Sinch: more content, better quality, every month.
Case Study . Sinch
Sinch localizes three brands into many languages, and AI translates most of the first drafts. Undertow works inside their Crowdin setup as their fractional language team. Every month we find the problems that keep coming back and tweak the system to fix the root cause. The AI gets better, the editing gets lighter, and Sinch can put more content through without growing the team.
"Everyone thinks the hard part of AI translation is picking the right model. It isn't. It's the glossaries, the style guides and the prompts that the model runs on. All these linguistic assets have to be rebuilt for a machine instead of for a person. Fine tuning all this is what we do with Undertow every month."
Alfonso González Bartolessis, Senior Localization Manager, Sinch
.png)
5M words
in Q1 2026. More than all of 2025.
2 QUARTERS
languages, same marketing team
67 fixes
shipped in the first month of the quality loop
12%
of AI words need real editing in the projects we review. That is where the linguists spend their time.
Brands: Sinch, Sinch Mailjet, Sinch Mailgun.
Languages measured: French, German, Spanish, Brazilian Portuguese.
Platform: Crowdin.
Scope: help centers, marketing, product and web, in-app and UI, subtitles.
Services: Custom AI Translation, AI Translation Quality, Fractional Localization Management..
What happens after the AI is working
Most localization stories start with a team drowning. This one does not. Sinch already had a strong localization function, a team that knew what good translation looks like, and AI producing most of their first drafts.
Once the machine is drafting at that scale, volume stops being the question. The question is whether any of it is getting better.
Three things made that hard to answer.
The glossaries and style guides were written for people
A human translator fills in what the glossary does not say. AI does not. It needs every rule and exception written down.
Everything was getting the same treatment
A help center article and a product screen carry very different risk. The pipeline treated them the same.
Improving the system meant doing it by hand
Someone had to read every mistake, decide which category it belonged to, and type it into a spreadsheet, one at a time.
.png)
We changed the system around the AI
The move to Crowdin took half the time planned
Their previous platform sent every piece of content down the same fixed path, with no way to adjust the AI. So Sinch chose Crowdin. Moving an operation this size, across three brands and everything they publish, usually takes a year. We ran the move from inside, working alongside their team.
"Moving all our localization tech stack to a single new platform was initially planned as a year of work. With Undertow working alongside us, it only took two quarters, and we spend the rest of the year improving the system instead of still building it."
Alfonso González Bartolessis, Senior Localization Manager, Sinch
Different content, different pipeline
Not all content carries the same risk, so we stopped sending it all through the same process.
2 Brands, 2 Help Centers
Automated AI translation at volume, each tied to its own brand guidelines.
Marketing and web
Its own workflow, separate from the help centers, with its own prompt. Marketing copy needs a different tone from a help article, and the prompt says so.
In-app and UI, from Figma and GitHub
Tighter pipelines, with human evaluation on the strings customers see most often.
Subtitles, with human review every time
Daniel Collado, Sinch's Localization Program Manager, built a solution inside Crowdin for subtitles, so our linguists only have to check that the timing and the line breaks work in each language and that the text still matches what is on screen.
Crowdin gives you the controls to build a different pipeline for each type of content. What it cannot do is tell you which type needs which pipeline. That decision comes from knowing the business, and we make it together with the Sinch team.
Glossaries and prompts rewritten for AI, not for people
A glossary written for a human translator is a reminder. A glossary written for AI is an instruction. That difference is where most of the quality comes from.
So we rebuilt Sinch's glossaries and style guides around the way the AI reads them. The glossaries are layered: one main glossary for terms shared across Sinch, plus a smaller one for Sinch Mailjet and one for Sinch Mailgun. A term can mean one thing in one brand and something different in another.
The prompts spell out the rules: how the content will be used, what changes from market to market, and which terms must never appear. Nothing is left for the AI to guess.
Every month builds on the last one
Setup happens once. This keeps going, and it adds up. Every month, the corrections the linguists made become permanent improvements to the glossary, the style guide and the AI configuration. Those improvements stay. The system never goes back to where it started.
SINCH × UNDERTOW × CROWDIN
The monthly quality loop
1
The AI drafts it
Pre-translation in Crowdin,tuned to the content type
2
A linguist reviews it
They change the text.No categories, no spreadsheet
3
Two separate AI agents
One classifies every change.One ranks the fixes by impact
4
Undertow ships the fixes
Glossary, style guide, AI configuration
The usual way to do this is by hand
A person goes through the output, finds every mistake and categorizes it one at a time. It works. It also asks a linguist to spend the afternoon on data entry rather than on the thing they are good at.
So we built AI agents to do that part
At Sinch the reviewing is done by our own linguists, a team we put together for Sinch's markets and their kind of content. They change the text, and nothing else.
One agent then reads every difference between the first AI output and the version the linguist approved, and puts each one in a category. A terminology error. A number or a date in the wrong local format. Something the AI could not know, because it could not see the surrounding content. Or simply a wording the linguist preferred.
A second agent takes that whole list and works out which language assets should change, what kind of fix each one needs, and how much difference it would make. It groups the fixes, ranks them by impact, and hands the language specialists a short report. So the specialists spend their time on the changes that matter most and are realistic to make.
They can still see the full list, not only the short report, so anything the first agent treated as a personal preference can be put back in. This is still work. But reading a ranked list is much faster than building one.
The short list is a suggestion, the decision sits with the language specialist
Our language specialists read the report and decide what should change. Nothing is changed in the glossary, the style guide or the AI configuration until the Sinch team has approved it. That is now the only part of the analysis where a person spends time.
"Categorizing errors by hand used to eat hours of linguist time, and it is work that is never really finished. Now the review happens, the agents do the classifying, and Undertow brings us a short list of changes to approve. That is the part that makes a difference over time."
Alfonso González Bartolessis, Senior Localization Manager, Sinch
Recurring problems become permanent fixes
A problem that appears once is noise. A problem that keeps coming back is a missing rule. Each one becomes a glossary entry, a line in the style guide, or a rule in the AI configuration. No rule works everywhere, and the AI will still get things wrong sometimes. What changes is that you stop correcting the same mistake every month.
Sinch decides how much editing each project gets
There are two levels. Standard Editing means the translation has to be correct, clear and compliant. If the AI already got it right, we leave it alone. Full Editing means we polish the style as well. Sinch chooses the level for each project, so the effort matches the risk and no linguist has to guess how far to go.
We built Sinch a dashboard. They open it whenever they want. It shows how the program is developing: how the numbers move month after month, what the work on the glossaries and prompts is doing to the AI output, and where the effort is going.
Because Crowdin exposes the AI workflow steps at a granular level, we can apply our operational and linguistic expertise where it actually moves quality, and keep refining the whole system as we go.
The service
Language Intelligence
All of this has a name. Language Intelligence is the work of making an AI translation system better on purpose: the monthly loop, the two AI agents, and the language specialists who decide what to change. Everyone wants better AI translation. This is the work that produces it.
We run it for Sinch inside Crowdin. We can run it on your setup, in whatever platform you already use.
Four things are different now
"Before, I could tell you how much we translated. Now I can tell you whether it is getting better, and show exactly what we changed to make that happen. That is a very different conversation to have with leadership."
Alfonso González Bartolessis, Senior Localization Manager, Sinch
The AI has instructions now, not hints
The fixes go into the glossary, the style guide and the AI configuration. Each fix covers a whole type of mistake, not only the one place where we found it.
Human attention now goes where it makes a difference
Human review is concentrated where the risk is: product interfaces, subtitles and the web pages that matter most. Help center content runs automated at volume. That split is a decision now, not an accident.
Volume stopped being the limit
Sinch processed 5M words in the first quarter of 2026, more than they had translated in the whole of 2025. AI drafts 65%, human translation covers 27% where the content is sensitive or complex, and translation memory handles the remaining 8%.
Localization now has a model it can offer other teams
Teams at Sinch without their own localization process no longer have to build one. They can send their work through the system the localization team already runs.
"We want to bring every team into our localization workflow. There are teams that don't have a fully defined localization process yet, and we want all their needs to be channeled through localization and Crowdin."
Alfonso González Bartolessis, Senior Localization Manager, Sinch