Blog

A new, better way think about translation quality in 2025

In 2025, localization teams are under pressure to translate more content, in more formats, faster than ever. They’re also under pressure to maintain, and maybe even improve, translation quality.

There lies the rub: Expectations are up. Budgets and bandwidth, not so much.

AI is often sold as the push-button fix to solve all these issues for low cost and with little effort. But the truth is more complicated than that: localization is still an investment, and great outcomes still depend on having the right processes and the right people in the loop.

And like any investment, it helps to define what you want in return. Quality matters, absolutely. But it’s only one component of what a localization program is expected to do—build trust, drive growth, and reflect your brand in every market.

It’s time to think about translation quality with a dual focus:

Meeting the expectations of your users
Maximizing the results you get from your localization program.

We spoke with Jan Dockal, our Language Quality Director, to explore what quality really means in translation and how to build a program that delivers it efficiently. Here’s his take on it.

Why translation quality still matters

It goes without saying that translation quality matters. If people can’t use your product, navigate your website, or understand your support content, they won’t stick around. Poor translations can trigger customer complaints, compliance violations, and lost trust—none of which are cheap to fix.

But as Jan explained, “Not everything needs to sound sexy.” A promotional text message or a one-off social post isn’t held to the same standard as a set of safety instructions —and it shouldn’t be.

That’s why the smartest localization programs don’t aim for “perfect” across the board. Instead, they set clear quality targets based on what the content is, who it’s for, and what’s at stake.

Forget “100% perfect, all the time.” Aim for fit-for-purpose quality.

“Fit-for-purpose” quality means the translation is good enough for the content to do its job, whether that’s informing, selling, supporting, or keeping your customers safe. Once you’ve hit a solid baseline, not everything needs to be flawless to be effective.

Chasing 100% perfection—not just accuracy, but brand alignment, cultural nuance, and emotional tone—requires more time, more hands, and more budget. And often, that level of polish isn’t necessary.

That’s where content tiering comes in. By grouping content based on visibility, risk, and business importance, you can apply the right level of effort where it matters most:

Top tier content

Here, you’ll find high-stakes content: safety instructions, compliance documentation, landing pages, and key marketing material. These require deep expertise, style guide alignment, human review, and often third-party QA.

Translations for markets with especially low tolerance for errors might also land here. Jan explains, “There are some markets which are known for very high quality expectations: Canadian French, Dutch, German, and languages like Japanese and Korean. These audiences have very high standards, want to sign off on content personally and tend to react badly to content that doesn’t meet their standards.”

This is the most expensive tier, because the price and process need to reflect the stakes. Premium outcomes require premium inputs. At this level, projects may not be a good fit for AI translation at all, further driving up costs.

Middle tier content

This might include help articles, onboarding flows, internal documents, or UI strings. This is content that matters to the user but isn’t legally sensitive or deeply branded. Here, AI translation with light human review or targeted QA can be a good fit, especially when context and terminology are well defined.

Bottom tier content

This covers items like user generated content, support content, and backend code comments. These rarely need linguistic polish. In many cases, raw AI translation is enough to get the job done—quickly and affordably.

Jan says, “Wanting premium quality everywhere is often a knee-jerk reaction. The smarter move is deciding where you really need it—and where you don’t. That’s how you meet customer expectations while keeping your budget under control.” Across all tiers, the principle holds: your quality expectations, the requirements of your content and your budget all need to match.

What actually works in 2025

To succeed globally in 2025, you need the right level of quality for everything you translate, at the right cost and speed for your business. A strong quality program brings together the right people, the right tools, and a shared understanding of what “good” means to make that happen.

Here’s how we help our clients build that kind of quality program.

Start with content tiers—and build from there

As we covered earlier, not all content needs the same process. Once you’ve categorized your content into high, medium, and low-stakes tiers, you can match each with the right mix of tools, oversight, and evaluation. This gives your team a scalable way to balance quality, speed, and cost.

Define what “good” looks like

Each tier needs its own definition of success. That starts with two conversations:

What does good quality look like for this tier?
How will we measure it? (Scorecard with the specifics: style, tone, terminology, grammatical errors)

These conversations don’t always happen early—or clearly—enough. Clients might feel disappointed by translations that technically “pass” QA, because the scoring system doesn’t reflect what they value most.

Jan explains, “We sometimes see clients wanting to increase their quality thresholds—like bumping the pass score from 90% to 97%—without changing scope, expectations, or processes. That rarely solves the root problem.”

If what you expect from a quality evaluation and how quality is being calculated are out of sync, you’ll continue being dissatisfied with the results.

Combine objective measurement with human insight

Some translation quality issues are easy to classify (wrong terminology, grammar issue), while others are more subjective (does it flow? does it sound natural? does it match the brand?).

According to Jan, “Quality is always part objective, part subjective. You need a model that accounts for both—error scoring and an overall assessment of how the content actually reads.” That might look like:

Using structured scoring to track measurable issues (quantitative)
Adding an overall quality rating (such as 1–5 stars or a qualitative summary) to reflect the reviewer’s holistic impression (qualitative)
Identifying trends across both dimensions over time, rather than overreacting to any single score

And everyone needs to be aligned and agreed. This is especially important when a new in-market reviewer joins, or a stakeholder’s expectations shift mid-stream.

Engage the right resources

Even the best setup needs the right people in the loop.

That’s why we:

Select specialized linguists and post-editors who understand not just the language, but the subject matter, brand voice, and content goals for sensitive or high-impact content
Match qualified evaluators to the task, based on domain, not just availability.
Recommend Linguistic QA (LQA) or third-party review when the risk or visibility calls for it
Provide tech consulting to help you integrate AI and QA workflows into your existing systems without creating chaos

Build strong processes and maximize supporting assets

High-quality translation starts long before a word is translated. It begins with clear inputs, smart tools, and a repeatable process everyone understands.

Here’s what we focus on:

Smart tools, clean data, and structured prompts: We use high-quality training data to tune engines and build effective prompts for large language models (LLMs). This helps AI output start closer to your target quality—less cleanup, more confidence.
Translation memories and style guides: Well-maintained TMs and concise, example-driven style guides support consistency and accuracy, whether the content is handled by a linguist, an AI model, or both. According to Jan, “A focused guide with five pages and good examples will take you further than a bloated one full of definitions no one can digest or remember.”
Context-rich workflows: Screenshots, product descriptions, and terminology guides aren’t “nice to have”—they’re essential inputs. Shared context keeps translations aligned with product reality and user expectations.
Process clarity and consistency: We help clients onboard reviewers and editors the right way, define quality thresholds per content tier, and track performance trends over time—not just individual errors.

When your inputs, workflows, and people are aligned, quality becomes a strategic advantage.

How to make quality measurable (and repeatable)

You can’t improve what you can’t measure, and that’s where structured quality models come in. The most commonly used are:

MQM (Multidimensional Quality Metrics): A customizable framework that breaks down quality into categories like accuracy, fluency, terminology, locale conventions, and more. MQM lets you track specific types of errors across projects, teams, and languages—making it ideal for large-scale programs with varied content types.
TAUS Dynamic Quality Framework (DQF): Offers real-time feedback, integration with CAT tools, and dynamic tracking of performance across vendors or engines.

Most companies adapt these models to fit their real-world needs—changing the weight of certain error types, adding or removing categories, and setting different pass/fail thresholds based on content type.

How AI is improving quality in 2025 and beyond

Right now, human linguists are still the gold standard. But AI can help them do their jobs better. At the same time, AI translation is improving at almost dizzying speed—and the hype is moving even faster.

As Jan explains, “Clients tend to think that with ChatGPT, they no longer need to invest in localization. But AI isn’t going to replace your language quality program—it just replaces the agent performing the work. You still need to feed it good data, good context, and a clear outcome. The rules haven’t changed.”

At Acclaro, we’re all in on using AI to boost quality, speed, and scalability. But we also believe in separating what’s real from what’s just noise. Here are the advances that are truly improving translation quality in 2025:

Better accuracy and fluency: Advanced neural models are producing smoother, more natural-sounding translations, even in complex sentence structures.
Improved contextual understanding: AI engines are getting better at interpreting idioms, tone, and nuance—especially when paired with well-crafted prompts and domain-specific input.
Greater language coverage: Support is expanding for languages that were previously underserved, helping you maintain consistency and reach across all markets.
AI microservices for specific tasks: Instead of applying AI to every step, we use targeted AI for high-impact tasks—like terminology extraction, TM cleanup, or style rephrasing—where it measurably improves output and efficiency.
Automatic Quality estimation (AQE) with LLMs: Large language models are being tested to identify segments likely to contain errors. This could reduce review time and help reviewers focus on what matters most.

Not on the list: AI is not replacing human linguists completely. The future (at least the near future) is a partnership between AI on one hand, and skilled humans on the other.

If you’re still guessing at quality, we should talk

Quality doesn’t mean perfect—it means purposeful. It’s about building a strategy that fits your content, your markets, and your goals. The right mix of people, process, and tools can help you scale without sacrificing what matters most.

Connect with us to talk about how we can help you drive quality through smart systems, skilled teams, and AI that works for your business.