AI localisation without losing the brand's voice

AI can translate a campaign into several languages before a human reviewer has opened the original brief.

The result may be grammatically sound, tonally polite and strategically unrelated to why the campaign worked.

Localisation is not the reproduction of sentences. It is the preservation of meaning and brand character inside a different cultural and decision context.

AI can accelerate that work. It cannot decide alone what must remain stable and what should change.

The practical answer is to work in layers. Lock the product truth and central idea. Adapt the rhythm, idiom, proof and action. Then give a local reviewer authority over the meaning.

“Translate this in our tone” sounds efficient. It is also how a confident brand can become a very polite stranger.

Define voice as choices

“Confident, warm and modern” is difficult for a person to apply and extremely easy for a model to imitate generically.

Describe voice through decisions:

• We make claims precise rather than dramatic.
• We explain complexity without performing expertise.
• Humour observes business behaviour; it does not mock the audience.
• We use short assertions followed by useful context.
• We admit uncertainty where evidence ends.
• We avoid motivational language and inflated futurism.

Then add examples and anti-examples. Voice becomes more portable when it explains what the brand chooses under pressure.

Build the source pack

Give the system:

• Business and product truth
• Audience and market context
• Central campaign idea
• Approved terminology
• Claims and evidence
• Voice principles
• Positive examples
• Anti-references
• Required disclosures
• Output format

Mark which source phrases are fixed, which are adaptable and which require transcreation rather than direct translation.

A useful AI request makes those boundaries visible:

LOCALISE FROM: [source language]  TO: [market, language and audience]
ASSET IN CONTEXT: [ad, page, subtitle, interface or script, plus what surrounds it]
PRESERVE: [product truth, central idea, fixed terms and required disclosures]
ADAPT: [rhythm, idiom, proof, formality and call to action]
TRANSCREATE IF NEEDED: [humour, wordplay or cultural reference]
DO NOT: [prohibited claims, stereotypes, awkward literal phrases and anti-references]
RETURN: [three alternatives with a short explanation of material choices]
FLAG: Any claim, phrase or cultural choice that requires factual or local review.

The model proposes options. The factual owner and local editor remain responsible for deciding which option is usable.

Separate four tasks

Translation

Preserve literal meaning for factual, legal or functional content where precision dominates style.

Adaptation

Change syntax, idiom and emphasis so the message reads naturally.

Transcreation

Rebuild expression around the same strategic effect when humour, wordplay or cultural reference cannot travel directly.

Market localisation

Adjust proof, offer, channel, CTA and decision context to the local business reality.

Do not ask one prompt to perform all four without showing which standard applies to each element.

Use a localisation matrix

Generate alternatives, not one “correct” answer

Ask for several approaches with reasoning:

• Closest to source
• Most natural conversational version
• More formal or category-appropriate version
• Transcreated version preserving the strategic effect

A local reviewer can then compare choices rather than repair a single output word by word.

Protect terminology

Maintain an approved glossary with:

• Product and feature names
• Category terminology
• Words that remain in English
• Transliteration rules
• Prohibited or misleading equivalents
• Pronunciation notes
• Legal or regulatory wording

Update it when real customer language reveals a better choice. A glossary is a learning system, not a monument.

Review in context

Do not approve strings in a spreadsheet alone. Place language in the actual ad, interface, subtitle or landing page.

Spreadsheets are excellent at holding words. They are less gifted at showing that a button no longer fits or a joke now takes nine seconds to explain.

Check:

• Line length and crop
• Reading time
• Speech rhythm and pronunciation
• Visual-text relationship
• CTA clarity
• Formality across the journey
• Whether support can continue in the same language

Brand voice breaks when a friendly local advertisement leads to a stiff generic form and an unprepared support team.

Give local reviewers authority

Local review should cover more than grammar:

• Is the tension true here?
• Does this sound like a person or translated brand copy?
• Is the humour socially appropriate?
• Does the proof matter?
• Is the action realistic?
• Which phrase changes the brand’s character?

Record material changes so the central team learns instead of treating every market edit as an exception.

Measure meaning after launch

Compare by market:

• Message comprehension
• Search and query language
• Comment themes
• Proof engagement
• Conversion and lead quality
• Repeated support questions
• Brand-description language

A high-performing local version may reveal an insight worth returning to the global brand.

Where AI helps most

AI is valuable for:

• First-pass alternatives
• Terminology consistency
• Long-form adaptation
• Subtitle and format variants
• Comparing tone across versions
• Flagging inconsistent claims
• Maintaining structured language assets

Humans remain essential for consequential meaning, cultural context, humour, evidence and final approval.

For campaign-level India adaptation, read Localising an international campaign for India . For production governance, use human approval gates .

Scaling a brand across languages? We can build the context and review system before generating the variants .