Translation can make a sentence readable. Localization has to make an entire system work.
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
Wissen & Sprache · Gewerkton field lessons
Translating words is not the same as localizing a system.
Shipping a voice-first construction documentation platform in 27 languages exposed the failures hidden behind fluent text: invented legal meaning, punctuation errors, broken interfaces and evidence flattened into prose.
linguistic realities
A multilingual defect, instruction or daywork report must remain operational across trades and languages while its original evidence stays unambiguous.
where fluent errors survive
The difficult last ten percent includes drifting terms, unsupported legal implications, wrong punctuation and compact labels expanded into sentences.
Fluency became camouflage for invention
A translator introduced plausible legal terminology that the source had never claimed. The lesson: missing information must remain missing, and translated text must stay constrained by the defined role of each product term.
A hyphen is product behavior
Malay and Indonesian reduplication showed why related languages cannot be treated as interchangeable. Punctuation changes meaning, search, tokenisation and display—and a global cleanup rule can make text less correct.
AI providers, not 13 shortcuts
BYO-AI spans providers and selectable processing in the EU, the US and Asia, including mainland China. Regional choice reduces lock-in; it does not replace language engineering.
Structured evidence beats polished prose
Localization holds when observation, interpretation and action remain separate. Concise interface text can point to a richer record without forcing the whole spoken account into a label.
That distinction becomes painfully clear when a product moves beyond a handful of languages. Gewerkton, a voice-first construction documentation and defect management platform, currently carries content in 27 languages. The work has exposed problems that rarely appear in a translation spreadsheet: a translator quietly inventing legal terminology, reduplication hyphens behaving differently in Malay and Indonesian, character budgets breaking interfaces, and large language models producing text that looks finished while preserving subtle errors.
These are not merely linguistic curiosities. Construction documentation is evidence. A note may need to retain who recorded it, what it concerns, the original audio, a photograph, a deadline or a signature. If that structure dissolves into free-form prose, even elegant translation cannot restore what was lost.
The resulting lesson is simple: translating words is one layer of the job. Localizing a documentation system means designing language, interfaces, data structures and regional infrastructure together.
multilingual translation management software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
One product, 27 linguistic realities
Gewerkton was born in the German market and has its deepest commercial integration there, including GAEB, REB, XRechnung and DATEV. Yet its intended markets are global. Its 27 content languages sit alongside a choice of regional AI providers in the EU, the US and Asia, including mainland China.
That combination matters because language is never an isolated presentation layer. It intersects with the way a team captures information, the place where data is processed, the form in which evidence is retained and the tools through which people act on it.
A multilingual menu is comparatively easy. A multilingual workflow is harder. A defect recorded in one language may need to be understood by another trade in another language, while the original evidence remains unambiguous. A dictated daywork report must still behave like a report rather than becoming a polished paragraph with its operational fields blurred together. An instruction must remain connected to its original audio rather than surviving only as translated prose.
This is why “supports 27 languages” is less a statement about vocabulary than a demand placed on the product’s architecture.
As an affiliate, we earn on qualifying purchases.
The KVKK incident: when fluency conceals invention
One of the most instructive failures was quiet. A translator introduced legal terminology in connection with KVKK. The wording looked plausible. That was precisely the problem.
A visibly broken sentence invites review. An invented term delivered in a fluent, confident register may pass unnoticed because it resembles the language readers expect from formal material. The error is not necessarily grammatical. It is epistemic: the text appears to know something that its source never said.
This risk is especially acute when translators or language models encounter material that sounds legal, regulatory or contractual. They may try to complete the pattern. A phrase that seems incomplete is expanded. A familiar concept is substituted for a narrower one. A label acquires implications absent from the original. Fluency becomes camouflage.
The remedy is not simply to demand more literal translation. Excessive literalness creates different failures. Instead, the source has to constrain what the translated version is allowed to claim. Terms need to remain attached to their defined role in the product. Missing information must remain missing rather than being filled with a plausible invention.
This is also a documentation lesson. A structured record distinguishes captured evidence from later interpretation. It can preserve an original audio file while allowing a transcript, translation and task assignment to exist as separate layers. Prose notes tend to collapse those layers. Once observation, inference and action are merged into one paragraph, translation has no reliable boundary to respect.
As an affiliate, we earn on qualifying purchases.
Malay, Indonesian and the importance of a hyphen
Malay and Indonesian reveal a different class of problem. Both use reduplication, and hyphen handling can look like a minor typographic question. In a multilingual product, it is not minor at all.
Punctuation participates in meaning, search, tokenisation and display. A hyphen may determine whether a repeated form is recognised as expected or treated as two fragments. An automated rule that appears harmless in one language can distort another. A global cleanup that standardises punctuation may therefore make the text less correct.
The broader point is that related languages cannot safely be treated as interchangeable variants. Similar vocabulary does not guarantee identical editorial rules. Nor does a shared interface justify applying one language’s preferences to another. Localization begins where convenient assumptions stop.
These details also demonstrate why language quality cannot be judged solely by reading isolated sentences. A phrase might be acceptable in a document but fail inside a compact label. It might read correctly yet sort, wrap or search incorrectly. It might fit on a desktop and become unusable on a device held on site. Linguistic review and product review have to meet in the same workflow.
voice-first documentation platform
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Character budgets are part of meaning
Interfaces impose limits that ordinary prose does not. A short source label may expand dramatically after translation. A button that fits neatly in one language can wrap in another, push neighbouring controls aside or obscure the relationship between a field and its value.
This cannot be solved responsibly by telling every translator to “make it shorter.” Compression changes tone and can remove distinctions. Abbreviations may be obvious to one audience and opaque to another. Even when all the words technically fit, the visual hierarchy can fail: the least important phrase dominates the screen while the action a site worker needs becomes harder to find.
Character budgets should therefore be treated as product constraints, not last-minute editorial requests. The language team needs to know where text appears and what job it performs. The interface needs enough flexibility to accommodate linguistic variation. And reviewers need to inspect meaning in context, not approve strings in isolation.
The issue is particularly sharp for a voice-first system. Speech allows a person to capture more detail than they would normally type on site. The product must then turn that input into usable evidence, defects, reports or tasks without forcing the full spoken account into every visible label. The answer is structure: concise interface text can point to a richer record without pretending to contain the whole record.
Why an LLM misses the last ten percent
“Just run it through an LLM” is attractive because the first results often arrive quickly and read smoothly. For broad coverage across 27 languages, that speed is useful. It is not the same as completion.
The difficult last ten percent contains the defects most likely to survive a casual reading. Terminology may drift between screens. A formal phrase may acquire an unsupported legal implication. Punctuation may follow the wrong language’s convention. A compact label may become a sentence. Two product concepts may be translated as if they were synonyms. The output remains fluent throughout.
Large language models are particularly persuasive when they repair apparent gaps. Yet localization sometimes requires the opposite discipline: preserve the gap, retain the source term or flag an uncertainty. A model optimised to produce a satisfying answer may silently resolve ambiguity that the product team needed to see.
This does not make language models useless. Gewerkton itself is built around a broad AI strategy: BYO-AI support covers 13 providers, customers can bring their own keys, and the processing region can be selected across the EU, the US and Asia, including mainland China. The aim is to avoid vendor lock-in while allowing regional choice.
But provider choice does not remove the need for language engineering. Models can generate, transform and organise language. They do not eliminate the product-level work of defining terms, protecting source meaning, testing layouts and keeping derived text connected to evidence.
Structured evidence beats prose notes
The most important conclusion reaches beyond localization. Construction records are stronger when information is captured as structured evidence rather than buried in narrative notes.
A prose note can mention a defect, a photograph, a deadline and an instruction in one continuous passage. A reader may understand it, but software has to infer which words represent the issue, which identify the action and which establish the timing. Every translation introduces another opportunity for those relationships to blur.
A structured record keeps the components distinct. The original evidence can remain intact. A translation can help another crew understand it without becoming a replacement for the source. A decision from a meeting can become a trade-sorted task list rather than remaining trapped in minutes. A dictated report can retain its report structure. An instruction can remain backed by its original audio.
This is the logic behind Gewerkton Field, the voice-first construction site app. Dictation becomes evidence, defects, daywork reports, takt information and portal activity. In housing and building construction, defects can carry a photograph and deadline, daywork reports can be dictated, and handover can include a signature on the device.
For infrastructure and tunnels, where durations are long and change orders numerous, original audio can support instructions. On wind farms and renewable-energy projects, distributed sites and rotating crews make offline capture important in dead zones. For data centres and industrial plants, many trades work in parallel under tight deadlines, so meeting decisions can be reorganised into trade-sorted tasks.
Across borders, EU, US and APAC teams may work on the same project in their own languages while the original evidence remains unambiguous. On Asian projects, Chinese, Korean and Vietnamese crews can work multilingually from capture through reporting, with data residency selected by region.
The model is another language
Documentation is not only verbal. Plans and models provide a spatial grammar: they establish where an observation belongs and how it relates to the work around it.
Gewerkton Studio is the browser workspace for plans and models. Where no model exists, the site team can create one in the browser. That matters because location should not be left as an ambiguous phrase if it can instead be connected to a shared project representation.
A sentence such as a site observation may be perfectly translated and still be operationally weak if nobody can identify where it applies. Spatial structure reduces the burden placed on language. Translation can explain the observation; the model can anchor it.
Gewerkton Cloud coordinates operations and model data between Field, Studio and third parties. The three product lines form one branded house: capture in the field, work with plans and models in the browser, and coordinate the movement of operational and model data.
Localization includes infrastructure and privacy
A language may be available everywhere while its acceptable processing location is not. For that reason, localization also involves regional AI-provider choice and data residency. Gewerkton offers an EU cloud or deployment on the customer’s own infrastructure. The choice belongs to the customer.
The marketing site applies a similarly deliberate approach. It serves 27 languages with zero trackers, no cookie banner and a fully egress-free architecture. Its media bank contains more than 51 self-produced clips and posters.
These details are easy to categorise as separate technical or publishing decisions. In practice, they shape whether a global product feels coherent. Language, evidence, processing region, interface behaviour and privacy architecture all reach the user at once.
A beta built through unusual engineering
Gewerkton is in beta now, with a public beta planned for fall 2026. That status should be stated without euphemism because the product and its localization systems are still being tested and refined.
The development method is unusual. A solo founder directs a fleet of coding agents using Codex and Claude. In one night, that fleet shipped 21 software packages, verified with negative controls and mutation tests. The episode shows how much implementation can be accelerated, but it does not make verification optional. If anything, greater production speed increases the importance of systematic tests and explicit constraints.
Localization follows the same principle. Generating 27 versions is not the finish line. The work is complete only when unsupported claims are caught, language-specific forms survive automation, interfaces accommodate expansion and the translated layer remains tied to the evidence beneath it.
What shipping in 27 languages teaches
The field report can be reduced to a few durable lessons.
- Fluency is not proof of accuracy. Plausible legal terminology can still be invented.
- Closely related languages require separate editorial judgement, including punctuation and reduplication rules.
- Character limits are part of localization because broken layouts change how information is understood and used.
- Large language models accelerate broad coverage, but the last ten percent requires terminology control, contextual review and testing.
- Original evidence should remain distinct from transcripts, translations, summaries and assigned actions.
- Structured records travel across languages more reliably than prose notes because their relationships do not depend entirely on wording.
The governing idea is captured by Gewerkton’s own line: “On site, what counts is what’s proven.” Translation can make that proof understandable to more people. Localization ensures the surrounding system does not weaken it along the way.
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
