How to Build an AI Translation Workflow (Playbook)

Machine translation is old; language models made it contextual and instructable. Enterprises now translate documentation, support content, product data, and communications at volumes that make full human translation impossible and unreviewed machine output unacceptable. An AI translation workflow resolves that with terminology control, quality estimation, risk-based post-editing, and evaluation. This playbook covers the build, following FISTA's AI enablement practice. Context is in ai translation and, for the multilingual web, nextjs internationalization guide.

What does the workflow do?

Stage	Function
Intake	Content from CMS, documentation, product data, support systems
Segmentation and context	Segments with surrounding text, content type, metadata
Terminology and memory	Glossary enforcement; translation memory matches
Translation	Machine translation or language model with constraints
Quality estimation	Per-segment confidence
Routing	Automatic, sampled review, or full post-editing by risk
Post-editing	Human correction in a translation tool
Delivery	Back to content systems; memory and glossary updates
Evaluation	Human ratings, error taxonomy, post-edit distance

Step 1: Define languages, content types, and risk

Specify languages and locales, content types and their risk levels, quality thresholds per type, turnaround needs, and brand and legal constraints. Risk classification drives routing: legal and medical content differs from support articles. See how to write acceptance criteria for ai.

Step 2: Build glossaries and style guides

Establish per-language glossaries of product names, terms, and do-not-translate items, and style guides covering tone, formality, locale conventions, and formatting. These are enforced in translation and used in evaluation. Without them, quality disagreements are unresolvable.

Step 3: Prepare content with context

Segment content at sentence or paragraph level while preserving structure and placeholders; attach surrounding context, content type, and audience; detect and protect code, variables, and formatting tokens. Context is the largest quality lever for language-model translation.

Step 4: Translate with constraints

Select the engine per language pair and content type by evaluation: dedicated machine translation, a language model, or a combination. Constrain with glossaries, style instructions, translation memory matches, and structured output that preserves placeholders. Route through the gateway for cost accounting and logging. Model selection guidance is in how to choose an llm provider and gateway design in how to build an llm gateway.

Step 5: Estimate quality and route

Apply quality estimation per segment, validated against human ratings on your content. Route by rules combining risk and estimated quality: high-risk content to full post-editing regardless; medium-risk to post-editing on low estimates and sampled review otherwise; low-risk to automatic delivery with sampling. Routing is explicit and auditable. See rules engine vs llm.

Step 6: Design post-editing

Post-editors work in a translation tool showing source, machine output, glossary hits, memory matches, and quality flags, and their edits update translation memory and feed evaluation. Capture edit distance and error categories. Review design follows how to build a human review queue.

Step 7: Integrate and deliver

Connect to content sources and destinations through APIs; preserve structure and metadata; update translation memory and glossary candidates; record provenance per segment (engine, version, reviewer). Provenance supports audit and continuous improvement. Integration patterns are in api-first development.

Step 8: Evaluate

Sample by language and content type; have qualified raters apply an error taxonomy (accuracy, terminology, fluency, style, locale); track automated metrics against references where available and post-edit distance in production; validate the quality estimator against ratings; and gate engine or prompt changes on the suite. Method is in the AI evaluation and testing whitepaper.

Worked example: software documentation and support content

A software company localizes documentation and support articles into a dozen languages. Glossaries and style guides are established per language with the localization team. Documentation segments carry heading context and product area; placeholders and code are protected. Evaluation per language pair selects a language model for most pairs and a dedicated engine for two. Quality estimation routes low-confidence segments and all legal notices to post-editors, while routine support articles flow automatically with sampled review. Post-edits update translation memory, and the error taxonomy reveals a terminology inconsistency in one language traced to a glossary gap. Turnaround for new articles falls from weeks to days, post-edit distance declines as glossaries mature, and human raters confirm quality holds across sampled content.

What does it cost to run?

Cost scales with word volume and the share routed to post-editing; translation memory and quality estimation reduce both. Value is measured in turnaround, languages covered, and quality ratings against the prior process. Drivers are in generative ai cost.

What are the common mistakes?

• Translating without glossaries or style guides.
• Segment-by-segment translation with no context.
• No quality estimation, so every segment is either fully reviewed or not at all.
• Legal and medical content flowing automatically.
• Ignoring translation memory.
• Evaluating with automated scores only.

How do you phase the rollout?

Start with one language pair and one content type of moderate risk, such as support articles, where volume is high and the cost of an imperfect translation is recoverable. Build the glossary and style guide for that pair, establish the labeled evaluation sample and the quality-estimation validation, and run the workflow with post-editing on everything for the first cycles so the estimator can be calibrated against real editor judgments. Once routing rules are trusted, reduce post-editing on high-confidence segments and expand to further content types by risk order, adding languages as glossaries and raters are in place. Legal and regulatory content joins last and keeps full post-editing regardless of estimated quality.

How do you handle terminology consistency?

Maintain a term base per language with approved translations of product names, legal terms, and brand phrases, enforce it in the translation prompt and in a post-check, and route deviations to reviewers. Consistency across thousands of strings is what separates a translation workflow from a translation tool.

How FISTA Solutions builds translation workflows

FISTA Solutions builds AI translation workflows to this playbook: risk-classified specifications, glossaries and style guides, context-rich segmentation, engine selection per pair by evaluation, validated quality estimation with explicit routing, post-editing integration, provenance, and human-rated evaluation. The AI enablement practice delivers the workflow and gateway, AI agents automate the surrounding content operations, and forward deployed engineers embed with your localization team. The record behind the work is 150+ projects across 12+ countries.

To scope a translation workflow, message FISTA on WhatsApp, or read how to build an ai content pipeline for the upstream content system.