "approve."

Define the possible decisions

A good gate should not force a false choice between yes and no.

Useful outcomes may include:

  • Approve — the work meets the criteria and may proceed.
  • Approve with bounded correction — a small identified change may be made without restarting the entire workflow.
  • Reject — the work does not meet the requirements.
  • Return for revision — the work is salvageable but needs another pass.
  • Escalate — the decision belongs to someone with different authority or expertise.
  • Pause — required evidence or context is missing.

These choices make the approval step part of the work system instead of ceremonial sign-off.

Decide who has authority before the workflow runs

An approval gate fails if everyone assumes someone else owns it.

For each gate, define:

  • who may approve;
  • what qualifications or role that person needs;
  • what criteria the person applies;
  • what evidence must be present;
  • what happens after each possible decision;
  • what is recorded.

This is especially important when multiple people or AI roles contribute to the same deliverable.

Keep an approval record proportional to the stakes

For a low-stakes personal workflow, an approval record may simply mean you deliberately review the final result before using it.

For professional or consequential work, it may be worth recording:

  • who approved;
  • when;
  • which version was approved;
  • what evidence was reviewed;
  • any conditions or unresolved risks.

The purpose is not bureaucracy for its own sake. It is to preserve accountability and prevent confusion about which artifact was actually authorized.

AI can support the gate without owning it

AI can help prepare an approval package. It can summarize differences between versions, check a rubric, list unresolved issues, or identify missing fields.

That can make human review more efficient.

But if the decision properly belongs to a person, the AI should not quietly convert "I prepared the evidence" into "I approved the action."

The separation of roles matters.

A six-question approval-gate design test

Before relying on a human approval gate, ask:

  1. Where does the workflow actually pause?
  2. Who has legitimate authority to approve?
  3. What criteria are they applying?
  4. What evidence will they see?
  5. Can they reject, revise, or escalate—not just approve?
  6. What record shows what was actually authorized?

If those answers are unclear, the gate is not yet well designed.

Approval is where accountability becomes visible

AI orchestration often involves many useful steps: planning, research, drafting, review, verification, and documentation.

The approval gate connects those steps to a responsible human decision.

It says: the system may prepare the work, but the person with the appropriate authority decides whether it is ready to cross the boundary into real-world action.

That is not a weakness in the workflow. It is one of the mechanisms that makes the workflow governable.