A marketing team runs a spring campaign. The creative is warm, optimistic, built around blossom trees – that very English shorthand for renewal, for the particular hopefulness of a season turning.
They localize it into Spanish. The translation is technically flawless, every word accurate. In Spain, it lands completely flat. Blossom trees aren’t a cultural reference point there. The concept simply doesn’t carry the same weight. The campaign was fluent. It just wasn’t culturally intelligent.
This is the failure that enterprise AI is producing at scale right now, and most organizations won’t see it coming. By the time the data arrives, the damage is already done.
SVP for Product at RWS Group.
The problem with looking good
AI-generated content doesn’t fail obviously. It doesn’t produce gibberish or grammatical errors. Scored against conventional quality benchmarks, it often performs as well as human-produced translation. It looks right. It reads well. That surface confidence is what makes the underlying problem so hard to catch.
The failure shows up as a lagging indicator.
A web page that retains visitors in one market but loses them in another. An email campaign whose call to action converts in English but stalls in Japanese, because “contact us” reads as an invitation to an English speaker and something closer to a command to someone from a culture where that directness feels unwelcome.
A product description that performs in one region and quietly underperforms in three others, with no obvious explanation in the copy itself.
By the time those signals surface in your analytics, you’ve already shipped the problem at scale. When AI is accelerating content production faster than any previous technology, scale arrives very quickly.
Fluency is a solved problem. Cultural intelligence is not.
Generic large language models are trained to produce fluent, coherent output across the broadest possible dataset. They are exceptionally good at this – they predict language patterns at a level of sophistication that genuinely rivals human performance on surface metrics.
They were not built to understand that the same message carries different emotional weight in different cultural contexts, or that a concept impactful in one market is meaningless in another.
Cultural intelligence can’t be added to a generic model after the fact. It has to be present from the start. In practice, that means something specific: it’s not about building a different foundational model. Only organizations with hundreds of millions to invest could contemplate that.
It’s about what surrounds the model. The linguistic assets that encode how a brand communicates. The translation memory capturing what has worked and what hasn’t across markets. Style guides, terminology databases, examples of high-performing and underperforming content.
Together, these give a model the context to do something a generic LLM can’t: produce content that transfers meaning, rather than just words.
Two people in a marketing team discovering they can generate French with a general-purpose AI tool is not the same as having a culturally intelligent content operation. The output might look identical. The performance in market will not be.
The human question
None of this argues for pulling back from AI. Quite the opposite. There’s a well-documented economic principle – Jevons paradox – which holds that as the cost of a resource falls, consumption rises. The same dynamic plays out in content.
As AI dramatically reduces the cost of producing localized material, the volume that can reach local markets increases significantly. The opportunity is there.
Volume without quality, though, is a liability. The answer isn’t to remove human expertise from the workflow; it’s to focus it.
In a well-designed AI content operation, human linguists and cultural specialists aren’t doing less – they’re working on the high-stakes pieces where judgment and nuance matter most, rather than being applied uniformly regardless of complexity.
The machine handles volume. Humans handle the work where getting meaning right is hardest.
What breaks this model is removing the human layer and assuming the machine will self-correct. It won’t, unless it’s been built to understand what it’s trying to achieve and to validate whether it has. Swapping a human translator for a machine translator is not the same as replacing a human-led process with an intelligent one.
Measure what the market tells you
Most enterprises currently measure AI’s contribution to content through cost reduction and speed. Both matter. Neither tells the full story.
The more useful question is how assets are performing in each target market. Not how quickly they were produced, or what they cost – but whether they’re driving the recognition, conversions and customer loyalty they were created to drive.
Every piece of content exists to do something. Reducing translation spend is only valuable if that saving unlocks more assets in more markets, and those assets actually perform. That’s the number worth tracking.
Enterprises that start asking that question tend to make different decisions about the tools they use, the role of human expertise in their workflows, and what it actually means to communicate across cultures.
We’ve featured the best AI website builder.
This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.
The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit
https://cdn.mos.cms.futurecdn.net/PAztEScphfxGJfYno5NjrL-2560-80.jpg
Source link
