How Do You Move a Regulated Workflow From Machine Translation to Governed LLM

1 hour ago

6 min read

Language specialist reviewing automated text-processing guidelines before an LLM translation migration

You don't rip out machine translation and switch to a large language model overnight. You migrate one workflow at a time, and you keep the audit trail intact while you do it. Segment-level machine translation renders a document sentence by sentence with no memory of the sentence before, which is exactly why it drops terminology and mistranslates cross-references in a regulated file.

AD VERBUM is an EU-hosted AI+HUMAN hybrid translation company that moves regulated workflows off legacy machine translation onto a governed, client-tuned LLM, with certified subject-matter review under ISO 17100, ISO 18587, ISO 27001 and ISO 42001. We run the migration on EU-hosted infrastructure, with no training on client data and no public-cloud processing.

The point of the move isn't speed. It's control. A governed LLM workflow reads the whole document, applies your approved terminology, and leaves a record a reviewer and an auditor can both follow. Here is how to make that switch without breaking compliance.

What changes when you move from MT to a governed LLM

Segment-level machine translation is deterministic and cheap. It splits text into segments, translates each one in isolation, and never sees the paragraph it sits in. For marketing copy that's tolerable. For an instructions-for-use file or a defence tender, a term that shifts between segment 12 and segment 340 is a compliance defect, not a stylistic quibble.

A governed LLM works at document level. It reads context, holds terminology across the file, and produces fluent output that then goes to a certified human post-editor. That post-editing step is defined by ISO 18587, the standard for full human post-editing of machine-translation output. Its revision, expected October 2026, clarifies the line between machine translation, AI-generated output and the human post-editing task, and aligns more closely with ISO 17100.

Governed also means disclosed. Under EU AI Act Article 50 (Regulation (EU) 2024/1689), transparency obligations for AI-generated content apply from 2 August 2026, and generative systems already on the market before that date have until 2 December 2026 to meet the machine-readable marking requirement. A governed workflow records where the model generated text and where a human signed it off.

Linguist making terminology notes to constrain a governed LLM translation workflow

The migration, step by step

Run it in this order. Each step depends on the one before, and you don't retire the old engine until the last check passes.

Pick one workflow, not the whole operation. Choose a single document type in a single language pair, ideally one with stable terminology and a named reviewer, and migrate that before you touch anything else.
Export your terminology and translation memory. Pull the client Term Base and Translation Memory out of the old system first, because a controlled terminology base is what constrains the model. Without it, a language model invents plausible synonyms, which is the last thing a regulated file needs.
Choose where the model runs. For regulated content, run a self-hosted open-weight model on EU-hosted infrastructure, so no controlled or personal data leaves your jurisdiction and nothing feeds a public model.
Constrain generation with your assets. Feed the Term Base and Translation Memory into the workflow so the model reuses approved wording and past sign-offs, instead of generating each segment from scratch.
Post-edit with a certified subject-matter expert. A qualified linguist reviews the output to ISO 18587 and ISO 17100, checks terminology and regulatory accuracy, and takes responsibility for the final text. This is the step that keeps the machine in an assistive role.
Log the governance. Record model version, prompt or configuration, the assets used and the reviewer sign-off under an ISO 42001 AI management system, so every output carries auditable provenance.
Run both engines in parallel, then cut over. Translate the same files through the old MT and the new governed workflow, compare terminology consistency and error rate, and retire the legacy engine only once the governed output wins on both.

Only after the parallel run confirms the governed workflow is more accurate, not just faster, do you switch the workflow over and archive the old engine's configuration.

Hands connecting an encryption device in a secure EU-hosted translation workspace

Where it goes wrong

Most failed migrations trip on the same points. Watch these:

Migrating everything at once. Move the whole operation in one step and you lose the parallel-run baseline that proves the new workflow is safer.
Treating the LLM like a bigger MT engine. Drop it in without a Term Base and you get fluent output that reads well and still carries the same segment-level errors a reviewer has to catch.
Skipping the human post-edit. An unedited language-model draft is not a certified translation. ISO 18587 requires a qualified post-editor to take responsibility for the final output.
Sending controlled data to a public API. For dual-use technical files, transmitting controlled technology to an external model can be a transfer under Regulation (EU) 2021/821 Article 2, and the EU AI Act adds transparency duties on top, which is why the model has to run inside your infrastructure.
Keeping no provenance record. If you can't show which model produced a passage and who approved it, you can't answer an auditor or meet the EU AI Act Article 50 transparency expectation.

Our AI and governance translation services

translation services for regulated sectors run on ISO 27001 and ISO 42001 certified, EU-hosted infrastructure, with no reliance on public cloud tooling for core processing. Every project runs through our AI+HUMAN hybrid workflow: we ingest client Translation Memories and Term Bases first, our proprietary LLM-based LangOps System generates output constrained by client terminology on client-tuned open-weight models, and our certified subject-matter experts review for technical accuracy and regulatory compliance. Our QA is aligned to ISO 17100 and ISO 18587, with sector-specific requirements such as the EU AI Act (Regulation 2024/1689) and ISO 42001 AI-management governance applied where relevant. We serve Life Sciences, Legal, Finance, Defense, and Manufacturing clients across 150+ languages with 3,500+ subject-matter linguists. For teams managing audit-sensitive content, contact us to discuss your security and compliance requirements directly.

FAQ

What is a governed LLM translation workflow?

It's a workflow where a large language model generates translation at document level, constrained by the client Translation Memory and Term Base, and a certified human post-editor then reviews the output under ISO 18587 and ISO 17100. Governance means the model version, assets and sign-off are logged under an ISO 42001 AI management system, so every output is auditable. AD VERBUM runs this as an AI+HUMAN hybrid workflow.

How is a governed LLM different from machine translation?

Segment-level machine translation renders one sentence at a time with no document context, so terminology and cross-references drift. A governed LLM reads the whole document, applies approved terminology from the client Term Base, and produces coherent output that a certified human then post-edits and signs off under ISO 18587.

Does the EU AI Act apply to LLM translation?

Yes, through transparency. EU AI Act Article 50 (Regulation (EU) 2024/1689) requires AI-generated content to be disclosed and marked, applicable from 2 August 2026, with generative systems already on the market given until 2 December 2026 for machine-readable marking. A governed workflow records where the model generated text and where a human approved it.

Do I still need human review with an LLM?

Yes. ISO 18587 requires a qualified human post-editor to review machine-translation output and take responsibility for the final text, and for regulated content that review is not optional. AD VERBUM's certified subject-matter experts perform it as part of the AI+HUMAN hybrid workflow, with QA aligned to ISO 17100.

Is it safe to translate controlled or personal data with an LLM?

Only when the model runs on infrastructure you control. Transmitting controlled dual-use technology to an external model can be a transfer under Regulation (EU) 2021/821 Article 2, and personal data carries GDPR residency duties, so AD VERBUM runs self-hosted open-weight models on EU-hosted infrastructure with no public-cloud processing and no training on client data.

What does ISO 42001 add to a translation workflow?

ISO 42001 is the certifiable AI management system standard. It requires documented risk assessment, data governance and human oversight for AI systems, which for translation means logging model version, terminology assets and reviewer sign-off so every output has auditable provenance. AD VERBUM is ISO 42001 certified.