Localization News · 16 October 2026
The Infrastructure Illusion
Why 229 AI solutions still rely on human scaffolding. Intento's 2026 state of translation automation report evaluated 229 AI solutions across 100 languages and found that zero realtime one-step solutions meet standard enterprise formatting and terminology requirements.
- 229
- AI solutions evaluated
- 100
- languages tested
- 57
- models, including MT engines and promptable LLMs
- 37
- vendors behind them
The gap on your desk
Two conflicting realities, one pipeline
The vendor brochure
Just plug in the API.
Effortless, immediate multilingual deployment, as if it were a switch you flip.
The desk, every morning
- Code tags stripped out
- Document formatting broken
- JSON structures corrupted
Raw AI translation models routinely fail basic enterprise hygiene tests.
If you manage an enterprise localization pipeline, you are caught between the marketing and the operations. The root of the failure is treating localization as a purely semantic exercise. It is a strict structural engineering discipline, and it is not only about the words.
The Intento 2026 benchmark makes that vulnerability hard to argue with.
The Intento benchmark
A hundred languages, three tests
Intento evaluated 57 models from 37 vendors, covering both traditional machine translation engines and promptable large language models, across 100 languages.
Test for basic semantic quality, meaning the translated sentence generally conveys the original meaning, and 54 of the 100 languages have a one-step solution that passes, assuming a reliable backup model is in place. The friction starts when you scale to enterprise requirements. Drag the fold along the line to raise the bar.
Each tile stands for one of the 100 languages. Tiles are not assigned to particular languages and their arrangement is illustrative. Only the counts are reported figures.
in the benchmark
Select a test to see how many languages still have a passing one-step solution.
The standard enterprise requirements are the absolute baseline for commercial production: adherence to client glossaries, brand tone, valid JSON structures and correct Unicode handling. Under those conditions the success rate falls from 54 languages to 38. Add advanced requirements such as enforcing gender neutrality and only 12 languages pass.
Engines compared
Where the paper tears
On raw error counts the two engine families look closer than you might expect. Traditional machine translation engines suffer 58 weighted errors per 10,000 words. Large language models sit at 31.
The catastrophic failures happen entirely within the invisible formatting. Select any line to see why.
Fail JSON validationShare of outputs
Broken tagsShare of outputs
Weighted errorsPer 10,000 words. One mark is one error.
That is what makes the formatting layer so expensive. Consider a continuous localization pipeline pushing daily updates for a software product. A single corrupted JSON file or one stripped HTML tag breaks the client's software build. The application will not compile.
At that moment, raw semantic accuracy becomes irrelevant. It does not matter how beautifully the sentence was translated if the software itself is broken.
The orchestration harness
A protective layer around the engine
The industry response is to stop deploying raw engines. The orchestration harness is middleware wrapped around the core translation engine, handling pre-processing, prompt tuning and post-processing, because the model cannot be trusted with raw code.
Before a string reaches the large language model, the harness masks the sensitive syntax, swapping complex HTML or JSON structures for simple placeholders the model can ignore. When the translation comes back, post-processing repairs the tags and restores the code around the newly translated text.
Send an or a or .
Illustrative example of the mechanism. The strings and the failure shown are not Intento test data. Green diamonds mark where a person works on the process.
A fair question follows. If a language model fundamentally cannot hold structural integrity, and the fix is a heavy harness, automated quality estimation and a secondary model for automated repair, are we really automating efficiently? Industry data shows 60% of initially failing LLM translations only pass the benchmark after the full three-step cycle of translate, check and repair.
That can read less like streamlined innovation and more like expensive duct tape over the foundational flaws of the underlying models. The commercial data in the report puts a number on the duct tape.
What the safety net costs
Cheap land, expensive foundations
On a raw basis, large language models look almost free. They cost a median of $4.59 per million characters, against a median of $24.99 per million characters for traditional machine translation engines.
Then the required workflow arrives. A two-step solution that catches basic errors multiplies the cost by 2.5 times compared with the base model. The full three-step solution, meaning initial translation, linguistic quality assurance and an automated repair loop, has a median cost of $590 per million source characters. Pull the pile down the scale to see where the money goes.
Cost per million characters, shown on a logarithmic scale. The depth of each marker is proportional to the logarithm of the reported median.
96%
The Intento linguistic quality assurance step accounts for roughly 96% of the $590. Automated quality estimation at that level is resource intensive. The methodology deploys three separate models from different vendors to review each translated sentence, and a finding is accepted only when they reach a 90% agreement threshold.
Think of buying a cheap plot of land in a swamp. The seller will proudly tell you it was a bargain, and it was. Pouring a reinforced concrete foundation, driving steel pilings to bedrock and building the structural supports to carry a safe building is what costs a fortune.
In that picture, the orchestration harness is the surveyor. It maps where the solid ground lies, the code tags and formatting constraints, before anyone pours the translation concrete. The linguistic quality assurance loops are the steel pilings. Without the surveyor and the pilings, enterprise content sinks into the mud and fails in production.
This is the trap to warn your procurement team about. They read mainstream press releases about five-dollar translation APIs and walk into budget meetings expecting an 80% reduction in localization spend. They ignore the infrastructure needed to make cheap translation safe in a live production environment. Between the two reported medians, $4.59 and $590, the premium for safety is roughly 128 times.
Voice
Where there are no brackets
If text localization keeps failing on invisible structural code, what happens when you strip the text away? There are no brackets in human speech. xAI recently released Grok Voice 3.0, a major update to its speech-to-speech model designed for real-time voice agents. It is deployed on live customer support calls for Starlink across 20 supported languages.
The mechanism explains the speed. A speech-to-speech model removes the intermediary text generation step. It listens to the acoustic input, processes the semantic intent directly and generates an acoustic output. Drag the fold to take the text out of the path.
- 70%
- of customer support calls resolved without a human in the loop
- 8¢
- per minute, priced to compete with OpenAI's GPT Live 1 and Google's Gemini 3.8 Live
- 20
- supported languages
It is not only answering basic questions. The system runs multi-step customer service workflows: it verifies customer identity, runs troubleshooting protocols for the physical satellite hardware, processes account credits and initiates shipping for replacement equipment.
Voice AI scales at this pace because it bypasses the structural problems of written pipelines. A voice agent does not preserve HTML character entities, parse nested code or protect proprietary software syntax. Its main requirements are acoustic fluency and conversational logic.
The 70% resolution rate still deserves professional skepticism. Twenty languages in a controlled call center is a small footprint beside the 100 written languages Intento benchmarked, and the languages in that 20 have vast volumes of training data online. The real test is acoustic accuracy on regional dialects, local accents and technical hardware terminology, especially at the outer edges of those 20 languages.
A dropped syllable during a troubleshooting step can frustrate a caller as fast as a broken HTML tag breaks a build, maybe faster, because the user is waiting on the phone. If the AI hallucinates a step while you are fixing your satellite dish, you will ask for a human. The tolerance for latency and misunderstanding in live audio is low.
The big technology companies concentrate on the 20 most profitable languages because that is where the return on their enormous compute investment is. Building speech datasets for regional dialects is expensive and slow.
Language data
Who builds the data when the market will not
Public institutions cannot pick the top 20 and ignore everyone else. They are obliged to solve localization for the long tail of low-resource languages, for citizen equity, accessibility and strict legal compliance. Select a branch to follow where each approach gets its data.
Commercial tech giants focus on the 20 most profitable languages, because that is where the immediate return on their compute investment sits. Building speech datasets for regional dialects is expensive and time-consuming.
Commercial language service providers also did not see immediate demand for languages such as Manipuri and Tangkhul. It did not justify the cost of building the first datasets.
The sector often treats human translation data as a cheap commodity to scrape from the open web. That can lead to training sets contaminated with poor-quality machine translation, which means training AI on bad AI output.
The Commission's internal eTranslation platform surpassed 1 billion translation requests this year. It supports all 24 official EU languages.
Because it handles sensitive government communications, legal drafts and public procurement, it cannot depend on commercial cloud APIs. Sending unreleased legal directives to an external vendor's server would be an unacceptable security risk. The Commission guarantees strict data security and non-retention policies in line with the EU AI Act and GDPR, using custom neural machine translation models hosted on its own infrastructure: sovereign AI.
The models are trained on decades of high-grade, human-curated translations produced by the Directorate-General for Translation, the DGT. That parallel data gives precise domain adaptation for European legal and administrative terminology, which generic models trained on open web data do not have. The Commission treated its translation memory as a sovereign asset, and built an engine that outperforms generic models on specialized institutional content without risking data leakage or copyright infringement.
The Digital India BHASHINI Division recently signed a memorandum of understanding with the state of Manipur. The aim is to build digital language resources for local indigenous languages such as Manipuri and Tangkhul, languages the big commercial vendors have historically ignored.
Web scraping is not an option, because the digital footprint of these languages is too small. There are not enough Manipuri websites to train a model. Instead, the project relies on crowdsourcing spoken and written dialect data directly from local communities, educational institutions and native speakers, through the BHASHINI Mitra initiative.
That grassroots data is being used to build real-time English-to-Manipuri translation, automatic speech recognition and text-to-speech tools. The models feed citizen-facing public portals for essential healthcare information, educational resources and legal services.
It marks a shift in how the industry acquires language resources. State-sponsored data collection is building the parallel corpora that will eventually open these regions to broader digital services, creating digital inclusion from the ground up.
But gathering, validating and structuring dialect data before a model can begin training takes an extensive logistical effort, and it takes human expertise. Which brings the story back to the people running these systems.
The people in the loop
A vacancy notice shows where the job is going
Amid the hype about fully automated workflows and autonomous agentic AI, the world's largest in-house translation service is hiring human linguists. The EU Directorate-General for Translation recently opened recruitment for a Swedish translator based in Luxembourg and an Irish translator based in Brussels.
The job title is the story. It is not simply translator. It is Language Technology Coordinator. The explicit duties are testing, deploying and raising awareness of translation applications and AI tools across the whole translation department. The coordinator rigorously evaluates new language models, assesses their output quality and safely integrates them into the daily work of the human staff.
In the photograph, the paper structure on the desk stands for the orchestration harness. Select a point on it to see where a person works.
That vacancy proves that automation is not eliminating linguistic roles in complex enterprise environments. It is elevating them. A professional linguist is becoming a workflow architect and an AI governor, the person configuring the orchestration harness.
You cannot automate the oversight of an automated system. Someone has to interpret the error metrics, and someone has to understand both the linguistic nuance of the target language and the rigid technical constraints of the deployment pipeline. The job has never been more complex.
The value of a localization professional is no longer bilingual fluency alone. It is structural orchestration, evaluating quality metrics and safeguarding brand safety inside complex AI pipelines.