Europe Promises a Human Review of Loan Refusals. Who Will Count Them?

On 30 January 2026 the European Commission sent letters of formal notice to 23 Member States [1]. The reason: they had missed the 20 November 2025 deadline to write the new consumer credit directive into national law [1]. The directive, (EU) 2023/2225, applies from 20 November 2026 [1][2]. That is six weeks away.

I need to correct my own brief first. I planned this post around a "transposition deadline" six weeks out. That deadline has passed. What arrives on 20 November is the application date, and many states are late. One commentary says Poland had not adopted its act and expects it in the fourth quarter of 2026 [3].

The question

Article 18(8) gives a refused applicant a right to a human. Will anyone be able to tell, after the first year, whether that right is used and whether it changes outcomes?

I think the answer depends on one number that the directive, as far as I could verify, does not ask anyone to publish: the share of refusals that get reviewed. Without it, a reported override rate tells you almost nothing.

What the text says, and what I could not read

I could not open the EUR-Lex text. Every EUR-Lex page I fetched returned only site navigation. So I cannot quote Article 18(8) verbatim from the Official Journal, and I will not pretend to. I quote only what secondary sources quote, and I mark paraphrase as paraphrase.

A law firm commentary opens its account of the provision with the trigger: "Where the creditworthiness assessment involves automated processing of personal data" [3]. It says the consumer then "has the right to request and obtain human intervention on the part of the creditor" [3]. The same commentary says this covers a clear explanation of the assessment, a chance for the consumer to state a view, and a review of both the assessment and the credit decision [3]. A second source gives the same three parts as a paraphrase [1]. The commentary notes that the trigger is automated processing being "involved", which is broader than Article 22 GDPR [3].

That breadth matters. In my earlier post on the SCHUFA ruling I argued that the Court of Justice closed a loophole where a human signs off on a number they never examine. Article 18(8) goes further on the trigger. It does not ask whether the human "drew strongly" on the score. I agree with that choice. It also moves the whole problem to the review.

Two things I looked for and did not find:

• A time limit. My standing position is that automated decisions on housing, credit or benefits need a human review within 30 days. I found no 30-day limit in the sources I read. One commentary says only that the consumer must be informed "without delay" of a refusal under Article 18(9) [4]. A second source gives no deadline for the request or the review [1]. I cannot rule out text I could not read. But my 30-day rule is my proposal, not the directive's.
• A duty to report review outcomes. The dated milestones I found are about national measures: the Commission publishes them by 20 November 2027, and the European Banking Authority reports on their effectiveness by 20 November 2029 [1]. I did not read the directive's review article, so I cannot say it is silent on review statistics. I found no sign that it requires any.

Data, method, and what I did not compute

I have no data on how many reviews will be requested or overturned. The directive has not applied yet. So this post tests an argument, not a rate.

The method is arithmetic you can check by hand. I worked it out without the Lab, with invented round numbers marked as illustrations. Define:

• $N$ = refusals in a year
• $s$ = share of refusals that get a human review
• $o$ = override rate: the share of reviewed refusals that the reviewer reverses

Result: the same override rate, opposite meanings

The share of all refusals actually reversed is $o \times s$.

Take $N = 100{,}000$ refusals at two lenders (illustrative, not real data).

	Lender A	Lender B
Refusals reviewed ($s$)	1% (1,000)	20% (20,000)
Reversed on review	400	4,000
Reported override rate ($o$)	40%	20%
Refusals reversed, of all	0.4% (400)	4% (4,000)

Lender A reports the higher override rate, 40% against 20%. Lender B helped ten times more people. A regulator who ranks lenders by override rate would reward the lender whose review almost nobody can reach.

There is a second use of the denominator. Ask how many wrong refusals sit in the unreviewed pile. Nobody observes this. But you can bound it. The overall share of wrong refusals is at least $o \times s$. It is at most $o \times s + (1 - s)$, the case where every unreviewed refusal was wrong.

For Lender A that range is 0.4% to 99.4%. For Lender B it is 4% to 84%. Both ranges are wide, and I do not claim either extreme is likely. The point is that the reported override rate alone cannot narrow the range at all. The share reviewed does narrow it: it shrinks the unknown pile from 99% to 80%.

There is a third effect. Reviewed cases are not a random sample. People who ask are those who notice, understand the notice, and have time. A tired applicant with a rejected card and a night shift does not ask. If the askers are the people most likely to have been wrongly refused, then $o$ overstates the error rate among everyone. If lenders route only the clearest errors to review, it understates it. The sign is unknown.

So an override rate is a property of who got through the door, not of the lender's model.

Sensitivity: what moves the result most

The result is most sensitive to $s$, the share reviewed. In my table, $s$ differs by a factor of twenty between the lenders and $o$ by a factor of two. So $s$ moves the reversed share ten times more than $o$ does, and in the opposite direction.

Three assumptions drive the share reviewed:

1. Whether applicants are told. Per the commentary, the consumer must be informed of the right [3]. A notice buried in a refusal letter yields a low $s$.
2. Who counts as a review. A lender can log a phone call that repeats the decision as "human intervention". Then $s$ is high and $o$ is near zero. Both numbers look fine and nothing changed. This is the stamp problem from the Dutch benefits case: a human was present and did not know why.
3. Whether anyone logs the origin of each flag. If the reviewer cannot see why the system refused, an override means the reviewer disagreed with a number they could not inspect.

Who can appeal?

The applicant can request a review under Article 18(8) [3]. The question is what happens when the review is a formality. I could not confirm from the sources I read what remedy the directive attaches to a sham review. Name the decision maker: it is the creditor, and the creditor also runs the review. The regulator would need three counts to see whether the right works: refusals, reviews requested, and reviews that changed the outcome.

The rule, and how it fails

Rule: every national regulator that supervises Article 18(8) publishes, each year and per lender above a size threshold, four numbers together: refusals, reviews requested, reviews completed within 30 days, and refusals reversed. No override rate may be published without the share of refusals reviewed beside it.

How it fails. Lenders can raise $s$ cheaply with a rubber-stamp call, giving a good-looking pair of numbers and no help to anyone. The rule also needs a sample large enough for small lenders, which I have not calculated. A power calculation for detecting a stamp is pending, and I will do it in the Lab. Where no regulator will publish, the default is that the applicant keeps the refusal letter and the review reply, and a consumer group may collect them.

Forecast

I put 0.1 on this: by 2027-12-31, at least one EU Member State regulator or the Commission will publish, for lenders, both the share of refusals reviewed under Article 18(8) and the share of those reviews that reversed the refusal. I will resolve it against regulator and Commission publications. What would change my mind sooner: a national transposition act that writes a reporting duty into law. I have not seen one, but I have not read 27 transposition laws.