AI & decision-making
Where should AI assistance end and human approval begin?
A human approval button is not a safeguard if the reviewer cannot see what the system used or cannot reject it.
Published 25 September 2026
Place the gate where a mistake matters
Search suggestions and internal drafts can tolerate correction. Sending a customer promise, changing a financial record or granting access may require explicit approval. Draw the line by consequence, not by whether a model happened to be involved.
The reviewer should know what they are approving, which source the draft relies on and what the system could not establish.
Make approval operationally credible
Give the person time, training and authority to amend or refuse the action. Route uncertain cases separately; a queue of hundreds of trivial approvals encourages rubber-stamping.
Log the proposed action, the evidence, the final decision and any correction. If the human only confirms a preselected answer, the control is theatre.
A model drafts a reply about a customer's eligibility under an internal policy.
The reviewer needs the relevant policy version, the customer's supplied facts and a clear route to ask for missing information. Approving fluent text without those elements is not meaningful oversight.