How I localized AI-generated emails for international markets without losing the human touch

Earlier this year, I was handed an AI-generated content project with a deceptively simple goal: adapt email messages for international audiences.

This wasn’t my first time navigating global nuance. With an MBA in International Business and experience working on a global consulting project in Portugal, I’d already seen how messages land differently depending on culture, tone, and language. But this was my first time applying that lens to AI content generation in my MarTech AI role at HubSpot — and it was more complex than expected.

We already had an AI-generated email prompt that worked well in English—conversational, friendly, and context-aware. The challenge? Making it work in Spanish and French without sounding robotic, clumsy, or culturally off-base.

Sounds easy. It wasn’t.

The Hidden Complexity of "Just Localizing"

What we were really doing was asking an AI model — trained predominantly in English — to speak other languages as naturally as a native marketer would.

Our first attempts fell flat.

Example (original AI output in Spanish):

Here’s what we aimed for in English:

“I saw you were scoping around the platform and that you were interested in speaking with us. Would you like to meet on one of the following days?”

This is the original output in Spanish:

“Estuve revisando tus interacciones en nuestra plataforma y quería ofrecerme como tu punto de contacto.”

In English, it translates to:

“I reviewed your activity and wanted to become your point of contact.”

While grammatically correct, this sounded invasive in Spanish — like we were watching the user too closely. It didn’t feel natural. One reviewer called it “creepy.”

Here’s another example:

Again, it’s technically accurate, but it’s redundant and robotic. It’s the kind of phrasing that makes a reader stop and go, “Did a bot write this?”

The takeaway: Even when the translation is accurate, the tone can be off. And tone is everything in marketing.

The Shift from Translation to Language-aware Prompt Design

At this point, I realized we needed more than AI outputs — we needed a system for guiding the AI to think like a multilingual marketer.

I built a language-portable prompt framework — a structured prompt that could adapt across languages while respecting each one's unique grammar, tone, and cultural context.

Here’s What Changed

Instead of one static prompt, I broke the logic into variables:

We also added clear, language-specific rules.

Example (Spanish):

Example (French):

Why This Shift Mattered

In English, a friendly CTA might look like:

“Would you be available for a brief conversation on one of the following days?”

We tried directly translating it into Spanish:

“¿Quieres agendar 15 minutos para hablar sobre lo que estás buscando?”

It was grammatically correct, but it sounded too casual and unprofessional in a B2B context. Not pushy, just slightly off-tone.

So, we reworded it to be friendly but formal:

“Si te parece bien, podemos agendar una conversación breve esta semana.”

This translates to:

“If it works for you, we can schedule a short chat this week.”

Here’s another example in French:

The second version adds value to the CTA. Not just time — but purpose.

Backing It Up With a Stakeholder Questionnaire

Localization isn’t just a linguistic issue — it’s a business alignment issue.

To get it right, I created a simple stakeholder intake doc and shared it with marketing ops, regional marketers, and content leads. The goal was to align early on tone, content boundaries, and regional sensitivities.

These are some of the questions I asked:

We got some pretty interesting insights.

For example, in some regions, stakeholders preferred not to reference the recipient’s company type in the copy, even though that was common in English (e.g., “I saw that you help startups with HR”).

The localized alternative became more general:

“Entiendo que están buscando formas de mejorar sus procesos internos.” (“I understand you’re looking to improve internal processes.”)

The results of this survey helped create clarity between content, ops, and regional marketing teams — and dramatically reduced our revision cycles.

The Final Product: Human-sounding Emails at Scale

With the updated prompt and intake framework, the new outputs were instantly better.

Before:

After:

And stakeholders responded positively:

Even better, we didn’t need to write separate prompts for every campaign. The same core framework now powers AI-generated messages in multiple languages — with consistent quality.

Takeaways for Marketers

Whether you’re working on AI copy, global ads, or multilingual content, here’s what I learned:

1. Don’t just translate — localize for intent.

Literal translations will get you “technically correct” content. But only localization will make it land.

2. Use prompts like creative briefs.

Include tone, formality, CTA style, gender neutrality, and other language rules as variables. Don’t leave nuance to chance.

3. Build language-aware templates.

Languages behave differently. Plan for things like verb conjugations, pluralization, and sentence rhythm upfront.

4. Get feedback early.

Use a stakeholder intake doc before generation, not after. You’ll avoid rework and misalignment later on.

5. Aim for a real, human tone.

If your AI output doesn’t feel like something you would write to a customer, it won’t convert. Read it aloud. Would you hit send?

AI localization is a marketing skill now.

This project taught me something that has stuck with me since: The future of global marketing isn’t just about scaling content — it’s about scaling context.

The companies that succeed with AI won’t be the ones who generate the most content. They’ll be the ones who generate the most resonant content because they know how to prompt for it. And that starts with understanding the languages your customers speak — in more ways than one.