20 Years of the AGG – and Now the Algorithm Gets a Say

Marcus

Aug 26

7 min read

On 18 August 2006, a law came into force in Germany that had primarily been portrayed as a bureaucratic monster: the General Equal Treatment Act (Allgemeines Gleichbehandlungsgesetz, AGG). Critics predicted waves of lawsuits, mountains of paperwork and a labour market so afraid of litigation that employers would barely dare to reject anyone. Twenty years later, the sober conclusion is this: the wave of lawsuits never came, the law stayed. And recruiting practice learned.

Most of the classic sources of error have now been cleared away. Hardly any German job ad appears without “(m/f/d)”, age requirements have disappeared from job profiles, application photos have become optional, and structured interview guides are standard. Anyone still writing “young, dynamic team” today is doing so not out of ignorance, but out of carelessness.

There is just one catch: in more and more processes, the decision about who even makes it as far as the interview guide is no longer made by someone who has attended AGG training. It is made by a ranking. And that is where the real anniversary question begins.

What twenty years of the AGG have actually changed

The figures are worth a look because they show both the impact and the limits of the law. In 2025, Germany’s Federal Anti-Discrimination Agency recorded a record 13,067 counselling enquiries, 15 per cent more than in the previous year.

The largest single block came from working life: around 3,600 enquiries, including discriminatory job advertisements, rejections during recruitment processes and harassment at work. By characteristic, racism accounted for 43 per cent, followed by disability and chronic illness at 27 per cent and gender at 22 per cent. Age accounted for 12 per cent.

Rising numbers are not necessarily a sign of failure, but of awareness: people increasingly know that they have rights and where to turn. The other side of the statistics is equally noteworthy: the feared avalanche of litigation never materialised. The AGG works primarily as a preventive instrument, through wording, processes and documentation. It is a law that does its job when organisations observe it, not when people have to sue under it.

The section every TA team should know

Anyone who wants to understand why algorithmic pre-selection is legally more delicate than human selection only needs to read one section: Section 22 AGG. It governs the burden of proof, and it does so in favor of applicants. Once an applicant presents facts that suggest discrimination based on a protected characteristic, their part is done. The employer must then prove that no violation occurred. Not make it plausible. Prove it.

The Baden-Württemberg Regional Labor Court demonstrated just how low that evidentiary threshold can be (judgment of 7 November 2024, case 17 Sa 2/24). A communications manager born in 1972 applied for a role seeking a “Digital Native” and an “absolute team buddy”. He met the professional requirements, was rejected and was awarded €7,500 in compensation. The court’s reasoning was that “Digital Native” describes people who grew up with digital technology and therefore carries an age and generational reference that cannot simply be argued away. Two English words were enough to constitute an indication of discrimination. The employer was unable to rebut it.

And now for the uncomfortable translation into algorithmic recruiting: a score is also a form of wording – just one that nobody can read aloud.

Why the black box hurts the employer, not the applicant

The common reflex is:

An algorithm is neutral; it knows neither age nor origin. Formally, that may be true. In practice, it is irrelevant.

What matters are proxy variables: characteristics that correlate with protected characteristics without naming them. Year of graduation. A cap on years of professional experience. Gaps in a CV. The software someone has worked with. Place of residence. The language variety used in a cover letter. Anyone optimising for “cultural fit” while training on successful existing employees is building a machine that reproduces the current workforce structure. That is not a bug. It is the objective function.

Legally, this is the territory of indirect discrimination: a neutral criterion that in practice disadvantages a protected group. This is where the black-box logic turns against the employer. If a conspicuous pattern in rejection rates can serve as an indication – for example, applicants over 50 passing the screening significantly less often than younger candidates with comparable profiles – the burden shifts to the employer. “We do not know why the system decided that way” is not a defence at that point. It is an admission that the employer cannot provide the counter-evidence.

Responsibility cannot simply be passed on to the vendor either. Under employment law, the employer remains responsible for discrimination in its own selection process, even if the technical pre-selection is carried out by a third party.

The same idea is currently being tested at scale in the United States. In Mobley v. Workday, the judge held in March 2026 that an employer cannot escape liability simply by delegating traditional employer functions such as hiring to a third party. The class action covers applicants aged 40 and over since September 2020; court filings refer to 1.1 billion applications rejected through the platform. US law is not German law, but the underlying question is the same – and it is heading our way too.

A brief look across the border

Because this blog is not only read in Germany, the findings cannot simply be transferred one-to-one. Austria has a very similar framework under its Equal Treatment Act, including an easing of the burden of proof for those affected and the Equal Treatment Ombud’s responsibility for these issues. Organisations recruiting in both markets can therefore apply largely the same review logic.

Switzerland is different. There is no general anti-discrimination act governing private employment relationships; explicit statutory protection is largely centred on equality between women and men. That may sound like less risk, but it is not a free pass. Organisations recruiting into the EU from Switzerland, working within a group with a German parent company or using software from an EU provider can still encounter the same requirements through the supply chain. And the reputational part of the problem – a selection process that cannot explain its own outcomes – has no national border anyway.

The anniversary reform – and what it does not regulate

Fittingly for the anniversary, the German Federal Cabinet approved an amendment to the AGG on 6 May 2026. For recruiting, one change matters above all: the deadline for asserting claims under Section 15(4) AGG is to double from two to four months. The reform also includes broader protection against sexual harassment outside employment relationships, a narrower version of the church exemption and a change in terminology from “Alter” to “Lebensalter” for age.

In the professional debate, this is widely regarded as a mini-reform. It contains neither a right for associations to bring collective actions nor the one-year deadline called for by the Federal Anti-Discrimination Agency. And the draft contains no separate provision for algorithmic decisions at all.

For practice, this means two things. First, the window in which a rejected application can turn into a legal case doubles – screening logs, scores and reasons for rejection therefore need to remain traceable for longer. Second, anyone waiting for a specific statutory clarification on AI is looking in the wrong place. The AGG is technology-neutral. It already applies today, to every ranking and every applicant tracking system. The EU AI Act’s high-risk obligations for recruiting systems have meanwhile been postponed until December 2027 – but the prohibition of discrimination was never dependent on that timetable.

The review checklist that follows from twenty years of case law

Six questions that belong in every review of an AI-supported selection process:

Which characteristics actually enter the model? Not which ones appear in the marketing material. Year of graduation, caps on years of experience, postcode and CV gaps are all candidates for a proxy-variable discussion.
Who looks at who gets screened out? Almost every team reviews the shortlist. Hardly any systematically reviews the rejections. That is exactly where the evidence can emerge.
Do you analyse outcomes by relevant characteristics? A simple rate comparison across age and gender groups in the funnel takes half a day and can become some of the strongest exculpatory evidence available in a dispute.
Is human review documented or merely claimed? “A human checks it” only helps if you can show when, by whom and with what outcome – including the cases in which the system’s recommendation was overridden.
What does the vendor contract say? Audit rights, access to bias-testing results and assurances on non-discrimination. If none of this is contractually agreed, the employer carries the risk alone.
Does the job-ad language pass the “Digital Native” test? “Digital Native”, “entry-level candidate with initial experience”, “high-energy”: anything that points to an age cohort rather than a capability can become an indication of discrimination with a price tag.

What remains

In twenty years, the AGG has achieved something that is easy to overlook precisely because it has become invisible: it changed the language of recruiting. Phrases that nobody would have thought twice about in 2006 now instinctively raise eyebrows.

The next round is harder because the decisive wording no longer sits in the job ad; it sits in the model. Yet the task is still the same: being able to explain why someone did not progress. Anyone who sees this merely as a compliance issue is underestimating it. It is a fundamental question of the profession.

Sources

Transparency note on the use of AI

AI-based tools were used to support the creation of this article. They were used primarily for research and information structuring, language and grammar editing, translation, and in some cases for creating or editing illustrations and visual material.

The concept, substantive statements, assessments and conclusions are the author’s own and reflect his personal views, professional experience and convictions. All content is editorially reviewed and approved before publication.

Articles on this website are not created or published automatically by AI. AI is used as a supporting tool; full substantive and editorial responsibility remains with the author.