AI translates in seconds. Whether the result is usable is decided earlier, by your terminology. Without approved terms to work from, a language model makes a fresh choice every time a technical term comes up, and the choice varies. This guide explains why AI translation needs a termbase, what inconsistent terminology actually costs, and how to build a termbase in five steps that holds up in day-to-day operation.
Why AI translation stays inconsistent without a termbase
Terminology management is the systematic collection, definition and upkeep of a company's technical terms across all its working languages. The result is a termbase: a central database that records the approved translation, definition and usage context for every term. For AI-assisted translation, it is the single most important quality foundation there is.
The reason sits in how the models work. A language model picks terms by statistical probability, not by your terminology. For most technical terms, three or four plausible translations exist. Which one the model picks depends on context, on the prompt, and sometimes on nothing more than chance. Across a single document, this rarely shows. Across a hundred documents, three departments and five language pairs, it turns into systematic drift: the same term appears in several variants, scattered across your website, contracts, documentation and support articles.
The market has caught on. Around 55% of large enterprise buyers now require translation models trained on their specific domain. Adaptive AI systems learn from glossaries, translation memories and style guides to get there. The catch: every one of these systems is only as good as the terminology it is fed. A neglected termbase makes even the best model unreliable.
How to spot terminology drift in your own content
- The same product or technical term shows up with different translations in quotes, on the website and in the documentation.
- Review rounds keep circling back to the same handful of terms.
- Every department orders translations separately, with no shared glossary.
- Your terminology lives in spreadsheets, old PDFs and the heads of a few experienced colleagues.
- A new supplier or a new tool noticeably shifts the vocabulary.
If three or more of these apply, your company is most likely publishing inconsistent content already. Every AI-generated text adds volume, and with it, drift.
What inconsistent terminology actually costs
Inconsistent terminology creates costs on three levels: direct correction costs, risk costs and brand costs. The first two can be put in numbers. The third builds up over time and is often the most expensive of all.
Direct costs: review rounds and re-checks
Every term deviation caught in review triggers a loop: query, clarification, correction, re-approval. Caught after publication, it also triggers a re-check and replacement of the affected documents. A model calculation: if a 40-page technical manual has to be re-checked because a safety-critical term was translated two different ways, you are quickly looking at 15 to 25 hours of checking, coordination and re-approval, spread across the technical team, documentation and the supplier. Building a starter glossary of your 100 most important terms usually costs less than a single loop of that kind.
Risk costs: when terms touch liability
In technical and regulated industries, terminology is not a matter of style. If an operating manual describes the same safety-critical action with two different terms in two documents, it opens room for interpretation that nobody wants: not the manufacturer, not the operator. In regulated industries, the potential follow-on costs of wrong terminology outweigh the cost of professional terminology management many times over. That is exactly why the question "how do you manage terminology?" has become a standard criterion in many tenders.
Brand costs: a brand that sounds different in every language
A translation memory keeps your sentences consistent. Terminology keeps your brand consistent. Without a termbase, your company sounds slightly different in every market: the claim is sometimes translated and sometimes not, the core product term flips between variants, the tone shifts. Customers rarely name the effect, but they feel it. Recognisability is a matter of terms, and terms can be controlled.
Building a termbase in five steps
A working termbase comes together in five steps: taking stock, prioritising, the do-not-translate list, an approval process, and integration into the translation workflow. The order matters, because each step builds on the one before it.
Step 1: Take stock
Gather everything that holds terminology today: marketing spreadsheets, glossaries from past translation projects, style guides, product databases and the knowledge of long-serving colleagues. The goal is not a perfect inventory. It is an honest picture of where things stand, including the places where your sources contradict each other.
Step 2: Prioritise your core terms
Not every word needs an entry. Start with the 50 to 100 terms that are business-critical: product names, safety-related terms, contract language, core marketing vocabulary. In practice, this selection delivers around 80% of the benefit. Bring the technical teams in early. They know best which terms carry real weight.
Step 3: Set the do-not-translate list
Define what must never be translated: product names, brand terms, protected designations. This do-not-translate list is the simplest and most effective part of any termbase. Also record the forbidden variants explicitly, so the database holds both the correct form and the forms people keep reaching for by mistake.
Step 4: Define the approval process
Without clear ownership, every termbase dilutes within months. Decide who may propose new terms, who reviews them, and who has the final say. What works in practice is one central owner with decision-making authority and a lightweight process: proposal, expert review, approval, distribution. Record a definition, the context and one example sentence for every term.
Step 5: Integrate it into the workflow
A termbase only earns its keep once it lives inside the translation workflow. Connected to your translation memory, CAT tools and AI translation, every deviation is flagged automatically before a text goes out. Corrections from review flow back into the database, so the same mistake does not come back. This is where most home-grown solutions fail: the list exists, but nobody works with it.
Terminology management with tolingo
As a translation agency holding four certifications (ISO 17100, ISO 18587, ISO 9001 and ISO/IEC 27001), we take on the full build and operation of your terminology: from term analysis and approval processes through to integration with translation memory and AI-assisted workflows. Your termbase grows with every project and remains your asset. In our Premium service under ISO 17100, a second specialist translator additionally revises every translation, terminology consistency included.
You can find the full scope on our terminology management page, along with translation memory and post-editing. Or talk to us directly: we will gladly walk you through how a termbase fits into your existing processes, with no strings attached.
Frequently asked questions about terminology management
What is the difference between a termbase and a translation memory?
A translation memory stores whole sentences from past translations and suggests them when content repeats. A termbase fixes individual terms, complete with definition and context. The translation memory delivers efficiency and sentence-level consistency; the termbase delivers consistency of terms and brand. They only reach their full effect together.
How many terms does a termbase need at the start?
50 to 100 business-critical terms are enough to begin with: product names, safety-related technical terms and core marketing vocabulary. In practice, this selection covers most of the critical cases. The termbase then grows with every project.
Is terminology management worth it for smaller companies?
Yes, as soon as you translate into more than one language on a regular basis, or several people and tools work on your texts. A starter glossary takes limited effort and usually costs less than a single larger round of corrections.
Does a termbase work with AI translation?
A termbase is what makes AI translation consistent in the first place. Modern workflows hand the approved terms straight to the model and check the output against the database automatically. Deviations are flagged before the text reaches approval.