Translating content with AI: the quality workflow we use

Translating a website or a catalog with AI today costs a fraction of what it used to cost to hand everything to a translation agency. The problem is that the raw output swings between good and embarrassing, and readers notice precisely on the pages that matter most. To translate content with AI while keeping publishable quality you need a workflow, not a copy-paste: in this article we describe the one we use on multilingual projects.

What AI translates well and where it fails

Modern machine translation handles informative, structured texts well: product descriptions, blog articles, help pages, transactional emails. The sentences are correct and the meaning comes through. The weak points are elsewhere:

• tone of voice: translated text drifts toward a neutral register, and if your brand has a recognizable voice it gets lost along the way;
• industry terminology: without instructions, the same technical term gets rendered three different ways on three pages;
• wordplay and cultural references: slogans, headlines and jokes need rewriting, not translating;
• sensitive content: legal texts, terms of sale and claims about health or safety require precision that must be checked by a competent person.

Hence the rule we follow: AI does the first draft, a person decides what gets published.

The workflow in five steps

On the multilingual projects we manage, the process is this:

1. Glossary preparation. Before translating a single line, we lock down the terms that don't get touched: brand name, product names, technical terms with their official translation, and the words to avoid. The glossary goes to the AI in the instructions and gets handed to the reviewer.
2. Tone instructions. With the translation request we attach explicit guidance: formal or informal register, how to address the reader (in languages that distinguish formal from informal address, the choice changes everything), length and style of headings.
3. Translation in blocks. We translate by homogeneous sections, not the whole site in one go: product pages together, institutional pages together. Similar context keeps terminology consistent.
4. Proportionate human review. Not every page deserves the same investment. Home, sales pages and legal texts go through a native speaker; low-visibility informational pages get a lighter review. Distributing the review budget well is the part of the workflow that most affects the quality-to-cost ratio.
5. Check in context. The final verification happens on the site, not in the document: text overflowing buttons, headings that wrap badly, dates and currencies in the wrong format. A correct sentence in the wrong place is still an error.

The native speaker is still needed, but in a different role

The AI workflow doesn't eliminate the translator: it changes their job. Instead of translating from scratch, the native speaker reviews, corrects the tone and adapts what the machine rendered flatly. On the cost side this shifts the investment from quantity to quality: you pay less for volume and concentrate the expertise where the text sells or creates legal obligations.

There are cases where we recommend starting directly with a native speaker: taglines and campaign materials, content for culturally distant markets, regulated industries. There, translation is really rewriting, and having AI do it only to redo it costs more than doing it well once.

The technical side: where the translations live

A quality workflow also rests on infrastructure. Translations need to live in one place, with a structure that lets you update the Italian version and know which pages in the other languages have fallen behind. On the site you then need the basics of multilingual SEO: separate URLs per language, correct hreflang tags, a sitemap that includes all versions. A well-built site, like the ones we deliver in our websites and eCommerce service, handles languages at the structural level, not with a plugin bolted on top: it makes a difference both for Google and for whoever has to maintain the content over the years.

A practical tip for anyone starting now: choose your languages based on data, not ambition. Look at which countries your traffic comes from and where you already ship: two languages maintained well are worth more than five left half-done.

Need a multilingual site done right?

We design multilingual websites and eCommerce with the right technical structure and a translation workflow that combines AI, glossaries and human review. If you want to take your site to new markets with texts you won't have to be ashamed of, book a free call: we'll look at your content together and propose the workflow that fits your budget.