The two previous articles describe problems that don't sit well with manual work: the decisions that expire when the rules change and what stops being true when a fact changes.
This one describes the reverse, and not out of modesty: there is a part of the work that no system solves, and confusing it with the other part is the fastest way to end up with a flawless tool and an untouched problem.
The line is fairly sharp.
What a system does solve, and it isn't a small thing
This deserves to be said first, because the lazy conclusion to an article like this would be "so none of it matters".
The record that doesn't forget. What was decided, when, by whom, and about what. This is exactly what an organisation cannot reconstruct afterwards, because the memory of a decision lasts less than its consequences.
The relationships between decisions. Which document depends on which classification, which classification rested on which article. That network is what makes it possible to answer "what does this change affect?" — a question that, done by hand, only whoever remembers the entire chain can answer.
And detection of what has expired. If every decision knows what it rested on, a change at the source can flag what needs reviewing. Without that relationship, the work is either review everything or review nothing.
The three have something in common: they are problems of memory and of linking, not of judgement. And they are exactly the ones an organisation loses to the passage of time, not to a lack of capability.
What it doesn't solve, and the article says so
Here is the other side, and it has support in the text.
Deciding what a system is for. The intended purpose isn't inferred from the software: it's decided. And the entire classification hangs from it.
Assigning oversight. Article 26(2) requires assigning it "to natural persons who have the necessary competence, training and authority". A tool can record who has been designated; it cannot give anyone authority. If the person designated cannot stop the system without asking permission, no feature fixes that.
Adopting proportionate measures. Article 4 requires measures "taking into account their technical knowledge, experience, education and training and the context the AI systems are to be used in". That judgement of proportion belongs to whoever knows their own team.
And deciding what's acceptable. What a system can do without intervention, what always gets reviewed, what doesn't get automated. These are risk decisions, and the risk is borne by an organisation, not by a product.
The line, in one sentence
A system can hold the trail of a decision. It cannot hold the decision.
And because accountability travels with the decision, not with the trail, no tool displaces it. It's the same reason why the conformity seal so many people are looking for doesn't exist: what can be demonstrated is diligence, and diligence belongs to whoever decides.
How the difference shows up in practice
With two questions, and they work both for evaluating a tool and for evaluating the process itself.
What happens when something changes? If the answer is "someone remembers", the memory problem is still open, and that's where a system helps. If the answer is "what's affected gets flagged and someone reviews it", it's solved in the part that can be solved.
Who is accountable for this decision? If the answer is a person's name, together with what they can and can't do, that's fine. If the answer is a product's name, there's a problem that no product fixes — and one that will become visible the day someone asks, because the question gets put to the organisation.
Why this matters before deciding anything
Because the two halves are bought in different ways.
The memory-and-linking half is a tooling problem: it grows over time, degrades silently, and is solved with structure.
The judgement-and-accountability half doesn't improve by buying. It improves by deciding: who is accountable for what, what counts as acceptable, what gets reviewed. It can be done with a spreadsheet, and without those decisions, no system fixes it, however good it is.
So the sensible order is the one almost nobody follows: decisions first, which are free and hard; then the structure to keep from losing them, which is expensive and easy. Do it the other way round and you get a tidy file for a judgement nobody made.
Content in accordance with Articles 4 and 26 of Regulation (EU) 2024/1689, as amended by Regulation (EU) 2026/1744 (Official Journal of the EU, 24 July 2026).
This article is for informational purposes only and does not constitute legal advice.