Meaningful human intervention: when does a signature on an AI decision count?

Two kinds of decision, two very different regimes

Before you can say anything meaningful about human intervention, one distinction must be clear, and it is precisely the distinction that is missing from almost everything written on the subject. Legally, it makes all the difference whether an AI system prepares a decision about a thing or about a person.

If it concerns a thing, Article 22 GDPR does not come into play. A setpoint in the greenhouse, a trip sequence, the release of a batch, a field from a consignment note: there is no data subject, and therefore no automated decision within the meaning of that article. That does not make human checks pointless there. It means the basis for them is your own liability, your quality scheme and your professional expertise, not privacy law. Naming that difference honestly is better than suggesting a legal obligation that does not exist.

If it concerns a person, the picture changes completely. Who is rostered, who gets which task, who is rejected, who is put on an improvement plan: these are decisions about natural persons, and Article 22 GDPR already applies to them today, without a transition period and without postponement. What exactly that means for recruitment and HR policy is set out in our article on AI in recruitment and workforce planning.

• Decisions about a thing: energy control, route optimisation, rejecting a package, extracting a field from a document. Article 22 GDPR plays no role here.
• Decisions about a person: recruitment and selection, pay grading, task allocation based on individual behaviour, performance appraisal, rostering that affects income. Article 22 GDPR already applies here, to the extent that the decision is based solely on automated processing.
• Borderline cases: the same planning software that today distributes trips across vehicles and tomorrow assigns trips to drivers based on their past performance. The same screen, the same planner, a different regime.

What the test really is

The standard does not come from the text of the law itself but from the guidelines on automated decision-making and profiling that the European data protection authorities adopted in 2017 and last revised in February 2018, known as WP251rev.01. Two passages from them do the work.

The first closes the back door. The controller cannot avoid the Article 22 provisions by fabricating human involvement: if someone routinely applies automatically generated profiles to individuals without any actual influence on the result, that is still a decision based solely on automated processing.

The second states the requirement in positive terms. To qualify as genuine human involvement, oversight of the decision must be meaningful, 'rather than just a token gesture'. It must be carried out by someone 'who has the authority and competence to change the decision', and that person must consider all the relevant data in the analysis.

There is a third sentence, and it is the one most often skipped, although in practice it delivers the most: as part of its data protection impact assessment (DPIA), the controller must identify and record the degree of human involvement and the stage of the process at which it takes place. The duty to record is therefore neither new nor optional.

The two cases in which the regulator says: this does not count

The Autoriteit Persoonsgegevens (Dutch Data Protection Authority, AP) made this concrete in its guidance on online and game-based assessments in recruitment and selection. It first sets out the positive criteria. The reviewer must be competent, which means understanding the logic behind the automated system and grasping, at least in outline, how the decision comes about. And the reviewer must be authorised: able to change or overrule the outcome, and to take other relevant information into account, including data the system has not processed. The AP adds that this competence and authority must not only be granted on paper, but must also actually be exercised in practice.

Then come two examples in which, according to the AP, there is no meaningful human intervention, and these two are the sharpest tool for a self-assessment. The first: the reviewer is formally involved in the process, but has no discretion of their own and must always follow the automated outcome. The second: the reviewer's only role is to convert a particular score from the computer into a negative decision.

Read that second example again alongside your own way of working. An employee who adopts a system outcome and puts their name to it fits that description exactly. The signature does not make the decision human; the ability to turn it into something else does.

One nuance needs to be added straight away, because most articles miss it. The AP points out that the human intervention a data subject can request after an automated decision is different from human intervention during the decision-making process. The first is a safeguard for a decision that was already permitted under an exception; the second prevents the prohibition from coming into play at all. Anyone who confuses the two thinks an objection procedure is enough, while the problem lies in the process itself.

What a Dutch court made of it

This is no longer theory. On 4 April 2023, the Gerechtshof Amsterdam (Amsterdam Court of Appeal) gave judgment in three related cases brought by drivers against Uber and Ola about access to information on automated decision-making. The case with reference ECLI:NL:GHAMS:2023:793 concerned drivers whose accounts had been permanently deactivated after a fraud signal, and the question of whether those deactivation decisions were based solely on automated processing.

The court adopted the three elements from the guidelines word for word: oversight must be meaningful rather than just a token gesture, it must be carried out by someone with the authority and competence to change the decision, and that person must consider all the relevant data in the analysis.

Uber argued that in each case one or more members of its Risk team had carried out a manual investigation after the fraud signal, and submitted documents to that effect. The court held that those documents did show that one, two or three employees had made a note, but 'in no way' (translated) demonstrated that this act had been much more than a mere token gesture, or that those employees had considered all the relevant data in their analysis. Uber had also not made clear what the qualifications and level of knowledge of the employees concerned were.

Just as important is where the court placed the burden of proof. Uber bore a heightened duty to substantiate when disputing the claim that the decisions were based solely on automated processing, because the information about any human intervention lies largely within Uber's own domain. That reasoning is general, not something that applies only to platform companies: whoever holds the system, the logs and the work instructions must explain what happened.

Then there is the detail that captures the whole case in a single image. For three of the four drivers, Uber did not manage to do so. For the fourth it did, and the reason was surprisingly simple: a personal conversation had taken place with him before the deactivation. Two employees with internal procedures and written notes were not enough; one conversation with the person concerned was. To put this in perspective, the district court at first instance had held, on comparable facts, that there had been meaningful human intervention in all cases. So this is a real test on which opinions can differ, not a foregone conclusion.

When the score itself is already the decision

There is another ruling that shifts the centre of gravity, and it is less well known than it deserves to be. On 7 December 2023, the Court of Justice of the European Union ruled in Case C-634/21 on credit scores. A credit information agency automatically calculates a probability value for whether someone will be able to meet their payment obligations in the future, and passes it on to a bank.

The Court held that establishing that value is itself an automated decision within the meaning of Article 22(1) GDPR where 'a third party ... draws strongly on that probability value to establish, implement or terminate a contractual relationship with that person'. What was decisive was factual, not formal: according to the findings of the referring court, an insufficient value leads the bank to refuse the loan in almost all cases, so the value plays a determining role. In those circumstances, establishing the value must in itself be regarded as the decision. The Court also characterised Article 22(1) as a provision that lays down a prohibition in principle.

Note carefully what this judgment does not say. The Court did not hold that a decision is still automated when a person signs it off without real influence. That reasoning comes from the guidelines and from the Amsterdam ruling above. What the judgment does do is shift the legal centre of gravity to the point where the outcome is in fact determined, rather than to the point where the signature sits.

For a buyer, this has a consequence that is rarely discussed. In the case the Court assessed, the Article 22 obligation rested on the agency that calculated the score, and not only on the bank that used it. If you buy a system that scores employees or candidates, and your organisation follows that score almost always in practice, ask your vendor what role it sees for itself. An answer along the lines of 'we only provide advice' becomes weaker the more that advice determines the outcome in practice.

Six things that quietly hollow out your intervention

On 23 July 2025, the AP published practical guidance on meaningful human intervention in algorithmic decision-making. The document is sector-neutral, gives effect to the AP's awareness-raising task under Article 57 GDPR, and is expressly 'not a checklist' (translated). It structures the question along four axes: human, technology and design, process, and governance. Six of the factors it names almost always recur in operational environments.

• Workload. The AP puts it bluntly: the less time a reviewer has to reach a decision, the less meaningful the human intervention will be. An organisational culture with a set number of decisions to review per day, or other targets, works against it.
• Routine. If a reviewer's work becomes routine without actually influencing the result, it loses its function, according to the AP. It is the same idea as in the guidelines, but framed as a design question.
• Timing. Does the intervention only take place as a check on the output, at the end of the process? The AP notes that the decision then appears, in effect, to have been taken already, and that the algorithm actually plays a decisive role. Human intervention earlier in the process carries more weight.
• A model that cannot be followed. The AP points out that an algorithm can be so opaque that a reviewer, even with training, cannot know how it arrived at a result. In that case the reviewer cannot carry out the human intervention meaningfully. A black box is then not an acceptable residual risk but a barrier.
• How the system is talked about internally. Algorithms are often framed as more reliable or more precise than people, and that framing contributes to automation bias. The AP therefore considers it relevant to examine how an algorithm is talked about within an organisation: as a crucial component or as a supporting tool.
• What deviating costs the reviewer. If someone is held personally responsible for an incorrect decision, they are less likely to go against the algorithm. The AP therefore considers it desirable that a reviewer is authorised to go against the algorithm, but is not penalised for a mistake.

Two measures the AP itself proposes

The document does not stop at diagnosis. Two design measures transfer directly to an operational environment. The first: regularly give the reviewer a decision that has to be made entirely without the algorithm. The second: add a random sample of cases that were not selected to the cases the algorithm did select. According to the AP, this ensures that the reviewer cannot rely blindly on the algorithmic output.

One sentence from the same document deserves separate attention, because it directly refutes a common sales argument: the fact that an algorithm is, for example, very accurate does not mean that human intervention is no longer required. Accuracy and lawfulness are two different questions.

Finally, the AP warns against tunnel vision on the process of human intervention, and we agree. All the factors and questions together can mask a bigger problem, namely that the decision itself is unethical, or that using an algorithm for that decision is morally problematic. That question comes first.

How this plays out in planning, on the line and in the roster

The three situations below are illustrative, not client stories. They show that the same question gets three different answers in your three most important processes.

The planner who approves a proposed route plan falls outside Article 22 GDPR: the decision concerns trips and vehicles, not a person. Article 14 of the AI Act does not apply here either, because ordinary route optimisation is not listed in Annex III. What is real, however, is the haulier's liability for driving and rest times, regardless of who proposed the trip; we wrote about this earlier in our article on route optimisation and driving times. The tipping point lies elsewhere: as soon as the same screen assigns trips to named drivers based on their behaviour or performance, it is a decision about people. In that case Article 22 GDPR applies today, and the high-risk rules under point 4(b) of Annex III apply from 2 December 2027. That date is not in the original text of the AI Act: its Article 113 still gave 2 August 2026 for the Annex III systems. The move to 2 December 2027 comes from Regulation (EU) 2026/1744, the omnibus package that entered into force on 27 July 2026.

The quality officer who signs off an AI rejection does not fall under Article 22 either: a batch of product is not a natural person. Yet almost every one of the AP's quality criteria is failed here at once, and that is an audit question rather than a legal one. Two hundred confirmations per shift, each lasting a few seconds, are routine by definition; the reviewer only sees the output at the end of the process; and whether they understand, in outline, how the model reaches its verdict is rarely tested. Precisely those three points determine whether your control measure holds up in an audit, as we set out in our articles on reject classification on the packaging line and HACCP and AI. The AP's two measures can be carried over literally to this setting: periodically send batches to the reviewer without the system's verdict, and add a random sample of unflagged batches to the stream for review.

The rostering or performance system is the only one of the three where Article 22 GDPR applies with full force today. If a system scores warehouse staff on picking speed and reliability, and this leads to a shift allocation that affects income, or to a flag that leads to an improvement plan, that is a decision about a person. The AP's two negative examples settle it: does the team leader have discretion of their own, or do they convert a score into a decision? The Amsterdam ruling provides the lesson on evidence: notes in a file were not enough, the reviewers' qualifications had to be shown, and one conversation with the person concerned beforehand did make the difference.

Article 22 GDPR and Article 14 of the AI Act are not the same

These two are constantly confused, including in vendor material, and that leads to the wrong timetable. They are different instruments, with a different trigger, a different addressee and a different timing.

• Article 22 GDPR is triggered by a decision based solely on automated processing, about a natural person, with legal effects or a similarly significant effect. It applies to every controller, in every sector, and has applied since 2018. It is a prohibition in principle: meaningful human intervention ensures that the prohibition does not come into play.
• Article 14 of the AI Act is triggered when a system is classified as high-risk. It is a design and use requirement, not an escape route. Among other things, it requires the people overseeing the system to remain aware of the tendency to rely automatically on its output, and to be able to decide not to use the system or to disregard, override or reverse the output.
• On top of that, Article 26(2) requires the deployer to assign human oversight to people with the necessary competence, training and authority, as well as the necessary support. For Annex III systems, these requirements apply from 2 December 2027.
• What stands out is how close these sources are, even though each text chooses slightly different words. The guidelines ask for someone with 'the authority and competence' to change the decision, the AP tests for competence and authority, Article 26(2) of the AI Act speaks of 'competence, training and authority' plus the necessary support, and Article 14(5) uses the same series, although its Dutch-language version has 'bevoegdheid' (power to act) where Article 26(2) has 'autoriteit' (authority). Anyone who now organises their process around competence, authority and support for decisions about people has already done most of the work for 2027.

What to record, and why that is the heart of the matter

All the sources above point the same way: what matters is not what you have set up, but what you can show about it. The guidelines call for identifying and recording the degree of human intervention and the stage at which it takes place. The Amsterdam ruling shows what happens when that record does not exist, or shows no more than that someone made a note.

Six things are enough for this, and they fit into the system you probably already use:

• Who the reviewer is, what training they have had and when. The court explicitly asked about qualifications and level of knowledge.
• What information the reviewer was shown, and what information the system did not have but was relevant.
• How much time was available for the assessment, and whether there is a target number per day or per shift.
• At what point in the process the intervention takes place: before the outcome, or afterwards as a check.
• How often the reviewer deviated from the system outcome, and on what grounds. This is the most useful number in the whole list: if it stays at zero for months, you have a finding before a regulator or auditor does.
• What deviating means for the reviewer: is it valued, recorded or penalised?

Where to start in practice

Start with a list of decisions, not of systems. That is the reversal this subject calls for, because most inventories start with the software and therefore miss exactly the decisions that are prepared by a third party. Mark each decision as being about a thing or about a person.

Then take only the decisions about people and answer three questions for each: can the reviewer demonstrably change the outcome, do they have information the system does not have, and do they have the time to weigh it? Three yeses are the start of a defensible process. A single no is a concrete point for improvement, not a reason to switch the system off straight away.

Next, measure for one month how often reviewers deviate. It costs little and yields more insight than any policy document. If the answer is zero, you know what the conversation needs to be about.

And make sure the knowledge is in place. The AP writes that when artificial intelligence is used, a reviewer must have sufficient AI literacy and must understand how the technology works at a basic level, its limitations, and its possible benefits and risks. Since this summer, that is no longer the same as what the AI Act requires, and that difference is precisely the pitfall. Article 4 of the Act, applicable since 2 February 2025, originally required 'a sufficient level of AI literacy'. Regulation (EU) 2026/1744, in force since 27 July 2026, rewrote that provision as a duty 'to support the development of AI literacy', with the express addition: 'This obligation does not require providers or deployers to guarantee any specific level of AI literacy of any individual.' The two tracks therefore diverge rather than converge. Supporting measures under Article 4 do not in themselves meet what the AP asks of your reviewer under the GDPR, because that requirement concerns an actual level of knowledge in the person who signs. It remains the cheapest step in this whole list, but you take it for the GDPR, not only for the AI Act. How to approach the broader classification and documentation question is covered in our article on the AI Act in operational processes.