Skip to content
FAS Localize

News & views

How Many Steps Does AI Translation Need?

Dao HuyBy , Senior Project Manager5 min readFacts checked Nov 16, 2026
A person's hands on a laptop keyboard with code on screen, standing in for an automated AI translation pipeline step.
Photo by Christina Morillo on Pexels

How many steps does AI translation need? Intento's 2026 benchmark of 229 solutions in 100 languages found the answer depends on the quality bar. A one-step solution plus a qualifying backup meets basic quality in 54 languages. Standard and advanced quality usually need a post-edit or a quality-assessment-and-repair step added on top.

How Many Steps Does AI Translation Need at Each Quality Tier?

Fewer languages can rely on a single translation step as the quality bar rises. For basic translation quality, 54 of the 100 languages have a one-step solution plus a qualifying backup model. That drops to 38 languages for standard quality. Only 12 languages reach advanced quality with one step.

Quality tierLanguages that pass with one step
Basic54 of 100
Standard38 of 100
Advanced12 of 100
One-step-solution coverage by quality tier, Intento 2026

FAS Localize (faslocalize.com) runs its own version of that same idea day to day. Its AI localization pipeline uses automated checks plus a QE score to decide the route. High-confidence segments auto-approve. Only uncertain ones reach a native linguist.

What Did Intento's 2026 Report Actually Test?

Intento published The State of Translation Automation 2026, an interactive report that evaluated 229 solutions. Those solutions were built from 57 AI models and machine translation engines, translating from English into 100 languages. The 57 models came from 37 vendors, including Alibaba, Amazon, Anthropic, Cohere, DeepL, DeepSeek, Google, Microsoft, Mistral, OpenAI, Tencent and xAI.

The team built the test set between June and September 2026. Intento's language engineers translated four English business texts, totaling 2,978 words, into 100 languages. They then assessed 10 requirements across three quality levels: Basic, Standard and Advanced.

Does Post-Editing Alone Fix Translation Quality?

Not by much on its own. Intento found that automatic post-editing alone, without identifying specific errors first, reduces translation errors by only 7%. Among the evaluated LLM solutions that initially missed the general quality bar, 24% met it after translation plus post-editing. 60% met it after translation, quality assessment and repair instead.

The gap comes down to the quality-assessment step. A model that only rewrites a draft is guessing at what to fix. A model that scores the draft first knows exactly where to spend the repair pass.

Why Does a Quality-Assessment-and-Repair Step Cost So Much More?

Adding a dedicated quality-assessment step plus targeted repair cuts translation errors by 54%. It also raises the cost. A workflow that adds that assessment and repair step costs a median of about $590 per million source characters. Quality assessment itself accounts for most of that added cost.

Even one-step solutions vary widely before any extra step is added. Intento's report puts the median cost per million source characters at $4.87 with LLMs, $10 with traditional MT engines and $25 with LLM-based MT engines. Error rates also vary: about 1.9 times between different models on the same harness, and about 2.4 times between different harnesses on the same model.

Is Real-Time Translation Fast Enough Without Extra Steps?

Usually not at the standard tier. Intento's report finds that one-step real-time solutions do not meet every Standard requirement for accuracy, terminology, tone and formatting. Traditional MT engines translate in a median of about 2 seconds per 1,000 source words. LLM-based solutions take about 66 seconds instead, too slow for a live chat widget to also run a quality check mid-reply.

Do Low-Resource Languages Need More Translation Steps?

Almost always, yes. In 10 of the 11 low-resource languages Intento assessed, even basic translation quality required additional review and correction steps. Separate academic research on quality estimation makes a related point: QE is described as essential for assessing machine translation quality without a reference translation, particularly in domain-specific and low-resource language scenarios.

How Many Steps Should a Team Actually Plan For?

Match the number of steps to the quality bar, not to a single model's reputation. See also our related post: LLM judge versus COMET for translation.

Match the content type to the pipeline depth

  • Internal or gisting content: one draft is usually enough.
  • Customer-facing copy: add a post-edit pass to catch tone and terminology misses.
  • Regulated, legal or high-stakes content: add quality assessment and repair, the step Intento's data shows actually moves the error rate.
  • Real-time chat or live captions: budget for basic quality only, since a review step adds too much latency.

Teams that want that routing logic without building it themselves can look at FAS Localize's AI localization pipeline. Automated checks plus a QE score decide the route. Native linguists verify only the uncertain segments, and results sync back into the client's repository, CMS or API. Every FAS Localize job still gets a fixed quote before any work starts, so an extra step never arrives as a surprise line item.

Frequently asked questions

How many languages can pass with just one AI translation step?

54 languages, at the basic quality tier. In Intento's 2026 benchmark of 229 solutions across 100 languages, 54 languages meet basic quality with a single translation step plus a qualifying backup model. That number drops to 38 languages for standard quality and 12 for advanced quality.

Does post-editing alone fix AI translation quality?

Only partly. The report found automatic post-editing without a specific error list cuts translation errors by just 7%. Adding a dedicated quality-assessment step before repair cuts errors by 54% instead.

Why is a QE-and-repair pipeline so much more expensive?

It adds a real cost per word. A full workflow that adds quality assessment and repair to the translation costs a median of about $590 per million source characters. Quality assessment itself accounts for most of that total cost, Intento's 2026 report found.

Can real-time AI translation skip the extra steps?

Mostly, by necessity. One-step real-time solutions don't meet every standard requirement. MT engines translate in a median of about 2 seconds per 1,000 words, versus about 66 seconds for LLM-based solutions, leaving little time for a review step.

Do low-resource languages need more pipeline steps?

Yes, almost always. In 10 of the 11 low-resource languages Intento assessed, even basic translation quality needed additional review and correction steps.

Sources

  1. Intento Publishes The State of Translation Automation 2026, an Interactive Benchmark of 229 AI Translation Solutions in 100 Languages, Intento. Accessed Nov 16, 2026.
  2. The State of Translation Automation 2026, Intento. Accessed Nov 16, 2026.
  3. Domain-Specific Quality Estimation for Machine Translation in Low-Resource Scenarios, arXiv. Accessed Nov 16, 2026.
  4. Vietnamese Localization Pipeline: AI + Native QA, FAS Localize. Accessed Nov 16, 2026.
  5. Vietnamese Translation Rates & Localization Costs, FAS Localize. Accessed Nov 16, 2026.

Want localization to run on autopilot?

Talk to us about your pipeline

Dao Huy

Senior Project Manager · EN · ZH · FR to Vietnamese

Your point of contact from brief to delivery: he scopes the job, sets up the AI and native-review workflow, and keeps quality, deadlines and budget on track.