Assisting or Deciding? A Test for When an AI Agent Has Replaced an Official’s Judgment
METHODS · WORKING PAPER — Governance. The second paper of the Agentic Government programme. Empirical and analytical; anchored to dated, citable law as of July 2026. Builds directly on Who Acted, Under Whose Authority?
Public administration has a favourite reassurance: the human still decides. An official signs, so the decision is human, so the citizen’s rights are intact. But between an official who used a tool and an official who ratified a determination they could not reconstruct, there is a line — and almost nothing in current practice locates it. This paper offers a test that does: a way to establish, for a specific decision, whether the human was assisting-and-deciding or signing-and-ratifying. The distinction is not academic. It is the difference between a decision the law can treat as human and one that is, in substance, solely automated while wearing a human signature.
The stakes are legal, not rhetorical. The right not to be subject to a solely automated decision, the duty to give reasons, and the requirement of meaningful human oversight all turn on whether a human genuinely decided. If the test for “genuinely decided” is nothing more than “a human signed,” those protections are satisfied by the exact failure they exist to prevent.
Why “the human signed” is the wrong test
Because a signature records that a person approved, not that a person decided. The ceremonial-human failure — a human placed at the end of a process whose substance they cannot reach — produces a signature identical to the one a genuine decision produces. An audit that checks for a human sign-off cannot tell them apart. Worse, the reassurance scales in the wrong direction: as agents become more capable and delegate more, the official is handed cleaner, more finished determinations, and the appearance of a considered human decision improves precisely as the reality of one recedes. The better the agentic chain, the more convincing the ceremony. So “was there a human?” is not merely a weak test; it is a test that reads best exactly when the underlying decision is least human.
What genuine deciding actually requires
An official is deciding, rather than ratifying, only when a bundle of conditions holds. Remove any one and the human role hollows out while keeping its shape:
- Reconstruction — the official can reconstruct how the determination was reached: what the agents did, on what basis, within what authority. (This is the delegation record from the first paper; without it, nothing below is possible.)
- Comprehension — they actually understand the basis, not merely possess a file they did not read. A determination too complex or too voluminous to comprehend in the time available is not being decided.
- Time — there is enough time per case for genuine consideration, not a queue that guarantees rubber-stamping.
- Alternatives — the official can reach a different outcome than the one the chain produced, and that path is real, not nominal. A review that can only confirm is not a decision.
- Authority and protection — their departure from the agents’ output carries structural force and costs them nothing. An official who fears overturning the system’s recommendation is not deciding against it.
- Engagement — on the record, the official actually engaged: considered the specifics, and where they concurred, concurred for stated reasons rather than by default.
Deciding is the presence of all six. Ratifying is a signature with one or more missing. The test does not ask whether the official could in principle have decided; it asks whether, for this decision, the conditions of deciding were actually present.
The test: three questions that locate the line
For any specific decision, three questions place it on the assisting-or-replacing line, and each has an observable answer:
1. Could the official have reached a different outcome — and does the record show they could have?
If the only available action was to confirm, or if overriding was formally possible but practically penalised, unsupported, or untraceable, the human was not deciding. Watch the override rate across comparable decisions: a human step that never diverges from the agents’ output, across a meaningful volume, is assisting nothing — it is being assisted past.
2. Could the official reconstruct and comprehend the basis in the time they had?
If the determination arrived as a finished output the official could not walk back to its reasons, or arrived faster and in greater volume than any human could consider, comprehension was impossible and the decision was, in substance, the chain’s. The tell is the ratio of decisions to review time: when it crosses the point where per-case consideration is physically impossible, ratification is structural, not occasional.
3. Where the official concurred with the agents, is there a recorded reason — or only a signature?
Genuine concurrence leaves a trace of engagement: the official considered this case and agreed for a stated reason. Ratification leaves only assent. A file of determinations in which every human concurrence is a bare approval, with no recorded engagement, is a file of ratifications regardless of how many signatures it holds.
Three affirmative answers — a real alternative, genuine comprehension, recorded engagement — mean the official decided, and the law may treat the decision as human. One or more negatives mean the agents effectively decided and the human ratified, and the decision should be treated, for the citizen’s rights, as what it is: solely automated in substance.
Why the answer must be per-decision, not per-system
A tempting shortcut is to certify a system as “human-in-the-loop” and treat every decision it produces as human. This fails, because the same system produces genuine decisions and ratifications depending on the conditions at the moment of decision. On a quiet day with a comprehensible case, the official decides. On a surge day with a queue and a finished determination, the same official, at the same terminal, ratifies. The line runs through decisions, not through systems, which is why the test is applied to a decision and read across a population of them — through override rates, decision-to-time ratios, and the presence or absence of recorded engagement — rather than granted once to a deployment and assumed forever. A system-level certificate of human oversight is exactly the kind of permanent badge that admissibility rejects: standing that persists past the conditions that justified it.
What live law assumes, and where the test fits
The law already turns on this distinction without giving a way to apply it. Data-protection law’s protection against decisions “based solely on automated processing” hinges on whether human involvement is meaningful or merely token — but supplies no test for which it is, leaving the line to be drawn case by case, usually not at all. The EU AI Act’s Article 14 requires meaningful human oversight of high-risk systems, including those used in essential public services and justice — but “meaningful” is an adjective in search of a test, and under the Digital Omnibus these public-sector high-risk obligations were deferred to December 2027, so even that requirement does not yet bind. The three-question test is offered as the missing operationalisation: a concrete, observable way to decide whether a given human involvement was meaningful, that a regulator, an auditor, or a court could actually apply after the fact. It does not add a new duty; it makes an existing one checkable.
The honest tension
The test must not become a demand that officials override agents to prove they are deciding. An official who agrees with a well-founded determination has decided just as genuinely as one who overturns it — concurrence is a decision when its conditions are met. The point is not a quota of disagreement; it is that the conditions of deciding were present and, where the official concurred, that concurrence was engaged rather than automatic. Nor should the test become a paperwork ritual that manufactures recorded “reasons” to pass an audit while changing nothing — that would simply relocate the ceremony. The safeguard against both is that the three signals are read across a population, not a single file: manufactured engagement and reflexive override both show up as patterns, and both are as visible to the test as honest deciding is. The aim is narrow and firm: an agentic administration may let agents assist, but it may not let them decide while a human is credited with the decision.
What this paper does not settle
Two questions remain open. First, the thresholds — how low an override rate, or how high a decision-to-time ratio, should shift the presumption from deciding to ratifying — are matters for evidence and rulemaking, not assertion, and will vary by decision type; this paper argues that the signals are the right ones, not that any particular cut-off is. Second, the remedy when a decision is found to have been ratified rather than decided — whether it is voidable, must be re-decided by a genuinely deciding human, or triggers a duty to redesign the chain — is a legal question this programme will take up next. The line this paper draws is meant to be usable now; what follows from landing on the wrong side of it is work still to do.
FAQ
What is the difference between an official assisting and an official deciding?
An official is deciding when they could have reached a different outcome, could reconstruct and comprehend the basis in the time they had, and recorded genuine engagement where they concurred. When those conditions are missing, the agents effectively decided and the official ratified — even though a human signed.
Isn’t a human signature enough to make a decision “human”?
No. A signature records approval, not deciding. The ceremonial-human failure produces a signature identical to a genuine decision’s, so a signature cannot distinguish them. The three-question test is what distinguishes deciding from ratifying.
How can you tell, after the fact, which one happened?
Through three observable signals read across comparable decisions: whether the official could and did sometimes reach a different outcome (override rate), whether the basis was reconstructable and comprehensible in the time available (decision-to-time ratio), and whether concurrence carried recorded reasons or only a signature.
Does this require officials to disagree with AI to prove they decided?
No. Agreeing with a well-founded determination is a genuine decision when its conditions are met. The test asks whether the conditions of deciding were present and whether concurrence was engaged — not for a quota of disagreement.
