What an Arbeitszeugnis is, for readers outside Germany
In most countries a reference is optional, short, and increasingly replaced by a confirmation of employment dates. In Germany the reference letter is a formal document with a fixed structure: company letterhead, personal data, a complete description of the tasks, an assessment of performance in several categories, an assessment of conduct towards superiors, colleagues, and customers, a reason for leaving where appropriate, and a closing formula with thanks and good wishes. It is signed by a person senior to the employee, printed on paper, and handed over with the last payslip.
Employees take it seriously because future employers in Germany read it, and read it in a particular way. A missing thanks in the closing formula, a conduct clause that names colleagues before superiors, or the absence of a sentence about reliability are not oversights in the reader's eyes. They are signals. That is why a letter that a manager considers fine can trigger a legal dispute, and why HR teams in companies with 50 to 500 employees spend, in my experience, two to three hours per letter when writing by hand, often more for senior roles.
Section 109 GewO: benevolent and truthful
Section 109 of the Trade Regulation Act (Gewerbeordnung, GewO) establishes the right to a reference and sets its two conditions: it must be clear and comprehensible, and it must not contain features or formulations intended to make a statement about the employee other than what the wording says. Case law adds the principle that has shaped the whole practice: the letter must be benevolent, so as not to obstruct the employee's future career, and it must be truthful, so that the next employer is not misled.
Those two duties pull in opposite directions, and German practice resolved the tension by developing a language in which a benevolent surface carries a truthful grade underneath. The labour courts have accepted this system to the extent that they routinely decide disputes by reference to it. An employer who wants to say that performance was average writes a sentence that reads pleasantly and is understood by every German HR professional as a grade of three. An employer who writes a flat negative sentence has breached the benevolence duty. An employer who writes a glowing letter about an employee who was dismissed for cause has breached the truth duty, and could in theory be liable to the next employer.
The coded language, in brief
The grade is carried by the intensity adverb in the summary performance sentence. The formula that translates roughly as performed the tasks to our fullest satisfaction at all times is the top grade. Drop at all times and it becomes a two. Drop fullest and keep full and it becomes a three. Drop to our satisfaction and the reader sees a four. Below that, the phrase about efforts to perform is understood as a fail. The same scale runs through the individual categories: working method, knowledge, resilience, results, and the conduct clause.
Beyond the grade scale there is a second layer of conventions. The order in which conduct is described, superiors first, then colleagues, then customers, is standard, and a deviation is read as a criticism of the relationship that was moved down. The closing formula is expected to contain regret at the departure, thanks, and good wishes for the future. Leaving any of them out is a signal. Passive constructions and an emphasis on traits irrelevant to the job, such as punctuality for a senior manager, are read as damning with faint praise.
This is the reason AI fits the routine well. The formulae are standardised, extensively documented, and present in the training data of every major model. Given a grade per category, the model produces the correct sentence with high reliability. What it cannot do is decide the grade, and given only free-text keywords it will guess one, fluently, and quite possibly wrongly.
The routine in four steps
Six categories, a grade of one to five each, three bullet points on tasks and projects, and a one-line reason for leaving. Ten minutes for the manager, no prose required.
The prompt template turns the sheet into a complete letter in the standard structure, with the formula that matches each grade. Employee referred to by a placeholder, not by name.
Someone with Zeugnis experience reads the draft against the six checks below and against the personnel file. Inconsistencies go back to the manager, not into the letter.
Name, dates, and letterhead are inserted in the HR system, not in the AI tool. Printed, signed by the manager or HR lead, handed over.
The sheet is the part that takes discipline. Managers who are used to writing a paragraph about a departing colleague have to be asked for grades instead, and the first few will resist because a grade of three feels harsh on paper. It is not harsh, it is honest, and the model turns it into a perfectly benevolent sentence. A sheet with honest grades produces a letter that survives a challenge. A sheet with inflated grades produces a letter that misleads the next employer.
Names and data: what goes into the tool
The reason this deserves its own section is that reference letters are where I most often find shadow AI in HR teams: a colleague under time pressure, a free account, a full name and five years of performance history pasted into a public model. The letter comes out fine. The data protection breach is invisible until someone asks. An AI policy for the HR team names the allowed tools, and the GDPR page explains what a data processing agreement is and why it is the line.
Six checks before signing
- Tasks complete and in the right weightThe task description covers the whole period and the most responsible tasks come first. A missing responsibility is read as a hidden criticism.
- Grade formulae consistentThe individual categories and the summary sentence carry the same grade. A summary at grade two with a working method at grade four is the most common reason for a challenge.
- Conduct clause in standard orderSuperiors, colleagues, customers. Any other order is read as a signal about the relationship moved down.
- Closing formula matches the gradeRegret, thanks, good wishes for a grade of one or two. Thanks and good wishes for a three. Omissions are noticed by every German reader.
- No secret codes, no irrelevant traitsNo praise for sociability in a letter for an analyst, no punctuality for a director. Nothing the employee could read as a coded message.
- Anonymised in the tool, personalised outsideThe draft history in the AI tool contains no name. The final document was completed in the HR system.
Reference letter processes rebuilt in HR Operations roles, including the reduction of manual HR admin from 30 to 5 hours a week at an international software company. Practical knowledge, not legal advice. More about me
Reference letter prompt template with assessment sheet
The six-category assessment sheet for managers, the prompt template that turns it into a complete Arbeitszeugnis in the standard structure, the grade-to-formula table for the review, and the six checks as a one-page list. In German with English annotations for international managers.
Frequently asked questions
Is there really a legal right to a reference letter in Germany?
Yes. Section 109 of the Trade Regulation Act (Gewerbeordnung, GewO) gives every employee the right to a written reference on leaving, and on request a qualified one that assesses performance and conduct. Refusing it, or delivering a poor one, regularly ends in the labour court, and the courts usually order a letter equivalent to a good grade.
Can the AI write the letter from a few keywords?
It can write fluent Zeugnis German from keywords, which is exactly the problem: fluent and wrong. The routine uses a structured assessment sheet from the manager with a grade per category. The model translates the grades into the correct formulae. The manager still has to decide the grades.
Will the employee notice the letter was AI-drafted?
Not if the sheet was filled in honestly and a person with Zeugnis experience reviewed the draft. The formulae are standardised anyway, that is the point of the coded language. What employees and their lawyers notice is inconsistency between the overall grade and the individual sentences, and that is what the review catches.
Do we need a works council agreement for this?
For a text routine where a person enters anonymised grades and reviews the output, in most cases not, because no employee is monitored and no decision is automated. If the tool pulled performance data from your HR system and produced grades itself, that would be a different matter and a case for the AI agreement.
Which tool should we use?
One with an enterprise data processing agreement, because the assessment sheet contains personal data even when the name is left out. Microsoft Copilot, ChatGPT Enterprise, or Claude for Work all work. The German Zeugnis formulae are well represented in all major models. A specialised Zeugnis generator adds little.
Does this apply to interim references and English references?
Interim references (Zwischenzeugnis) follow the same rules and the same routine. An English reference for an international employer is a different document without the coded grading. The prompt set contains a variant for it, but the German original remains the legally relevant one.