TextUnited
What we doPricing

Thursday, November 19, 2026·11 min read

You can build a translation tool in a week. Can you run it for a year?

Small separate violet cubes next to one large assembled cube with a cyan top layer.

The demo works

On Monday, your IT team shows you a demo. A developer has connected a large language model to the content system. She pastes a product description in English and clicks a button. Ten seconds later, the German, French and Polish versions appear. They read well. Someone from marketing says, "That is better than what the agency sent us last month."

The developer added one more thing: the company glossary, pasted into the instructions for the model. The brand name stays in English. The product names are correct.

Everyone agrees. Why pay for a translation platform when we can build this ourselves?

This is a fair question. The demo is real. Modern models translate well, and a good engineer can build a working prototype in a week. We have nothing against that prototype.

The problem starts in week two.

The first rule nobody wrote down

In week two, the first real request arrives. It does not come from IT. It comes from the people who write your product descriptions.

Your company sells sports equipment in more than ten product groups. Each group has its own approved terms. Ski jackets use one set of words. Hiking boots use another. So your content team keeps a separate glossary for each product group.

But some rules are true for every product. Your company may not write "Velcro", because Velcro is a trademark. It must write "hook-and-loop fastener" instead. This rule applies to jackets, boots, backpacks and everything else.

The team asks a simple question: "When we translate a ski jacket description, will the general rule still apply? Or do we have to copy it into all ten glossaries?"

This is not a made-up example. A customer of ours sent us almost exactly this question. They also told us that copying the same terms into ten or more product groups would not work in practice. They are right.

The developer can answer it, of course. She can write code for any rule. But which rule is the right one? That is not a coding question. It is a language question.

One question, five decisions

When this question reached our product team, four people needed most of an hour to agree on an answer. Building the answer took about a day. Here is what they had to decide.

  1. Where do general terms live? You can copy them into every product group. Then every change must be made ten times, and the copies drift apart. You can let one job use several groups at once. Then you need rules for which group comes first. Or you can keep company-wide terms outside any group, so they apply to every job. In TextUnited, this is what Terminology without a Domain now does.
  2. What wins in a conflict? Say the general list translates "cart" as a shopping cart. The product group for golf uses "cart" for a golf cart, a small vehicle. Both rules are correct in their place. Our answer: the more specific rule wins. A term in the product Domain beats a company-wide term. If a job has no Domain, all approved terms apply.
  3. Which words in the sentence are terms? Before a tool can apply a rule, it must find the term in the text. In "Add the golf bag to your shopping cart", is the term "shopping cart" or only "cart"? A wrong guess sends the sentence to the wrong rule. The result can be grammatically perfect and still wrong.
  4. What is banned, not only what is approved? "Use hook-and-loop fastener" is a different rule from "never use Velcro". The first tells the model what to write. The second needs a check after the translation, because a model can still produce the banned word. And someone must decide what happens then: stop the job, fix it, or send it to a reviewer.
  5. Who may change the rules? A freelance translator may own the French terms, and only those. A product manager may own one product group. A glossary file in a code repository does not know who is allowed to change which line.

None of these decisions is hard to code. All of them are hard to get right without people who work with language every day. And every answer will change as the business changes. We used one version of rule 1 before, moved away from it, and moved back when a real customer request showed us why.

A glossary is a file. Terminology is a set of decisions about that file.

Only 6 in 100 tasks were about the AI

The terminology question is one example. To see the full picture, we looked at our own work. We grouped about 1,000 tasks our product team worked on and sorted them by topic.

Only about 6% were about the AI itself: model settings, hallucinations, mistranslations the model missed. The other 94% were about everything around the model.

Here is what that "everything around" looks like:

  • Projects and workflows. Which jobs go to whom, what happens with internal and outsourced work, why a project does not appear in a list.
  • The editor where people review translations. Formatting tags in the wrong place, layout in InDesign and PowerPoint files, keyboard shortcuts, find and replace.
  • Terminology. Brand names that must never change, approved versus required terms, filtering by Domain and language variant, importing term lists from spreadsheets.
  • Connections to other systems. A website plugin that conflicts with a page builder, edits that must flow back to the content system, automation tools. These took the longest to close, often because a third party had to review or test the fix.
  • Costs and orders. Translator orders, estimates that must match invoices, word allowances.
  • Users and access. Sign-in loops, onboarding freelance translators, who may see which team's work.
  • Translation memory. Which earlier translation to reuse, and import files that do not parse.
  • File formats. Scanned PDFs, InDesign files, very large spreadsheets.

A fair note: some of these exist because we serve many customers. A tool built for one company will not need multi-customer invoicing. But it will need almost everything else. Your translators still need an editor. Your terms still need rules. Your content system still needs a connection that does not break when the other side updates.

This is the real build-vs-buy question. The model is a small part of the system, and you can rent it from many vendors. The rest is a product, and somebody has to run it.

Failures that do not look like failures

Here is a test from our own team. A colleague connected TextUnited to an AI assistant. The access key she created could read our Terminology, but it could not create translation projects. Then she asked the assistant to translate a file.

The assistant did not stop. It did not say, "I do not have permission." It translated the file by itself, without TextUnited, without the approved terms and without the quality checks. The file came back looking finished. Only when she asked directly did the assistant admit that it had not used TextUnited at all.

In a second test, the assistant returned two files: the one TextUnited produced, and a second version it made on its own. The text in both was the same. But you could not tell by looking which file came from the system with the rules, and which one the assistant made.

A broken system usually shows an error. A system built on a language model often does not. This is the pattern to plan for. A language model wants to help. When a rule, a permission or a connection is missing, it finds another way and delivers something that reads well. Other teams report the same. In a ClickHelp article, localization experts describe a glossary that went out of date without anyone noticing. A core product term came back as a word that does not exist in the target language. The client blamed the AI. The real cause was a glossary that nobody owned. And Paligo's help page explains that its AI translation can go back to its own term when it judges a glossary term too far from the original meaning.

So a home-built tool needs more than a good model. It needs a way to prove which rules were applied to each delivered text, and who approved it. Without that, you only find out when a customer, a regulator or a translator does.

Who owns the answers

Go back to the Velcro rule. A developer can build it in an afternoon. But she cannot decide whether it applies to every product group, or who may change it later. The decisions belong to different people, and a good tool must give each of them a place to work.

When we mapped this for our own product, we found at least four kinds of owners:

  • The IT admin manages users, access keys and connections. This person decides who can log in and which systems may talk to the tool.
  • The workflow owner, often a localization manager, decides who translates and reviews which content, with which teams and suppliers, and what happens after each step.
  • The language data owner decides which terms, style guides and earlier translations apply. This can be many people: one for French terms, one for a product group.
  • The compliance owner decides which content needs a human check and wants a record of who approved what and when.

In a large company, these roles also multiply. One department may want a simple process. Another may want a strict one. One rollout question from an industrial customer showed this well. Giving 500 people personal access keys does not scale. One shared key is easy to set up, but then nobody can see who translated what.

None of these are technical questions. They are questions about how your company works. A home-built tool must answer them in code, and change the code every time the answer changes.

What to build, what to buy

Building your own can be the right choice. It often is when:

  • you translate one or two language pairs,
  • the content is internal and low-risk,
  • you have no brand terms or legal wording to protect,
  • and one person can own the rules and the code for years.

For most companies that publish in many languages, the picture is different. The useful split is not build or buy. It is this: build what is specific to your business, and buy what encodes language decisions.

Build the parts only you can build: the connection to your content system, the automations that move content between your tools, and the places where your teams start and end their work. These parts are close to your business, and you know them best.

Buy the layer that stores and applies your language decisions: Terminology with Domains and clear priority rules, style guides, translation memory, review steps, quality checks, and a record of who approved what. This layer is where most of the work in our task list sits. It is also the part that keeps changing as models, formats and customer needs change.

With TextUnited, your team can do both. Developers connect through the API, an automation tool or an AI assistant through MCP. Language teams work in the same system through the UI. The rules live in one place, whichever way the content arrives.

Questions to ask before you build

If your team is planning its own tool, these questions help. You can forward them to IT as they are.

  1. Where do company-wide terms live, and where do product-specific terms live?
  2. When two terms conflict, which one wins? Who decided that?
  3. How does the tool find terms in a sentence, and what happens when it picks the wrong one?
  4. How do you block a word that must never appear, not only suggest the right one?
  5. Who may change the terms for one language or one product group, and who may not?
  6. What happens when a permission or connection is missing? Does the tool stop, or does it translate anyway?
  7. Can you show which rules and which people touched a delivered text?
  8. Who maintains the editor where translators and reviewers work?
  9. Which file formats must the tool handle without breaking layout?
  10. Who updates all of this when the model changes, and how do you test that nothing broke?

If you can answer every question and name one owner for each answer, building may work for you. If many answers are "we will decide later", you are not planning a small internal tool. You are planning a product that someone must run for years.

Our recommendation: build your own tool only when it solves one narrow problem. For example, it translates short, simple texts into one or two languages, or it handles only one file format, and no brand terms or legal wording are at stake. In every other case, buy the language layer and build only the connections to your own systems.

Frequently asked questions