German Business Writing AI: Why Sie and du Are Not the Hard Part

TL;DR: German business writing AI fails in a predictable place. The model chooses Sie or du correctly most of the time, then sets everything around the pronoun to the average of every German business email it has seen: the greeting, the closing, how many Konjunktiv softeners cushion the ask, whether the request is a verb or a noun. Your reader in Hamburg registers the mismatch before the second line. Telling the model to be "formal but approachable" changes nothing, because that phrase describes thousands of writers and none of their habits. Your German register lives in countable choices. Count them from emails you already sent, hand them over as rules, and every future model executes them more faithfully than the last.

You ask Claude for a follow-up to a supplier in Stuttgart. The draft opens with "Sehr geehrte Frau Dr. Weber," runs three Konjunktiv clauses before the request, and closes with "Mit freundlichen Grüßen". You have written to Frau Weber every week for two years. You open with "Hallo Frau Weber," ask in the first sentence, and sign off "Beste Grüße".

Nothing in the draft is wrong. All of it is somebody else's.

Why Does German Business Writing AI Get Sie and du Wrong?

Mostly, it does not, and that is what makes the failure hard to see. Ask a current model for a German business email and the pronoun will usually be correct, the verb will agree, and Sie will be capitalised. The register around the pronoun is where the draft goes wrong, and the register is what the reader is actually reading.

Sie and du are two words with a whole system attached. Duden's guidance on capitalising du and ihr in letters is a small illustration: Sie is always capitalised, du may be capitalised in letters and emails as a courtesy, a lowercase du signals more closeness, and mixing both spellings in one message is an error. Three writers who all correctly use du with a colleague will make three different capitalisation choices, and each choice says something about the relationship. The model makes one choice for all of them.

The same holds downstream of the pronoun: which greeting opens the email, whether the ask arrives as "Bitte schicken Sie mir" or "Ich würde mich freuen, wenn Sie mir schicken könnten", and which of a dozen closings signs it. Each is a habit. A model trained on all German business writing has the average habit, which belongs to nobody.

Sie and du Are the Visible Layer of a Deeper Register System

German business register moves on six parts, and each writer sets them separately for each reader. Address is one. The other five do most of the work.

Every row is a register decision, and every default in the middle column is the statistical centre of German business correspondence. Spell-checkers pass all of it. The reader who has known you for two years does not. Even the closing carries a signature: the set of accepted formulas is small and public, so a writer who signs "Viele Grüße" to clients and "VG" to the team has a habit, and a model that signs "Mit freundlichen Grüßen" for both has erased it.

Three Emails, One Writer, Three Registers

Take a project manager in Osaka who runs a supplier relationship in Munich and writes German every day. Three emails before lunch.

To the supplier's account manager, Herr Braun, after three years of weekly contact: "Hallo Herr Braun," one sentence of context, the ask in sentence two as "Könnten Sie mir die aktualisierte Stückliste bis Donnerstag schicken?", then "Danke!" and "Beste Grüße". Sie throughout, with surnames; first-name terms were never offered and never needed.

To a director at a prospective partner she has met once: "Sehr geehrter Herr Dr. Vogel," one line referring to the meeting, the ask wrapped as "Ich würde mich freuen, wenn wir einen Termin für ein erstes Gespräch finden könnten", and "Mit freundlichen Grüßen" above her full title block.

To her two German-speaking teammates, in the shared channel: "Hi zusammen, kurze Frage:" and the ask, no closing at all. Lowercase du, always.

Same writer. Three registers. All three are "formal but approachable" in a way no adjective can distinguish, and all three are correct. Feed the brief for the first email to a model with a tone description and it produces the second. We described the general version of this shift in why your writing voice changes when you switch languages; German makes it sharper because the pronoun system leaves no neutral ground to hide in.

For a fast read on how far your own drafts sit from your real writing, the Humanity Score test compares a piece of AI output against human text in about a minute.

Six Places the German Draft Slips

Across German drafts from ChatGPT, Claude and Gemini, the same six slips recur, and none of them is a grammar error.

1. The pronoun drifts. The email opens "Hallo Anna," on du terms and shifts to "Könnten Sie" by paragraph two, usually because the prompt was in English and "you" carries no register. A German reader experiences the drift as a sudden formalisation.
2. The greeting is one notch too formal. "Sehr geehrte Frau Weber" to a weekly contact: correct for a stranger, cold for a colleague.
3. Softeners stack. Konjunktiv II plus "eventuell" plus "kurz" plus "gerne" in one ask. Each is polite. All four together read as a writer who is afraid to ask.
4. The ask goes nominal. "Wir bitten um Rückmeldung bis Freitag" from a person who would write "Sag mir bitte bis Freitag Bescheid".
5. The closing is interchangeable. "Mit freundlichen Grüßen" on every email, including the one to the teammate who has received "LG" from you for four years.
6. The rhythm flattens. Every sentence twelve to twenty words, a subordinate clause every second one, no "Kurze Frage:" line standing alone. This is the part of voice AI flattens most.

The first slip is the one people notice. The other five are the ones they feel.

"Formal but Approachable" Hands the Model Nobody's German

Adjectives point at a register the model already knows. Type "formell, aber nahbar" or "professional, warm, appropriate for German business", and the model retrieves its average of everything filed under those words. Your colleague who types the same phrase receives the same email. You two write nothing alike.

AI does not know your tone. It knows words that describe tone.

This is the failure we took apart for English in why tone adjectives are not your voice, and German makes it costlier. The register vocabulary is richer, so "formal" covers more ground and constrains less: it spans a Sie email with "Hallo" and "Beste Grüße" as well as a letter with "Sehr geehrte Damen und Herren" and a title block. And German business readers read register as information about the relationship, so a wrong notch is a message you did not mean to send.

The common workarounds leave the gap open:

• Uploading past emails as knowledge files. The model retrieves them when it decides they are relevant and paraphrases the content. It does not absorb the habits.
• Writing in English and asking for German. Produces English directness in German grammar: the ask lands in sentence one with no softener, and the pronoun drifts because "you" never told the model which one. The reverse problem, AI English that sounds translated, is the same failure with the languages swapped.
• Switching to a newer model. This one deserves its own section, because it is the one people bet on.

A Newer Model Executes the Average More Fluently

Model capability and user-data conditioning are separate layers. The lab improves the first with every release. It cannot supply the second, because the second is built from emails you sent and the lab has never seen them.

What a stronger model delivers is fidelity to instructions. Give it "formal but approachable" and it renders that instruction more thoroughly than the last model did: cleaner Konjunktiv, more consistent Sie, a tidier closing. It has moved closer to the average, not closer to you.

The same property works for you the moment the instruction carries something precise. A block that says "Sie with Herr Braun, greeting Hallo, ask in sentence two as a direct Könnten Sie question, one softener at most, close Beste Grüße, never Mit freundlichen Grüßen" is a set of rules a stronger model hits more accurately than a weaker one. The better the model gets at following nuanced instructions, the better it executes your profile. We laid out the general argument in why better AI models still don't know your voice. German is a case where it bites hard, because the distance between "formal" and your formal is measured in choices a reader cannot miss.

So the model dropdown is the last thing to change. Change what the model is asked to execute.

Measure Your German Register From Emails You Already Sent

Your register is countable. Each of the six elements above can be read off ten to fifteen sent emails in under an hour, and each count becomes an instruction with no adjective in it.

Separate first. German in one folder, every other language in another. Across the 320 samples in four languages we measured, AI output shifts markedly by language on formality, conciseness and expressiveness, so a profile pooled across languages inherits the wrong baseline for each. German is not yet in that published set, and we will not quote a German number we have not measured; the mechanism is the same. Within the German folder, label each email by relationship: long-standing Sie contact, first contact, du colleague, supplier, complaint. A profile averaged across your client mail and your team chat is a profile of nobody.

Then count, per register:

• Address. Sie or du, and for du, capitalised or not. Note what you never do.
• Greeting. Every opener you used, and how often. Write down the ones you never use.
• Softeners. Konjunktiv II asks versus direct asks. "Kurz", "mal", "eventuell", "gerne" per ask. Most writers have one favourite and a ceiling; write the ceiling as a hard limit.
• Ask construction. Verb or noun, and which sentence the ask sits in.
• Closing. The exact string per register, including abbreviations (VG, LG, MfG).
• Rhythm. Words per sentence, typical and range. Subordinate clauses per paragraph.

A finished block for one register:

Register: existing supplier contact, Sie, weekly. Greeting "Hallo Herr Braun," never "Sehr geehrter". One sentence of context, then the ask in sentence two as a direct "Könnten Sie ... ?" question. One softener maximum, "bitte" inside the question, never "eventuell" or "kurz". Verbs, not nouns: never "Wir bitten um". Sentences 8 to 18 words, no more than one subordinate clause each. Close with "Danke und beste Grüße". Never "Mit freundlichen Grüßen", never a title block.

One block per register, and a line telling the model which one applies. That is the whole method.

Deploying the Register Block in ChatGPT, Claude and Gemini

The block belongs where instructions persist, never in the chat window. In ChatGPT, that is custom instructions or the instructions field of a Project or a Custom GPT. In Claude, it is the custom instructions of a Project. In Gemini, it is the instructions of a Gem. Our guide to making ChatGPT sound like you covers where each block goes and how to stop it drifting over a long conversation. Keep one block per language, clearly labelled, or the model starts softening your English asks with the German ceiling.

This is what a Style Profile from MyWritingTwin.com does, per language. German samples get their own Writing DNA analysis, separate from English or Japanese samples, across more dimensions than the six above, and the output is a block of constraints rather than a paragraph of adjectives. Your Writing Twin for ChatGPT, Claude and Gemini is that block deployed inside the tools you already use. Neither profile is a translation of the other, because your two voices are not translations of each other. The keigo version of this problem shows how differently the same method plays out in Japanese.

How Do You Test Whether the German Draft Is Yours?

Give the model the brief for a German email you already sent, put its draft beside yours, and count. Five minutes, no tools, and the answer is a set of numbers rather than a feeling.

Pick one email per register from the last month. Write a two-line brief for each: who it went to, what you needed, what they already knew. Ask for the email in German and compare:

• Address: same pronoun throughout, same capitalisation of du?
• Greeting: yours, or "Sehr geehrte"?
• Ask position and construction: which sentence, verb or noun?
• Softeners: how many, and which?
• Closing: yours, or "Mit freundlichen Grüßen"?
• Words per sentence: typical, shortest, longest

If the counts differ, the model does not know your register, whatever it says when you ask it directly. Run the same test after replacing the adjective with the measured block. The counts converge on the first draft and keep converging as you tighten single rules. Sie and du were never the hard part. The register was, and the register was never handed over.

Frequently Asked Questions

Why does my AI-written German email sound off when the grammar is correct?

Because address is a grammar question and register is a judgment question. The model gets Sie or du right and then sets greeting, softeners, ask construction and closing to the average of all German business text it has seen. A German reader hears the mismatch in the first line and reads it as information about the relationship.

Can I fix it by telling ChatGPT to be "formal but approachable"?

No. The phrase covers a Sie email that opens with Hallo as well as a letter that opens Sehr geehrte Damen und Herren, and the model chooses one for you. Adjectives describe an effect and carry none of the mechanics. Your register lives in the greeting, the softener count, verb versus noun for the ask, and the exact closing string. Give those as rules.

Will a newer model handle Sie and du better?

It will follow instructions about register more faithfully, which is not the same as knowing your register. The lab improves capability with each release and cannot supply your data, because that is built from emails you sent. A stronger model executes a vague adjective more fluently and a precise block more accurately.

Does my English Style Profile carry over to German?

No. Your German voice is a separate set of habits calibrated to different norms, including a pronoun system English lacks. MyWritingTwin builds a separate Writing DNA analysis per language from separate samples. An English profile applied to German produces English directness in German grammar.

How many German emails do I need?

Ten to fifteen sent emails across the registers you actually use is a practical floor for a manual block. Keep them separate from your English and Japanese samples and label which relationship each was written for, so the patterns read as habits rather than as one Tuesday afternoon.

See how far your AI drafts sit from your real German with the Humanity Score test, or upload your German samples and let a Style Profile do the counting.