Making AI replies sound like your brand in Turkish and English.

Many brand guides describe the voice in adjectives: friendly but professional, warm, never pushy. People who have absorbed the brand can work with that. A model can’t. Asked to be “warm”, it produces its own idea of warmth, which tends to mean extra exclamation marks and apologies for things nobody complained about.

Models follow examples better than adjectives. A tone guide for AI replies needs rules the model can apply, paired examples in each language, and checks in the evaluation set that show when replies drift.

Why do “friendly but professional” guides fail with AI?

Because every adjective leaves the decision to the reader. Two agents interpret “friendly” differently and correct each other over time; a model interprets it once, in its own way, and repeats that thousands of times. Brand guides also rarely say what to do in the moments that matter: a refund that is late, a delivery that failed, a request the policy refuses. Those are exactly the replies customers remember. Many brand guides were written for marketing copy, read by people who chose to engage, not for a customer whose parcel is lost.

What should a tone guide for AI contain?

Seven sections, each written as rules with examples. The rules below are illustrative; yours should come from your brand and service teams.

Then add paired replies, good and bad, for the situations that matter most. Take an illustrative pair for a late refund, for a customer whose record holds the title “Hanım”.

• Turkish, bad: “Değerli müşterimiz, yaşadığınız olumsuz deneyim için üzgünüz. Talebiniz ilgili birime iletilmiştir.”
• Turkish, good: “Merhaba Deniz Hanım, iade ettiğiniz ürün depomuza ulaştı ve kontrol ediliyor. Kontrol tamamlanınca ücret iadesi, ödeme yaptığınız karta yapılacak. Tutarın hesabınızda görünme süresi bankanıza göre değişir.”
• English, bad: “We apologize for any inconvenience this may have caused. Your request has been forwarded to the relevant department.”
• English, good: “Hi Deniz, your return has reached our warehouse and is being checked. Once that’s done, the refund goes back to the card you paid with. How quickly it shows depends on your bank.”

The good versions say what happened and what happens next. The bad ones say nothing a customer can use.

Models follow examples better than adjectives.

How do you handle siz and sen in Turkish?

Decide once, write it down, and apply it everywhere: email, chat, WhatsApp and the phone scripts agents read. Where the brand hasn’t decided, “siz” is the safe default. “Sen” from a company can read as over-familiar, especially to older customers or in a complaint, and a brand that uses “sen” in its campaigns should decide separately whether service follows.

Consistency within a conversation matters as much as the choice. Generated text drifts: “Talebinizi aldık, en kısa sürede sana dönüş yapacağız” starts with “siz” and ends with “sen”. Don’t mirror a customer who writes “sen” unless the brand has chosen to. And never guess Hanım or Bey from a first name: names such as Deniz are given to women and men.

How do tone rules differ between Turkish and English?

English marks formality with word choice, contractions and first names; Turkish marks it in grammar, and pairs courtesy titles with first names rather than surnames. Politeness formulas don’t travel well: “Anlayışınız için teşekkür ederiz” is ordinary courtesy in Turkish, while “Thank you for your understanding” can read as presumptuous after a service failure. So write each language’s examples natively instead of translating them. Set length limits in sentences, not characters, because Turkish words run long. Decide too whether the assistant speaks as “we” or “I” in each language.

How do you test tone?

Put tone into the evaluation set as checks, not impressions:

• Automatic checks: banned phrases, sentence length, exclamation marks and emojis, and form of address, which can be detected from Turkish verb and possessive endings.
• Rubric checks: is the apology specific, does bad news follow the pattern, is any promise backed by system data? A model can score these against the rubric, with native reviewers checking a sample in each language.
• Agents’ edit reasons: when agents edit a draft, they pick a reason such as tone, fact, policy or length. A rising share of tone edits shows which rule is failing, before customers notice.

Who owns the guide?

One named owner, usually the customer service lead, with brand or marketing owning the voice itself and legal reviewing wording for sensitive situations. The guide lives next to the prompts, is versioned like them, and every change re-runs the evaluation set. Review it when tone edits rise, when a new channel opens, or when the brand changes its campaign voice. For how the same rules apply on a chat channel, see WhatsApp customer service with AI.

Learn from your best agents

Collect twenty real replies your best agents are proud of, and twenty they would rewrite, in both languages. Turn the difference between the two sets into rules, keep the best pairs as examples, and run the new guide on a sample of past conversations before it reaches customers. Tone guides for suggested replies are part of customer operations work, and the same discipline shapes customer answers and product copy in commerce intelligence.