Human approval must preserve a real decision
Public-sector AI review needs evidence, time and authority to change the outcome. Design escalation and records before treating an approval as human judgement.
By 
Paul. AI and agentic engineering
A person approving an AI recommendation must be able to reach a different decision. For consequential public-sector work, that requires access to the underlying evidence, authority to intervene and a route for unresolved cases. Putting an approval button after a model's answer establishes none of those conditions by itself.
Jacob Kaplan-Moss's 1 October 2024 essay distinguishes assistive systems, which inform a human operator, from systems that take action themselves. He presents a personal ethical position developed from notes begun in 2021. This November 2026 retrospective takes that distinction as a design question. It is not a statement of universally applicable law or proof that human review makes an application ethical.
Locate the decision before designing the screen
Consider a proposed assistant helping an official examine an application for a public service. This is an illustrative workflow, not a deployed AL+ project. The assistant could locate relevant passages, identify missing documents and draft a summary. The official would still determine whether the evidence supports the decision under the applicable rules.
Write down exactly where the system's influence begins. Choosing which files appear first can affect what the official reads. Marking an application as incomplete can delay it even before anyone formally rejects it. A design that calls only the final signature a decision may overlook the earlier steps that shape the outcome.
Document the actions the assistant may take, the actions requiring authorization and the person or role allowed to authorize them. Separate permission to draft from permission to send a notice or alter a case record. These distinctions should be visible in the system's behavior and in the operating instructions.
That authority also needs an owner outside the software team. Developers can implement a permission model, but the institution must establish who is empowered to make the substantive decision. If nobody can explain who may reverse a recommendation, implementation should pause at that boundary.
Give reviewers evidence they can disagree with
A summary is a proposed interpretation. Preserve a route to the source passage and its document version so the reviewer can inspect what supports it. Distinguish an explicit statement in a file from an inference drawn by the model. A missing answer should remain missing rather than becoming a confident completion.
For the illustrative application, the review view could show the proposed conclusion alongside the documents used and any unresolved contradiction. The official should be able to inspect excluded material where their permissions allow it. Showing only supporting passages would make it difficult to discover that the recommendation omitted contrary evidence.
Access rules still apply. Giving a reviewer enough evidence does not mean granting access to every file held by the institution. If a case requires information the assigned official cannot lawfully or operationally access, the system needs an authorized escalation route. Expanding access merely to make the assistant work would change the problem rather than solve it.
There is an operating cost here. Meaningful inspection takes time, and a system that generates more recommendations than staff can examine can create a review backlog. Staffing and case allocation belong in the project estimate. A faster draft is useful only if the institution can complete the work that follows it.
Make disagreement an ordinary path
The reviewer should be able to correct the draft, request more information, reject the recommendation or defer the case to an authorized colleague. Those paths need defined destinations. A button labelled “escalate” is incomplete if it sends a message to an unattended queue while a deadline continues to run.
Record what evidence was available, what the assistant proposed and what the official decided. Keep the record proportionate to the case and the institution's data obligations. Recording every prompt forever would create its own disclosure and retention problems; the aim is to explain the decision, not collect an unlimited technical transcript.
Disagreement should not require an unofficial workaround. If an official must copy data into a private spreadsheet to avoid accepting a recommendation, the application has failed to support their authority. Review the written operating procedure alongside the interface: both should describe the same available choices.
An evaluation should include cases with conflicting documents and insufficient evidence. Ask whether the reviewer can stop the proposed action and get the case to the right person. Measuring how often staff click approve would leave the central design claim untested. No such evaluation was performed for this article.
Automation can have a legitimate place
The strongest counterargument is that requiring individual review for every action can waste scarce public resources. A deterministic acknowledgment or an administrative transfer under clearly specified conditions may be appropriate to automate. Human decisions can also be inconsistent, hurried or wrong. Adding a person is not a general cure for those problems.
The relevant boundary is the consequence and discretion of the action. A notice confirming receipt differs from a judgment about eligibility. Define which actions are routine, how exceptions are detected and who can stop or reverse them. Kaplan-Moss's ethical argument should prompt scrutiny of that boundary, rather than serve as a blanket ban on automation.
Nor does a well-designed review path resolve every question about fairness, accuracy or legal authority. A harmful policy can be executed with excellent records. The applicable institutional and legal assessment must examine the actual use, affected people and jurisdiction separately.
A useful brief for an AI agent workflow should therefore name one decision, its authorized human owner and the evidence needed to challenge the assistant. Before expanding the rollout, establish what happens when that person says the recommendation is wrong. That is the point where human approval becomes an operating responsibility.