Choose approval requirements from consequence, ambiguity, authority, and recovery, rather than adding a review button to every automation. Human attention is limited and should protect decisions where it can materially improve the outcome.
Approval policy is a business design
An approval policy defines which actions need a qualified decision before execution. Relevant factors include irreversibility, financial or customer impact, evidence quality, identity assurance, and whether the workflow can detect and repair errors.
For a proposed assessment, inventory actual effects: reading information, drafting a message, updating a record, sending external communication, cancelling work, or transferring value. Score their material properties qualitatively and assign an accountable policy owner.
No universal action list fits every business. A routine address correction can be low impact in one context and consequential in another. The actual deployment’s requirements and authority model must govern the decision.
Compare fictional operational actions
An assistant drafts a callback note for staff review. The draft does not itself contact the customer or alter a commitment, so its approval boundary can differ from sending the final message. Saving validated intake may also have a different policy from cancelling a booked visit.
A disputed refund recommendation has uncertain evidence and a material effect. A qualified reviewer may need to decide, while the application still enforces resource permission and amount limits. Review does not replace those controls.
| Action property | Policy question |
|---|---|
| Reversible low-impact update | Can deterministic checks suffice? |
| External promise | Is authoritative completion evidence available? |
| Ambiguous customer intent | Is clarification needed before proposal? |
| Irreversible or material effect | Who has authority to decide? |
| Weak recovery support | Should the effect remain manual? |
Use fictional failure exercises to assess each proposed scope. Ask what happens if the action is wrong, who notices, and whether the customer can be made whole under the actual process.
Review the queue the policy creates
A policy can look cautious while creating more proposals than staff can meaningfully review. Estimate volume and reviewer coverage. Remove unnecessary review from deterministic low-consequence preparation when approved, while preserving strong boundaries for consequential effects.
Approval requirements should evolve through evidence and authorised policy changes. Do not quietly reduce oversight because a pilot feels smooth. Record the rationale, tested controls, and remaining limits before expanding autonomous execution.
This framework is proposed. The HVAC case study explicitly distinguishes intake from confirmed appointments and dispatch. Pair it with manual-process suitability.
Approval policy FAQ
Does more human review always mean less risk?
Not if reviewers lack evidence, authority, time, or a meaningful ability to reject.
Can approval be replaced by deterministic checks?
Sometimes, under the approved scope and consequence assessment. Verify the actual invariant rather than assuming equivalence.
Who should own the policy?
The accountable business owner, with engineering implementing and testing the authority and execution boundaries.