When work is judged by its paper trail

Premise

Suppose an assessment can know nothing about a person beyond what has been recorded.

This piece starts from that assumption and looks at the present through it. It is an assumption rather than a conclusion, though only just. Whoever judges your work can’t see the effort, the circumstances or the intent behind it. They have to get at it some other way.

Two numbers from the same survey

In 2024 the European Working Conditions Survey asked people whether a computer programme plays a part in how they work.

Among the 26% of respondents who said that some elements of their work are decided by a computer programme, just over half — 15% of all respondents — said there is a clear procedure for raising an issue when they disagree with an automated decision.

That is worth pausing on. For about a quarter of the workforce, software makes decisions about their work, and more than four in ten of them cannot point to a clear way of objecting.

The other figures are more modest than the headlines might suggest. 17% say a programme monitors their performance to some or a large extent; 16% say a programme allocates their tasks; and 10% say a programme decides when they work. In financial services, the figures for monitoring and task allocation rise to 34–35%.

The survey itself adds a note of caution: these forms of management are not that widespread in European workplaces. That caveat belongs with the numbers, so it stays with them here.

What an assessment can actually read

This is not really about software. It is about what happens whenever someone’s work is judged.

Whoever does the judging can’t see the work from the inside. They see what it leaves behind: badge records, emails, tickets, documents, what was said in a meeting. So what an assessment reads is a stand-in for the work, and there is no way around that.

Using a stand-in is not the problem.

The problem is that the person being judged doesn’t get to see what was chosen to stand in for their work.

When a maths test comes back with a wrong answer marked, you can at least see which step went astray. Here nothing comes back to look at. There is just an outcome.

Obvious tasks leave the least behind

Take one plus one. Writing it down feels silly — who needs telling? Recording the obvious looks like padding, and a little patronising to whoever reads it. So it doesn’t get written down.

But once it isn’t written down, nothing shows that it was done. If someone later asks what you did, a task too obvious to note looks exactly like a task you didn’t do.

While things go well, the gap doesn’t show, because the records aren’t being read. They get read once a question has already been asked. And at that point the same record can prove that something was done — and also that this was all you did.

A rule that requires an explanation

The EU’s 2024 directive on platform work does something unusual.

Anyone doing platform work has the right to an explanation, oral or written, of any decision taken or supported by an automated decision-making system, without undue delay and in clear and plain language. They must be able to reach a contact person appointed to discuss the facts, circumstances and reasons behind the decision, and the platform has to make sure that person has the competence, training and authority to do so.

Where a decision affects the essential aspects of the working relationship, the reasons have to be given in a written statement no later than the day the decision takes effect. If the worker asks for a review, the reply must be sufficiently precise and adequately substantiated, in writing, within two weeks. If the decision infringed the worker’s rights, it must be rectified within two weeks.

The transparency rules ask for more than a label. What has to be disclosed includes the main parameters the system takes into account and the relative importance of those parameters, as well as how a person’s data or behaviour affects the decisions. People going through a recruitment or selection procedure must be told before it begins. And the rules on algorithmic management also cover people who do not have an employment contract.

Member States have to bring all this into national law by 2 December 2026.

This is not a recommendation: it comes with a set form and a deadline.

The wider version has been delayed

Something similar was due to arrive for jobs in general. Under the AI Act, systems used to recruit and select staff, to decide promotions and terminations, to allocate tasks, and to monitor and evaluate performance count as high-risk. The obligations for them were due to apply from 2 August 2026.

In July 2026 an amending regulation pushed that date back to 2 December 2027. It gives its reasons: standards, common specifications and guidance were late, and so was the setting-up of the national authorities that are supposed to supervise the rules.

Putting the duty to explain in place is itself running late.

So platform workers get the right to be told the reasons first, and people in ordinary employment get it a year later.

In Belgium, you can ask — on your way out

Belgium already has a rule requiring reasons, at the point of dismissal. It is set out in collective agreement no. 109, which covers private-sector workers with at least six months’ service; earlier contracts count towards the six months.

The details are worth a close look.

The employer does not have to give reasons unless asked. A worker who wants them must request them by registered letter. The employer then has to reply by registered letter, within two months, setting out the concrete grounds.

If the employer misses the deadline, or replies without concrete grounds, it owes a fixed civil fine of two weeks’ pay. If the dismissal is found to be manifestly unreasonable, the worker is separately owed compensation of three to seventeen weeks’ pay. The two can be combined, and both come on top of pay in lieu of notice.

Then comes the clause that alters the whole balance.

If the worker asked and the employer failed to answer properly, it is up to the employer to show that the dismissal was not manifestly unreasonable. If the worker didn’t ask, the burden of proof stays with the worker.

A single registered letter can shift the burden of proof.

The timing is back to front

	While the job continues	On the way out
Employer must give reasons without being asked	not established	no — the worker must ask
If the worker asks	not established	registered request, registered answer within two months
If the employer does not answer	—	two weeks’ pay, and the burden of proof shifts

The reasons don’t come while there is still time to put things right. By the time they do, there is nothing left to fix.

The duty to explain is tied to the end of the relationship.

Keeping records is work too, and it goes uncounted

If assessments run on records, then producing those records is part of the job: writing down what you did, deciding where to file it, choosing whom to tell, naming it so it can be found later. That is different from doing the work itself, and it takes time.

We could not find any public indicator of how much.

One figure does the rounds: that knowledge workers spend 60% of their time on “work about work”. It comes from a company that sells software promising to cut that time. In other words, an interested party is measuring the very problem its product is sold to fix.

This piece does not treat that figure as fact. What can be said is narrower, and more uncomfortable: where official statistics don’t measure something, a number from someone with something to sell fills the gap.

The better the work, the less it may leave behind

There is one more point underneath all this.

Maintenance and prevention work has an odd feature: the better it goes, the less it leaves behind. No outage means no outage report. A problem headed off before it develops leaves no incident to point to.

Someone can put in the same hours, and the better the result, the less there is for an assessment to read.

That says something about the measurement, not about the person.

Record-keeping as insurance

When something goes wrong, the record is what you fall back on. When nothing does, it sits unopened, taking up space.

And when you write something down, you can’t know which record will turn out to matter. So you write down a lot, and that breadth is the cost: in hours, and in storage.

The cost is certain and ongoing. The payout is rare, and it comes only after whoever is doing the reading has decided the record was needed.

That isn’t laziness, or waste. It is a sensible response to not knowing which record will be asked for.

Is it a fair deal?

This piece doesn’t have the answer. What it can do is set out what would need to be in place for it to be fair.

People would need to know the criteria in advance, and to be given the grounds afterwards.

Now set what has been established against that.

For people still in their jobs, we could not find data on either. On the way out, the reasons come if the worker asks for them — and if they are asked for and don’t come, the burden of proof moves to the employer. In platform work, from December 2026, the main parameters and their relative weight must be disclosed in advance, and a decision must be explained in writing within two weeks.

Same question, different answer depending on where you are.

Where the author stands

The work I did and the things an assessment reads did not take the same form.

I was not told which form was the right one. That part didn’t come back to me.

What remains possible is more limited. Even without being told what the question was, I can leave the work in a form that someone else can check: writing down what the evidence was, separating what was established from what wasn’t, and keeping a record of mistakes instead of deleting them.

That isn’t a set of model answers. Anything made public gets read again, against other criteria, and filtered again.

It is simply a way of not leaving the whole story of my work to a single verdict.

This isn’t advice. It’s a note of what I’m keeping an eye on.

What I got wrong in the research

I made the same mistake twice.

I took the article numbers for the platform work directive from secondary summaries. Three of them agreed, and I treated that agreement as proof. When I opened the official text, four of the numbers were wrong. The same happened with the survey figures: four numbers taken from search summaries didn’t match the report.

Both times, I trusted several sources because they said the same thing. Sources that copy each other agree without adding anything.

I also wrote that the AI Act’s employment obligations were already in force. They had in fact been pushed back to the following year.

The provisions and figures left in this piece have been checked against the official texts.

What this piece established

The platform work directive gives people a right to an explanation of automated decisions, a written statement of reasons for decisions affecting essential aspects of the working relationship, and a substantiated written answer to a review request within two weeks. Member States must transpose it by 2 December 2026
The AI Act’s high-risk obligations covering employment were postponed to 2 December 2027, and the amending regulation puts the delay down to late standards, guidance and national authorities
Of the 26% who said parts of their work are decided by a computer programme, 15% of all respondents said there is a clear procedure for raising a disagreement
In Belgium, workers must ask for the reasons for their dismissal by registered letter. If the employer does not answer properly, it is up to the employer to show that the dismissal was not manifestly unreasonable

What this piece does not claim

That employers want surveillance
That keeping more records has led to more dismissals
That AI is causing this change
That any particular way of behaving will keep a person safe