Skip to content

Building an AI Translation Workflow: What Happens at Audit, Build and Run

By Volkan Güvenç, founder of Alafranga Language Solutions
 
Most problems with AI translation are not engine problems. The output is inconsistent because the translation memory is not applied, the glossary is not enforced, the review step is a light proofread, or content that should never have gone to AI went anyway. Changing the engine rarely fixes any of this. Changing the workflow does.
 
This article describes how we build those workflows, step by step, in the three stages we use on our workflow consulting page: Audit, Build and Run. It is based on our own production work: large machine translation post-editing (MTPE) projects since 2017, starting with 2.2 million words of e-commerce product content for a European translation agency, delivered by linguist teams on three continents against daily batch deadlines, and our controlled AI-assisted translation workflow (SmartEdit), in client delivery since 2018.

 

▮Audit: deciding what goes to AI

The audit answers one question before any tool is chosen: which content can be drafted by AI, and under which controls. It works on the client's real files, not on a sample chosen to make AI look good.
 
  • Content classification. Every content type is assigned to one of three routes: AI draft with human review, AI draft with review and an independent revision by a second specialist, or fully human translation. The criteria are the consequence of an error, regulatory status, repetition and translation memory leverage, the quality of the source text, and confidentiality restrictions in client contracts.

  • Risk mapping. Where would an error have legal, safety or commercial consequences? A wrong torque value in a maintenance procedure, a mistranslated exclusion clause and an inconsistent UI label are three different kinds of risk and need three different controls.

  • Asset review. The condition of the translation memories and glossaries decides how much AI can achieve. We look for duplicate and contradictory translation units, outdated terminology, misaligned segments, glossaries without approval status, and TMs split across tools or vendors. A polluted TM applied with high priority produces confident, consistent errors.

  • Cost comparison. The comparison includes review time, revision, terminology work and project management, not only the engine cost per word. Some content gets cheaper with AI. Some does not, once the review it needs is counted honestly.

  • Deliverable. A written workflow specification: routes per content type, quality gates, responsibilities, tools and a roadmap. It is written so that the client's team can implement it and the client's own customers or auditors can follow it.

▮Quality gates: what is checked, by whom, at which step

The core of the specification is a set of gates. A typical AI-assisted route looks like this; the audit adapts it to the content.
 
Gate What is checked Who
Source preparation File format, segmentation, locked and non-translatable content, source errors flagged before translation Project manager, engineer
Pre-translation Translation memory matches applied first; AI drafts only the segments below the agreed match threshold, with the glossary applied Automated, configured per project
Automated QA Numbers, units, tags, placeholders, terminology, forbidden terms, length limits, consistency across segments QA tools in the CAT environment
Human review Meaning against the source, terminology, register, omissions and additions Specialist translator in the subject field
Revision Independent second check where the content route requires it Second specialist
Feedback Recurring AI errors fed back into prompts, glossary and TM Project manager and reviewer
 

▮Build: setting the workflow up in the client's tools

Once the routes and gates are agreed, the workflow has to work with real translators, real file formats and real deadlines. We build it in the client's environment rather than moving the client into ours.
 
  • AI drafting inside the CAT tool. GPT, Claude or DeepL are connected via API to memoQ, Trados or Smartcat. The translation memory takes precedence, the glossary is applied to every segment, and the AI output arrives in the same editor the reviewer already uses, with the same QA checks.

  • Context beyond the segment. CAT tools split text into segments; AI works better when it sees the whole document. We set the workflow up so that the full project or document is in the model's view, not one sentence at a time. Why this matters: the quiet conflict at the heart of translation technology.

  • Engine selection by testing. Which model is used for which content and language pair is decided by testing on the client's own material and scoring the results, not by default. The answer often differs between, for example, marketing copy and maintenance procedures.

  • Prompt design. Prompts carry what the TM and glossary cannot: register, audience, style rules, do-not-translate lists, unit and number formats, tag handling and product-specific conventions. They are versioned and documented like any other project asset.

  • Automated QA. Checks for numbers, tags, placeholders and terminology drift run before a reviewer sees the text, so that human review time goes to meaning, not to mechanical errors.

  • Platform setup. For multi-translator programmes: shared TMs across language pairs, vendor assignment, and a client portal where it is needed.

  • Migration. Moving off legacy tools and fragmented TMs means cleaning and consolidating translation memories. Consolidation changes match rates, and we state the expected effect on match quality before the move, not after.

  • Content system connections. CMS connections through XLIFF exchange or API, with routing by language pair or content type. ERP and other document systems are scoped case by case; if an integration is not feasible or not cost-effective, we say so before any work starts.

  • Handover. A tested workflow, full technical documentation and a briefing for the client's team. The client owns the configuration, prompts and assets.

▮Run: operating the workflow

Some clients run the workflow themselves after Build. Others publish content continuously and do not want an in-house language team. For them we operate the workflow:
 
  • Production. AI drafting and human review in every language, with a named reviewer per language, so that the same person sees the same product line over time.

  • Asset maintenance. TM and terminology kept current as products, regulations and the client's own wording change.

  • Quality reporting. A monthly report on error types, review findings and terminology changes. The report is what makes the workflow auditable and shows where prompts or glossaries need adjusting.

  • Pricing. A monthly retainer with an agreed volume band, a fixed price per product release, or a combination. No per-word invoicing, because the work is the operation of the workflow, not the word count.

▮Where AI translation workflows usually fail

  • AI applied to everything. The same route for UI strings, contracts and safety instructions.

  • TM bypassed. The engine translates segments that already have approved translations, and the new output contradicts them.

  • Glossary not enforced. The glossary exists, but nothing checks that the AI output follows it.

  • Review as proofreading. The reviewer reads the target text only, without the source, and misses omissions and meaning errors that read fluently.

  • No feedback loop. The same AI errors are corrected by hand in every project because they never reach the prompts or the glossary.

▮What stays human

Medical device documentation, texts with legal effect and safety-critical instructions run through our fully human workflow (HumanExpert) by default. Where a sworn translator's signature is required, the document stays with a sworn translator. An audit that recommends AI for everything is not an audit.
 

▮Questions about AI translation workflows

  • Can AI drafting be used inside memoQ and Trados without changing tools? Yes. The engines are connected via API, the translation memory and glossary remain in place, and reviewers work in the same editor and QA environment as before.

  • Does the TM still matter if the AI is good? More than before. The TM is the record of approved decisions; the AI should only draft what the TM does not cover, and a polluted TM spreads errors consistently.

  • How is the right engine chosen? By testing candidate engines on the client's own content and language pairs, scoring the output, and deciding per content type.

  • Is client content used to train AI models? No. Content is processed through API connections that do not use it for model training, and it is not used for any other purpose without written consent.

For the packages, scope and pricing, see AI translation workflow consulting. If you only need a single AI translation checked before publication, see Independent Review.