Automated translation workflows that know a tagline from a contract

A tagline, a legal clause and a product description go wrong in different ways, so they shouldn’t share a workflow. Here’s how to decide what to automate for each content type, how to check it, and where a person still needs to be involved.

My team looks after Phrase’s own localization, and we don’t use one workflow for it. Help Center articles are created and updated often and get strong leverage from our translation memory and termbase, so they go through automated translation with light human review or none. The blog gets the same treatment for a different reason: terminology isn’t a concern there, and speed and cost come first.

Product UI and marketing copy are different. We post-edit every word, because it has to be accurate and fit for purpose, and it still runs through automated pipelines that take routing, assignment and handoffs off our project managers’ desks.

The content decides how much review it needs.
Automation can take care of almost everything else.

Phrase’s research points the same way. A Phrase study of machine translation quality ended with a caveat: the share of output good enough to skip editing, and so the savings, will vary across domains, content types and language pairs.

This article turns that caveat into a method. Machine translation quality varies by content type, and so does the cost of getting it wrong, which is why automated translation workflows work best when each content type gets its own configuration. Below you’ll find a decision matrix, three worked examples (a product description, a legal clause and a tagline), the settings behind them in Phrase, and the mistakes that cost teams most.

A decision matrix for automated translation workflows

Every row answers two questions. How far can you trust the machine, and the checks that score it, on this kind of content? And what does it cost if something wrong gets through?

The first three rows get less attention than contracts and campaigns, but they’re where most of the work is. In my experience, support content and UI strings alone make up 60 to 80% of the workload for most localization teams, so small improvements to those workflows add up quickly.

Treat the matrix as a starting point. Your languages, engines and appetite for risk will move the lines.

When trust and cost point in different directions

Trust varies by content. Engines don’t perform evenly: Phrase’s documentation for MT Autoselect notes that an engine that does well on product descriptions might struggle with a medical report, which is why Autoselect weighs content domain as well as language pair. Automated quality checks vary too. Alon Lavie, Phrase’s strategic advisor on AI, has argued that quality estimation is least dependable for fine-grained, context-dependent judgments, especially in creative content.

Cost varies just as much. A clumsy product description might cost you a sale. A mistranslated termination clause can cost you a lawsuit. A tagline that lands badly in Japan costs something harder to measure and slower to repair.

The trouble starts when the two answers diverge, and legal text is the clearest case. Contracts are structured, formulaic and repetitive, exactly the kind of source text machine translation handles well, which is why you’ll find industry guides describing legal documents as ideal for it. On quality, they have a point. On cost, the picture changes: a high quality score estimates how good a translation probably is, and it says nothing about what happens if it isn’t.

Medical and other regulated content sits in a harder corner still, where the machine is less reliable and a miss costs the most. Technical documentation behaves more like product copy: structured enough for the machine to handle well.

The rest of this article follows three sentences through their workflows, from the most automatable to the least.

Product descriptions: let the checks do the sorting

“Water-resistant to 50 m. The rechargeable battery lasts up to 14 days on a single charge.”

This is the content automated translation was built for. It’s structured, it repeats across a catalog, and most of what matters can be checked: units, numbers, product names. It’s also where the volume is, so it’s where the savings are.

A workflow that suits it:

1. Pick the engine per language pair. Use the engine you’ve tested, or let MT Autoselect choose.
2. Feed it context. Translation memory keeps the catalog consistent, and a termbase protects product names and specifications.
3. Add AI post-editing where it earns its place. MT Optimize, a generative AI step in Phrase Orchestrator, refines fluency, grammar, terminology and formality. Test it on your own catalog before switching it on everywhere.
4. Check every segment automatically. A quality score such as Phrase’s QPS estimates how a professional reviewer would rate each segment, and QA checks catch problems such as missing tags and unapproved terms. Segments that clear both are confirmed; the rest go to a post-editor.
5. Sample what got through. Spot-check a share of confirmed segments every cycle. QA can run a first review of the sample and flag issues for a person to judge, so you find drift before your customers do.

Automatable isn’t the same as risk-free. A wrong battery life or a garbled safety note leads to returns, support tickets and, in some categories, regulatory questions. Keep the checks and the sampling even when results look good, and watch two things in particular: new product lines, which bring vocabulary your translation memory hasn’t seen, and the source feed itself, because errors in the source come straight through.

Legal clauses: high scores, mandatory sign-off

“Either party may terminate this Agreement upon thirty (30) days’ written notice to the other party.”

For a machine, this sentence is close to ideal. It’s formulaic, some version of it is probably already in your translation memory, and an engine will likely render it fluently. It’s also a sentence where “may” turning into “must,” or 30 days becoming 3, changes what your company has agreed to.

Fuzzy matches show how small the difference can be. As Martin Švestka, Phrase’s Director of Product for TMS, has pointed out, a stored translation of “the jury found the defendant guilty” will be a very high fuzzy match for “the jury found the defendant not guilty,” and nobody wants that one confirmed automatically.

So the legal workflow uses automation for consistency and routing, and leaves the decision with an expert:

1. Specialist legal linguists own the translation and the review. Regulated content needs people who know the legal systems on both sides of the language pair.
2. Translation memory and termbases are enforced strictly. A termbase doesn’t lock text, but QA can be configured so a job can’t be completed or passed to review until the approved terms are used.
3. MT, if you use it, is a draft for the linguist. Choose the engine per language pair and treat its output as a starting point.
4. Scores tell reviewers where to look first. They never decide whether review happens.
5. Sign-off is mandatory, and traceable. Phrase records which vendor and linguist were assigned to each job, with timestamps (linguists working for a vendor are anonymized), which gives you a trail for your own audits and your regulator’s.

Where a market needs an extra pair of eyes, Phrase Orchestrator can add review steps and conditional review loops for that market alone. Medical and other regulated content follows the same pattern, with more of the work falling to the specialist, because the machine’s draft is less reliable to begin with.

Automation still does valuable work here, and most of it is unglamorous: the right term every time, the right linguist on every job, and a record of who did what and when. The judgment stays with a person whose name is on it.

Taglines: where the score has least to say

“Made to be worn out.”

Read this imaginary tagline once and you get both meanings: the product lasts, and you’ll wear it all the time. A literal translation keeps one meaning at best. It will also read fluently, which is exactly what makes it risky. We’ve written about why fluent output is so hard to judge.

Here, both questions point the same way. The machine has the least to offer, the score has the least to say, and a miss is public. QPS is built to predict how a reviewer would rate a translation’s accuracy and fluency, so a transcreated line that departs from the source on purpose can score poorly and be exactly right, while a literal one can score well and fall flat.

Translation carries meaning from one language to another. Transcreation carries intent, and it’s allowed to change the words, the image or the whole idea to get there. So, taglines need transcreation.

1. Start with a brief for each market. Say what the line has to do, as well as what it says.
2. Use AI as a sparring partner. Auto Adapt, a generative AI step in Phrase Orchestrator, adapts content for an audience or tone and generates variants for a writer to react to. If your brand copy has its own content group, those variants draw on its style guide and rules, so they start closer to how your brand sounds in each market.
3. Let a transcreator or in-market writer create or choose the line.
4. Review it in market, with someone who knows the brand and the audience.
5. Use the score for triage at most, for example on longer brand copy around the tagline. Never as a gate.

Automation still has a job, as it does with Phrase’s own marketing copy: moving the work between writers, reviewers and markets without a project manager chasing every step.

Setting thresholds, and checking they hold

A quality threshold is a dial between cost and risk. Lower it and more segments are confirmed without a person; raise it and more go to review. In Phrase, you can set QPS thresholds in your pre-translation settings or in Phrase Orchestrator, and choose which segments get locked from editing, and on what basis.

Phrase’s 2024 study shows what the dial is worth. For one technology customer translating English into German, about 35,000 segments in all, a threshold of 90 would have confirmed 37% of segments without editing, and a threshold of 95 would have confirmed 17%. For a team paying post-editors to check everything, that’s a lot of money spent confirming work that was already good.

A score is still an estimate, so treat any threshold as a hypothesis to test against what your reviewers actually change. The QPS Insights dashboard in Phrase Analytics shows how often linguists edit segments at each score and how long they spend on them. If they rarely change segments that score 88, you’re paying them to confirm good work. If they keep correcting segments above your line, either the line is in the wrong place or the score isn’t reliable for that content:

• Set thresholds per content type and language pair. Lower-resource pairs deserve extra caution, because engines and scores are both less reliable there.
• Move in small steps. Lower a threshold gradually and check what post-editors change just below the new line.
• Accept that some content has no threshold. For legal and brand copy, the score informs reviewers. It doesn’t replace them.

How to set up automated translation workflows by content type

In Phrase, the setup follows the matrix, and content groups hold it together.

1. Give each content type its own content group. A content group links a project to its style guide, style rules and AI quality checks, so legal text and brand copy are never held to the same rules. Quality evaluation draws its checks from the content group and the target language.
2. Set up MT profiles per content type, and per language pair where it matters. Select your preferred engine, or let MT Autoselect choose among the engines you enable in each profile. Autoselect weighs content domain for 12 major source languages, including English; for other source languages, it selects on language-pair performance alone.
3. Attach translation memory, termbases and QA rules to each content type.
4. Add AI steps where testing supports them. MT Optimize for post-editing, Auto Adapt for audience and tone, or the AI Translation Agent, which produces context-aware translations with built-in review and refinement.
5. Choose your quality checks and thresholds. Phrase’s quality evaluation is a toolbox rather than a single tool: QPS scores, QA checks, AI checks grouped by content group into quality profiles, and Auto LQA for an automated first review. Pick what each content type needs, and set thresholds only where the content can take one.
6. Chain the steps in Phrase Orchestrator. It can route content by type, use quality scores as decision points, add conditional review loops and hand work between AI and people, which is where most of the project management time comes back.
7. Watch the results. Tag projects with domain and subdomain metadata, and the TM Threshold dashboard can filter your results by domain, subdomain and target locale. Review them regularly, alongside QPS Insights.

Mistakes that cost the most

• Treating automation as all-or-nothing. Plenty of content can’t skip review, but almost all of it can skip the manual handoffs.
• Relying on a single signal. Scores, QA checks and human samples catch different problems. Use at least two for anything customers will see.
• Setting it once and forgetting it. These settings are sticky. Phrase’s own data shows that before teams had a dashboard to guide them, the average translation memory threshold across projects had sat at 76 for years. Once the data was in front of them, it rose to around 86.
• Judging production quality with BLEU. BLEU measures how many words and phrases overlap with a reference translation. It’s useful for comparing systems in research, and of little use for deciding whether a segment is ready to publish.

Where to start

Pick one content type you’re probably over-reviewing and one you might be under-protecting. For most teams, that’s support or product content on one side and something regulated or brand-critical on the other. Give each its own content group and its own checks, then watch what your reviewers change for a month.

Then ask the question this article has been circling: which of your content can a score vouch for, and which will always need someone to put their name to it?

To see how the score works, start with the QPS explainer or the 2024 threshold study. To find the quality evaluation tools that fit your content, start small: run a sample through one or two of them and compare the results with your reviewers’ judgment.

Where the numbers come from

• Quality score figures: a Phrase simulation study published in September 2024. One technology customer, English to German, about 3,500 jobs, 35,000 segments and 457,000 words. Source: Alon Lavie, “Understanding QPS and its expected savings and ROI.”
• Translation memory threshold figures: Phrase Analytics data across projects, published in April 2025. Source: Martin Švestka, “Mastering TM and Phrase QPS thresholds for localization automation.”
• Workflow practice and workload estimate: Phrase’s localization team, October 2026.
• Product behavior: Phrase help center and developer documentation, as of October 2026.

Appendix: the metrics

• QPS (Quality Performance Score). Phrase’s AI quality estimation. It predicts the score a professional reviewer would give a translation under the MQM framework, from 0 to 100, without needing a reference translation. Phrase groups scores into four bands: 95 to 100 excellent, 90 to 94 very good, 75 to 89 good, and below 75 poor.
• MQM (Multidimensional Quality Metrics). The framework human evaluators use to classify errors by type and severity, across categories such as accuracy, fluency, terminology and style. It was designed for human evaluation, and applying it to automated evaluation has its own difficulties.
• Edit distance (TER, chrF). How much a post-editor changed the output. A direct measure of effort, and the best check on whether a threshold sits in the right place.
• COMET. A neural metric widely used in research to compare MT systems. Its quality estimation variants work without a reference translation.