AI & decision-making · JNAGA perspective
Where should AI assistance end and human approval begin?
A human in the loop is useful only when that person has enough context and authority to intervene.
Published 25 September 2026
Place the boundary at the consequential action
Summarising a document for internal consideration differs from changing a customer record, sending a commitment or deciding someone's access. The greater the consequence or difficulty of reversal, the stronger the case for explicit human authority. The task, not the presence of an AI model, should determine the control.
A review step should specify what the person is approving. Are they checking facts, tone, policy fit, permission to disclose information or the decision itself? Different checks need different evidence.
Give reviewers room to disagree
Show the source material, uncertainty and exceptions in a form the reviewer can understand. Make it easy to amend, reject or escalate. If speed targets discourage correction, the nominal approval may not provide the protection the business expects.
Record the final authorised action and the reason when it matters, without keeping unnecessary sensitive material. That makes it possible to learn from errors and explain what happened.
Revisit the boundary with evidence
Some tasks may become safe to automate more fully after testing; others will continue to need judgement. Review actual error patterns and operating conditions before changing the boundary. Do not equate a low number of reported problems with proof that no problems exist.
Imagine an assistant drafts a response to a customer's complaint. The text sounds confident, but the assistant has not seen the latest service history.
The reviewer needs the underlying record and an easy way to correct the draft. If the interface offers only 'approve' with no context, the human step is ceremonial rather than protective.