If you’re getting close to selecting an AI translation vendor, this is the last checklist you need before signing. In a sales pitch, anything looks possible. But the verbal pitch and the actual operating system can be very different things. Below are the items to verify in writing — not just hear in a presentation — across three areas: security, quality, and terminology management.
Security: Where Does Your Data Go?
In AI translation, security isn’t a technical concern — it’s a business risk. In manufacturing, finance, healthcare, and defense in particular, this is the area where the answer often determines whether AI translation is even an option.
Checkpoint 1: Will your source content be used to train AI models?
Verify whether the source content you submit for translation will be reused to improve or train AI models. A verbal “no, it won’t” isn’t enough. Confirm in writing that the contract spells out the scope of data use, and that the clause specifically addresses AI model training.
Checkpoint 2: Can the vendor choose which AI to plug in?
If the vendor is tied to a commercial translation platform, you’re locked into whatever AI that platform provides — and bound by that AI’s security policy, whatever it happens to be. A vendor running its own platform can pick an AI that matches your security requirements, and switch to a more appropriate one if the AI provider’s policy changes.
Checkpoint 3: If customer data is used for tuning, how exactly?
If the vendor develops or tunes its own AI models, ask whether your data is used directly as training material or indirectly (through synthetic or reconstructed data). Either way, one principle has to hold: data from one customer must never end up in another customer’s model in a form that could be traced back to the source.
Checkpoint 4: What security certifications does the vendor hold?
Confirm whether the vendor holds an information security certification such as ISO 27001. Just as important: ask whether the certification covers the full translation service, or only specific parts of the operation.
How Hansem Global handles this. Hansem Global runs every translation project under an ISO 27001-based security framework. Because the AI Workstation is our own platform, we integrate only with enterprise-grade AI services that don’t use customer data for model training.
When we develop domain-specialized models, we use synthetic and reconstructed data, so no individual customer’s data remains in identifiable form. For customer-dedicated models, that customer’s data is used solely in that customer’s dedicated model and is never introduced into other customers’ projects or general-purpose models.
Quality: Is the Output Coming From a Repeatable System?
You can’t judge AI translation quality from the output alone. Even when a single project comes back looking great, you need to know whether that result was produced by a repeatable system — or whether it just happened to land well that time.
Checkpoint 1: Are AI translation and AI QA connected?
Ask whether the workflow is “AI translates → human reviews” or “AI translates → AI checks the output → human reviews flagged items.” As we covered in the previous article, the fluent mistranslations produced by LLM-based translation are hard for humans to catch on their own. The key question: is there a multi-stage structure where AI QA first checks semantic alignment with the source, then routes mismatches to expert linguists for focused review (AI translation → AI QA → expert linguist PE)?
Some translation platforms have started offering AI QA features. But verify whether the accuracy and stability of those features are at production grade. A demo or pilot is the right way to validate this directly.
Checkpoint 2: Can QA criteria be customized?
Ask whether the vendor offers only standardized QA, or whether QA items can be added and adjusted to match your industry and document type. Reviewing safety warnings in an automotive manual and reviewing the tone of marketing copy are completely different jobs.
Checkpoint 3: How is quality measured?
Ask how translation quality is measured, how automated evaluation is combined with human evaluation, and what the response process looks like when a quality issue is found. “We guarantee quality” is something every vendor says. Ask for the specifics behind that guarantee.
Checkpoint 4: Can you trace the output of each stage?
You shouldn’t only receive the final output. Ask whether you can review the AI translation, the AI QA findings, and the linguist’s edits at each stage. Without that transparency, it’s hard to diagnose where a quality issue came from — and impossible to chart a meaningful improvement path.
How Hansem Global handles this. Inside the AI Workstation, AI translation → AI QA → expert linguist PE all run as a single connected flow, and customer-specific QA modules can be added and extended. The output of each stage is traceable within the platform, and we validate quality up front through pilot projects on your actual content.
Terminology: Is Consistency Maintained End to End?
Terminology management is the most underestimated factor in AI translation quality. Many companies think terminology is handled the moment a glossary is shared at the start of a project. The reality is more complicated.
Checkpoint 1: What does the vendor do when you don’t have a glossary?
Many companies don’t have a systematized glossary. In that situation, ask whether the vendor’s response is “send us your glossary and we’ll apply it” — or whether they have the in-house capability to extract domain-specific terminology from your documents and propose a structured glossary themselves. The early-project quality difference between vendors who can do the latter and vendors who can’t is significant.
Checkpoint 2: Do extracted terms flow automatically into translation?
If the glossary lives only as a separate file disconnected from the translation system, you end up relying on each translator’s individual attention. Verify that the terminology standard is automatically applied at the translation stage and re-checked for consistency at the QA stage.
Checkpoint 3: How is terminology managed across multiple languages?
When the same source content is translated into multiple languages at once, ask whether there’s a system that keeps terminology aligned across each target language. Translating from English into German, Japanese, and Chinese simultaneously — where the same source term ends up meaning different things in different languages — is a common problem.
Checkpoint 4: Who owns the translation memory and glossary, and how are they returned?
Translation memory (TM) and glossaries built during a project are valuable linguistic assets. Confirm in the contract whether you (the customer) own these assets, and exactly how they’re returned at the end of the engagement — in what file formats, and covering what scope.
How Hansem Global handles this. The AI Workstation’s terminology extraction feature can identify domain-specific terms even on large projects where the customer hasn’t provided a glossary, organizing them with multilingual equivalents and contextual notes. Extracted terms feed directly into the translation and QA stages, so consistency holds across the full workflow.
At-a-Glance Checklist
| Area | Key Checkpoint |
|---|---|
| Security | Will source content be used to train AI models? |
| Is the vendor on its own platform or tied to a third-party platform? | |
| Does model tuning use synthetic or reconstructed data? | |
| Does the vendor hold security certifications (ISO 27001 or equivalent)? | |
| Quality | Is there a 3-stage defense (AI translation → AI QA → expert linguist PE)? |
| Can QA criteria be customized for your industry and content type? | |
| Are quality measurement criteria clearly defined? | |
| Can the output of each stage be traced? | |
| Terminology | Can the vendor extract terminology when no glossary exists? |
| Are extracted terms automatically applied during translation and QA? | |
| Is there a system for cross-language terminology consistency? |
The Bottom Line: Good Questions Find Good Partners
The most important question when selecting an AI translation vendor isn’t “which AI does it use?” It’s what security framework, quality structure, and terminology system sits behind that AI.
Take the checkpoints in this article into your vendor meetings and ask them directly. The difference between vendors who can answer clearly and specifically — and vendors who deflect — will be obvious.
Hansem Global has built a system that can answer every one of these questions specifically. We integrate security, quality, and terminology management on a single platform — the AI Workstation — and we continue to evolve it in response to what real projects demand.
Adopting AI translation isn’t a technology decision. It’s a partner decision. The right questions are how you find the right partner.
This is the final article in our series, How to Adopt AI Translation the Right Way. Read the rest of the series:
- Part 1. Where Should You Start with AI Translation?
- Part 2. Why Do Results Differ So Much If Everyone Uses AI Translation?
- Part 3. How to Go Beyond the Limits of General-Purpose AI Models in Translation
- Part 4. Is AI Translation Effective for Every Project?
- Part 5. Where Do AI Translation Quality Failures Happen, and How Can You Prevent Them?