Use this when

The AI can prepare an action, but a person should make the final decision.

This is useful when the preparation is repetitive and the action affects a customer, a financial record or another consequential part of the business.

The pattern

INPUT
  ↓
AI CREATES DRAFT
  ↓
PERSON REVIEWS
  ↓
APPROVED?
  ├── No → Edit, reject or escalate
  └── Yes → SYSTEM EXECUTES AND LOGS

Example

An AI prepares a quotation from a customer's request and the current price list. A salesperson checks the products, price and terms before the quotation is sent.

Why it works

The AI removes repetitive preparation while the person keeps control over the consequential action. The approval also creates a useful place to learn which drafts need correction.

Good fit

  • Customer communication that commits the company to something.
  • Changes to financial, employment or contractual records.
  • Early pilots where the error patterns are not yet understood.
  • Low-volume work where a domain expert can review every result.

Poor fit

  • Review is slower than doing the task directly.
  • The reviewer cannot see the source information needed to judge the draft.
  • People approve automatically because there are too many requests.
  • A fixed rule could produce the result more reliably.

Variations

Approve only above a threshold

Low-risk, well-tested actions execute automatically. Unusual or consequential cases require approval.

Review a sample

The system executes allowed actions, while a person reviews a meaningful sample and every flagged exception. Use this only after the workflow is well understood.

Draft without execution

The AI prepares the content, but the person completes the action in the normal system. This is often the safest first pilot.

What this does not solve

Approval does not repair unreliable source data, vague policies or missing ownership. It also does not replace logging, reversal and incident handling.

Good enough to use when

  • The reviewer can see the draft and its important sources.
  • Approval is explicit and recorded.
  • Rejection, editing and escalation paths are available.
  • The executed action matches exactly what was approved.