Ditto

In financial services, a translation that's almost right is just wrong. The term either matches what that locale expects or it doesn't. A localization leader at a Fortune 50 company raised this with me at a content design meetup we hosted recently. Her team has spent years getting that precision right. As they start moving toward agentic workflows, she wanted to know how they hold onto it. What I showed her was locale-based style guides in Ditto. When text gets generated or edited for a locale, it carries that specific locale's guidance with it, instead of falling back on a model's default sense of what a correct translation sounds like. My sense is this is the part of agentic workflows that gets underplayed. Most of the attention goes to generation, to the speed of getting words on the screen. But localization isn't a step you bolt on at the end. It needs its own infrastructure, its own rules, its own way of being governed across every surface a product touches. Most teams I talk to are bracing for the same thing she was, the sense that automation and precision pull in opposite directions. I don't think they do. Automation doesn't lower the bar for content. It raises the bar for the infrastructure underneath it.

A solo content designer and a team of product designers logged over 1 million usage events in Ditto in May. They’re responsible for product copy across an entire fintech portfolio: multiple brands, 10+ markets, 20+ languages, 100M+ users. Every release used to mean manually tracing copy across Figma screens, spreadsheets, and translation files -- locale by locale. Translators got strings, not screens. And when a longer translation broke the layout, the team often didn’t find out until QA. Now, the manual parts are automated in Ditto -- when designs are updated, strings are extracted, deduped, and matched against their library of approved copy and past translations. New strings automatically get keys, then AI-translated using design context and brand- and locale-specific glossaries and guidelines. Sensitive content goes to human review -- and then it all ships. That spike in May is when they flipped it on. This is the kind of content infrastructure I get excited about -- something that enables a small team to support a huge product surface without lowering the bar for quality.

AI is unnervingly good at making bad output look finished. When the copy comes back polished and reads comprehensive, the polish becomes its own kind of noise — and makes it hard to evaluate on the criteria that actually matter. Is this on-brand? Compliant? Does it describe the product accurately? My sense is teams start answering those by how the output feels rather than whether it's measurably correct. We've watched it happen with teams building tooling for themselves: scanners that check copy against a style guide, search across strings, flag inconsistencies. It works beautifully for a while. Then someone adds one rule, it quietly stops catching something it used to, and because the result still looks confident, nobody notices for weeks. Accuracy degrades as the context grows. The polish never does. A system that's supposed to learn over time needs hard, precise rules to learn against. Without them, a good-enough first-pass from a markdown file is probably the ceiling. So the question on every project right now isn't "can we make it agentic?" It's: if this drifts at scale, would I know? Could I catch it before it ships? That answer depends on the boring infrastructure: rules, checks, benchmarks. It's the difference we keep seeing between systems that get better over time and ones that just produce more.

We've spent a lot of time talking about the importance of prioritizing localization in your product workflows. But even if you agree that it's important, the actual process of building a localization workflow can still be... daunting. We get a lot of questions about the realities of managing localization workflows, like... ❓ My team manages multiple products and languages, and the volume of copy is only increasing with AI iteration. How do I manage? ❓ How can I spot localization mistakes before they make their way into production? ❓ We have a hard enough time just localizing the words themselves, how do we also localize the reasoning and nuance behind them? We've written a (newly updated!) all-in-one guide to building a localization workflow in Ditto, to make this a bit easier for you. We cover... ⚪ How to centralize your text and translations into Ditto ⚪ How to choose which languages to support ⚪ How to get your text translated in seconds ⚪ How to enforce quality with locale-specific style guides ⚪ How to integrate those localized strings into development Check it out, on the Ditto blog 👇 https://lnkd.in/eqtEQeNG

We spent a long time asking ourselves what it would mean to localize *as you build* — rather than as a step you take at the end. The answer ended up being less about translation and more about plumbing. My sense is that the "AI translation" framing, while compelling, undersells the problem. A model can translate a string in seconds. What it can't do is make localization a coherent part of how a product gets built: where designers see localized text in context, engineers pull formatted locale-specific string files, and agents generating copy are working from the same underlying structure. The hard part isn't visible. It's whether there's a system underneath managing the strings at all — or whether they're just hardcoded into the product, one line at a time. So that’s what we built. Every string in Ditto can now carry a locale. The work was making that true everywhere product text lives: localization as a first-class citizen across the system. That's not a translation problem. It's an infrastructure one.

Most AI translation tools solve one step of the problem: generating a translated string. That part is basically table stakes now. What's harder is everything around it. Like making sure your French translations actually follow your brand's terminology. Or that your Japanese copy adheres to the formality conventions your team agreed on. Or that any of it connects back to the strings your engineers are actually shipping. And even when the translation is good, most teams are still generating it in one place and managing their strings in another — which means localization is always a separate step bolted onto the side of the actual workflow. The progression that matters isn't "manual → AI." It's: → Can the AI follow locale-specific guidance? → And does it live where your strings already live — not in a separate tool or step?

We don't need to just draft translations, we need to make sure they're on-brand, accurate, sound like a human. Customize your Ditto style guides to a locale, and enforce it with AI. Learn how 👇 https://lnkd.in/gtQJ3VgP

You have an AI translation sidekick waiting for you (for free) in Ditto. Want to learn more about it? 👇 https://lnkd.in/gfiwTnUV

We can all agree that localization is critical — every user, no matter where they’re logging in from, deserves the best possible experience. But for a lot of us, localization is an additional responsibility, or a last minute handoff to get a project over the line. That's why we've gone all in on bringing localization workflows into Ditto. Because when we say you need a system for your product copy, we mean for ALL your product copy. Learn more 👉 l10n.dittowords.com

The Ditto blog has been busy: So many new resources, all focused on helping you improve your localization workflows 🌎 Give these a read, or bookmark them for later... A newly updated, all-in-one guide to localizing with Ditto 👉 https://lnkd.in/eqtEQeNG A feature deep-dive on everything we've built to bring streamlined localization workflows to your Ditto workspace 👉 https://lnkd.in/eyQPbrdg The "why" behind this localization launch, and where we're going next 👉 https://lnkd.in/es4QytEV Once you work through these, don't worry – fresh new content coming your way tomorrow.