An approval button shows that someone clicked. It does not, on its own, show what they could inspect, whether they could disagree, or whether the action had already become difficult to stop.
The question is not simply whether a person appears in the workflow. It is whether that person has a meaningful opportunity to exercise judgement within a clear scope of authority.
My article Why I Built an Anthropic TV-AI set out a human-accountability principle: advisory outputs must remain distinguishable from decisions made by authorised people. This note develops the review conditions behind that principle, without making claims about a particular product’s performance.
Define the decision before presenting the recommendation
“Approve” is too ambiguous when several activities sit behind the same button. Is the reviewer checking a calculation, accepting a report for discussion, approving an operating change, or authorising an external commitment?
Give the review a specific object and version. State the decision sought, the proposed next action, the limits of the reviewer’s role and what remains outside it. A technical reviewer’s conclusion may inform a separate approver without itself authorising the action.
The NIST AI Risk Management Framework Playbook, GOVERN 3.2, addresses the definition and differentiation of human roles and oversight responsibilities. Its MAP 3.5 addresses human-oversight processes. Those are useful reference points; the practical questions here are not a certification checklist.
Let the reviewer inspect the basis—not only the summary
A useful review view separates the recommendation from its inputs. It identifies which statements are observations, calculations, assumptions or unresolved claims. Sources need to be available at the level needed for the decision, with their versions and limitations intact.
That does not mean showing everyone every confidential file. Where access is restricted, establish an appropriate review route and explain the limitation. A restricted source cannot simply be treated as verified because a generated summary sounds confident.
The reviewer also needs time, relevant expertise and a route to seek help. A formally available reject button does little if the person cannot inspect the material or is expected to clear an unmanageable queue without examination.
A fictional recommendation that should not pass unchanged
A meaningful review record would preserve the first briefing, the missing-period concern and the request for further evidence. Regenerating the briefing should create a new version, not erase the basis of the reviewer’s challenge.
When the missing logs arrive, the reviewer may reach the same recommendation, amend it or advise against it. What matters is that the outcome responds to the evidence and remains within the reviewer’s role—not that disagreement necessarily produces rejection.
Make disagreement consequential
The available responses should fit the decision. A person may need to amend a proposed action, request specialist assessment, defer a conclusion, reject it or refer it to someone with the right authority.
Test the actual effect of those responses. Does deferral prevent the proposed next step? Does a request for evidence return to a named owner? Can an unresolved objection remain visible after another authorised person decides to proceed?
Where proceeding is permitted despite an open concern, the decision record should explain the rationale, remaining uncertainty, limits, safeguards, follow-up owner and reconsideration trigger. The concern should not be relabelled “resolved” merely because a decision has been made.
It is equally important not to imply that a recorded rationale creates permission to bypass a required condition. Matters outside the reviewer’s authority need the appropriate escalation or decision route.
Keep the resulting instruction narrower than the discussion
The same discipline applies after approval. An authorised instruction should identify the action, version, conditions and scope actually approved. A broad positive assessment should not be stretched to cover additional spending, new recipients or a different activity.
For consequential actions, test that the workflow respects a refusal or deferral before execution. A timeout or unanswered message should not be presented as human approval. An authorised contingency arrangement, where one exists, needs its own recorded basis.
Review whether human oversight is working
A practical exercise is to introduce a clearly labelled test case with an incomplete source, a material contradiction or an out-of-scope action. Observe whether the reviewer can find it, challenge it and prevent the unapproved next step. Use a test environment; do not experiment on a live commitment.
Look at the reasons for amendments and deferrals, the response to unresolved concerns and whether reviewers can obtain specialist help. An approval rate or a count of human clicks cannot answer those questions by itself.
Nor should this become a target for maximising rejection. The aim is a review that can reach and explain the appropriate conclusion, including justified approval, without pressure to manufacture certainty.
Where this sits in the framework
Fracture identifies a gap in the basis or authority for the proposed action. Signal makes the concern visible. Spine gives an authorised person a real route to respond. Assurance preserves what was reviewed, what was decided and what stayed open.
The useful test is: Could this reviewer change what happens next—and would the record show why?
About this version
This article was prepared on 24 September 2026. It develops the author’s linked public writing with new explanatory examples. The figures, identifiers and scenarios are fictional; they are not programme evidence. References were checked on the preparation date.
Fracture → Signal → Spine → Assurance