GDPR-compliant organization
ISO 27001 certified organization
SOC II compliant organization
Smartcat routes your content through multiple machine translation engines and large language models, applies your glossary and translation memory to every job, and sends anything that needs human judgment to a reviewer.
Trusted by 1,000+ enterprise brands to power global content
Trusted by:
I want to translate from
to
Time of Smartcat usage
3 months
Translation accuracy
95%Smartcat’s AI translation software chooses the best algorithm for your language pair, learns from your edits, and gets better the more you use it.
Want to know how it works? Talk to our solutions expert.
Book a demo
Machine translation is the automated conversion of text from one language into another — today almost always neural machine translation (NMT), which learns from large volumes of previously translated text rather than from hand-written rules.
AI translation is broader. It adds large language models and AI agents that read context, apply your approved terminology and learn from your corrections. Smartcat combines both: machine translation for speed, adaptive AI for quality that improves as you use it.
Raw output still breaks in predictable places. Here is where — and what to do about each.
Terminology drifts between files
The same product name comes back three different ways across three documents, and nobody notices until a reviewer in-market flags it.
The same correction gets made five times
A fix applied in one project does not carry to the next, so reviewers re-edit the identical sentence in every new file.
Quality is uneven, and you find out late
An engine that handles English–Spanish well can fall apart on Japanese or Finnish. Without a measure, the discovery happens in review — or after publication.
Nobody can say what is safe to ship raw
The real question is not “is machine translation good?” but “which of my content can go out without a human pass?” Most teams have no evidence either way.
Idioms and marketing copy
Marketing copy is where machine translation most visibly fails. Idioms, wordplay and taglines get translated literally, because the model optimises for the likeliest rendering of the words rather than the intent behind them.
Run MT on marketing content to get a draft and a cost baseline, then budget a human pass for anything carrying the brand voice. Product descriptions and feature copy usually survive; headlines and campaign lines usually do not.
Legal and medical terminology
Legal and medical content is translatable by machine, but the cost of an error is asymmetric. Lock terminology in a glossary before the first run, and have a qualified reviewer sign off the output.
Smith+Nephew cut reviewer editing workload on regulated content by 70% using translation memory and glossaries — the review still happened, it just got faster. Do not publish regulated content from raw machine translation.
UI strings and placeholders
Software strings break machine translation in two ways. Placeholders like %s or {count} can be reordered or translated if the engine does not recognise them, and strings arrive without surrounding context, so “Open” can become a verb or an adjective depending on the guess.
Define your variable formats as custom placeholders so they are carried through untranslated, and check every placeholder is present in the target. Text expansion is the third issue — German and French often overflow fixed-width UI.
Brand names, numbers and tables
Set brand names, product names and model numbers to Do Not Translate in your glossary before any machine translation run — the setting keeps the source term as-is and excludes it from AI translation.
Numbers carry their own trap: decimal separators, date order and units differ by locale. Tables translate cell by cell, so text that fitted a column in English often does not after expansion.
| Type | How it works | Strongest at | Where it breaks |
|---|---|---|---|
| Rule-based (RBMT) | Hand-written grammar rules and dictionaries | Narrow, controlled domains with fixed terminology | Cannot generalise beyond its rules; output is rigid |
| Statistical (SMT) | Learns word and phrase probabilities from bilingual corpora | Broad coverage where data exists | Optimises fragments, so sentences read disjointed |
| Neural (NMT) | Encodes the whole sentence, then generates the target | General-purpose translation — what “machine translation” means today | Quality swings sharply by language pair; confident-sounding errors |
| LLM-assisted / adaptive | Adds context, your glossary and past corrections to the model's input | Terminology consistency and content that must sound like you | Slower and costlier per word than raw NMT |
| What it is | What it does | Who it's for | |
|---|---|---|---|
| Machine translation | An engine that generates translation automatically | Produces a first draft in seconds | Anyone needing speed or volume |
| CAT tool | The workspace a human translator works in | Segments files, applies translation memory and glossaries, tracks changes | Translators and reviewers post-editing MT output |
| TMS | The system that manages the work around the files | Assigns tasks, tracks deadlines, routes approvals, reports | Localization managers running many projects |
Most professional workflows run machine translation inside a CAT tool, with a TMS coordinating the work. Smartcat includes all three — see CAT tool.
There isn't one. Quality varies by language pair and content type, and the gap between two engines on the same text is usually wider than the gap between two versions of the same engine. An engine that wins on your German marketing copy can lose on your Japanese support articles.
So Smartcat routes across multiple engines and models instead of locking you into one vendor's output. An AI Translation Profile sets which engine handles each language pair, plus a fallback engine if the primary one fails. Switching is a setting, not a migration. See AI Translation Profiles.
Run multiple engines from one platform
Route each language pair to the engine that handles it best, without a separate contract, console or invoice per MT vendor.
Let engine choice follow the content
Profiles set the engine per language pair, with a fallback engine, and can be created per client, department or brand and applied to new projects automatically.
Train output on your own glossary
Approved terms and past translations are applied before the engine runs, so your terminology is in the first draft instead of being added in review.
Know which segments are safe to publish
Translation Quality Score (TQS) scores each AI-translated segment. High-scoring segments can be confirmed without review; low-scoring ones go to your team or a Marketplace reviewer.
92%
edit-free words
In 2026, 92% of words translated by Smartcat AI for paying customers were accepted without a single human edit.
280+
languages
available the moment you sign up
80+
file formats
supported for translation by AI algorithms
500K+
reviewers
providing high-quality translation services
70%
less reviewer editing on regulated content
We love having total visibility over our entire translation lifecycle on our Smartcat workspace due to the centralized nature of the platform and being able to gather all the linguistic assets we need, including translation memories and glossaries.
Start free trial
Barbara Fedorowicz
Translation department manager, Smith+Nephew
$1.20
per 1,000 words, down from $200–$300
Smartcat’s AI translation is a fantastic tool. Our ability to source content from a central repository and benefit from adaptive algorithms which incorporate translation memories and glossaries across all project teams means the business as a whole gains from each team’s translations.
Start free trial
Senior Manager
Stanley Black & Decker
for ease of setup
ease of use
company
compliant organization
Upload a real file, pick a language pair, and compare the export against what you ship today.
If you want to understand one paragraph in a language you do not read, this is the wrong tool — open a free text-box translator; it will be faster and the answer will be just as good.
Smartcat earns its keep when translation is a repeated process rather than a one-off: many files, many languages, terminology that has to stay consistent, and someone accountable for whether the output is publishable. Below that threshold it is overhead.
Pricing is in Smartwords, and reuse lowers the bill as your translation memory grows. Stanley Black & Decker went from $200–$300 to $1.20 per 1,000 words. See pricing.
SOC 2 Type II compliant. Encrypted in transit and at rest, isolated workspaces, role-based access, and SSO via Azure AD, Okta and ADFS on higher tiers.
They translate text, not a workflow. Neither applies your glossary and translation memory across every file, keeps your corrections for the next job, or hands work to a reviewer.
One engine is a bet on one vendor being best for every language pair and content type you publish. Routing removes the bet — and switching is a setting, not a migration.
A 1:1 consultation with a Smartcat specialist. No form queue.
When the cost of an error is higher than the cost of the review. Regulated content, contracts, safety instructions and anything carrying a claim get a human pass; internal documentation, support archives and high-volume product data usually do not.
Every machine-translated project can go through post-editing in the Smartcat Editor — reviewers see source and target side by side, and their edits feed back into your translation memories. No in-house reviewer for a language? Hire a vetted post-editor from the Marketplace without leaving the platform.
Full guide: machine translation post-editing (MTPE) · As a service: MT post-editing services
English into German, French, Spanish, Italian and Dutch. These pairs have the most training data, so raw output is often publishable for factual content after a light check. German and French also need a formal/informal choice — du vs Sie, tu vs vous — set in the Translator's system prompt. That works with LLM engines only, not Google NMT, DeepL or Microsoft Translator.
Japanese, Chinese and Korean. Segmentation and line-breaking differ, and specialised fonts may be substituted on export — check output against your brand fonts. Vertical Japanese (tategaki) and vertical Chinese columns are not supported; route that work to a human translator.
Arabic and Hebrew in image translation. Direction detection and mixed bidirectional ordering are supported, and text is re-rendered right-to-left. Watch the layout: mirrored interfaces, alignment, and punctuation at line ends.
Documents. 80+ formats, including DOCX, PPTX, XLSX, PDF and XLIFF 1.2/2.0, exported in the original format. PPTX speaker notes are skipped unless you enable Load notes at upload, and Excel formulas are not preserved — only cell text is translated.
API and strings. For UI strings, product data and CMS content, generate a key in Settings → API and call the same endpoints a paid account uses. API access is included in the 15-day trial and continues on Anticipate and Autonomous.
Websites. A JavaScript snippet by default, plus an optional SEO proxy that serves crawlers. hreflang is generated automatically, and the proxy never sits in front of your visitors.
The best machine translation software is the one that fits your language pairs, formats and quality bar — there is no single winner. Compare engine choice, glossary and translation-memory support, and post-editing.
Fit depends on the team. Small teams without IT need no engine setup or per-engine contracts; e-commerce and SaaS teams need CMS and PIM integrations; enterprises need per-brand profiles and SOC 2 Type II. Smartcat covers all three — for one-off paragraphs, a free translator is better.
Yes. Google Translate, DeepL's free tier and similar tools translate text and some documents at no cost, with limits on volume, file size and confidentiality. What they do not give you is your terminology applied automatically, a translation memory that lowers cost over time, or a review workflow. Smartcat's trial is 15,000 Smartwords over 15 days, no credit card. There is no unlimited free Smartcat plan.
Upload the file, choose the source and target languages, and download the result in its original format. In Smartcat, your glossary and translation memory are applied before the engine runs, and your AI Translation Profile decides which engine handles that language pair. Review the output in the editor, then export. For step-by-step document workflows, see document translation.
Accuracy depends on the language pair, the content type and how much of your own data the engine can use, so a single number misleads. In 2026, 92% of words translated by Smartcat AI for paying customers were accepted without a single human edit. It is strongest on high-resource languages and factual text, weakest on idiom and context-free strings. Translation Quality Score then rates each segment of your file.
Some of it can — the real question is which segments, not which tool. Smartcat's Translation Quality Score rates every AI-translated segment, so you can sign off the strong ones as they are and send only the weak ones to your own reviewer or a Marketplace specialist. Anything regulated or legally binding still deserves a human read, whatever its score.
It depends on the provider. Free consumer tools may keep what you paste in, which is why many legal and security teams rule them out for confidential material. Before uploading anything, check certification, encryption, workspace isolation and access control. Smartcat holds SOC 2 Type II, applies encryption at rest and in transit, keeps each customer's workspace separate, and assigns access by user role.
It depends on who you pay. Engine APIs such as Google, DeepL and Microsoft typically bill per character; agencies bill per word, often with minimum fees; platforms bill per word, per seat or by subscription. The bigger lever is reuse — content you have translated before should not cost full price again. Smartcat meters Smartwords: one per source word per target language.
Machine translation converts text between languages automatically. Modern systems are neural: the model encodes a whole sentence, then generates the target sentence, so word order and agreement hold together rather than being stitched from fragments.
Quality comes from the data the model saw in training — which is why the same engine can be strong on English–German and weak on English–Finnish. Smartcat adds your glossary and past translations to that input, which improves it further.
A machine translation engine produces the draft; a CAT tool is where a person checks and corrects it, with translation memory, glossaries and a record of every change. You need the engine for speed and the CAT tool for control. In Smartcat the two are one step — MT output opens in the editor with your terminology already applied. See CAT tool.
Yes. Your glossary and translation memory are applied before and during translation, so approved terms and previously approved sentences come through rather than being re-guessed. Every edit a reviewer approves is written back, which is what makes output improve with use. This is the main reason quality diverges between two teams using the same underlying engine.
Yes. Create an AI Translation Profile for each client, department or brand, give it only the translation memories and glossaries it should use, and apply it when you set up projects. By default, new translation memories are added to every profile — switch that off if clients must stay separate. Profiles are specific to one workspace and can't be shared across workspaces.
Both. Smartcat is the platform that runs machine translation with your glossaries, translation memory and engine choice. For the human side, you can hire vetted linguists and post-editors through the Smartcat Marketplace in the same workspace, or buy MT post-editing as a service. Most teams combine the two: machine translation for volume, people for whatever carries risk.
No. 100%+ translation memory matches consume zero Smartwords on all current plans, so repeat and near-repeat content costs nothing to process again. This is why per-word cost falls as your translation memory grows, and why teams with large memories see the biggest gap between list price and what they actually pay.
All the common ones — Word, PowerPoint, Excel, PDF and XLIFF — among more than 80 formats, including XLIFF exported from Articulate Rise, Storyline and Easygenerator. Scanned and image-heavy PDFs are OCR'd automatically, and standalone images go through image translation at 1,000 Smartwords per image. Bilingual DOCX with parallel columns is export-only: it can't be re-imported after editing.
Under the Data Processing Agreement, on written request Smartcat securely destroys or returns all uploaded content and personal data within a maximum of 30 days, unless law requires otherwise. Deletion is not automatic at contract end — the written request starts the clock. Workspaces remain isolated and access role-based for as long as the account is open.
Book a demo — a 1:1 consultation with a Smartcat specialist, no commitment, no form queue.
Smartcat. Smith+Nephew: reviewer editing workload down 70%, turnaround from 10 days to two–three days.
Smartcat. Stanley Black & Decker: translation costs down up to 70%, from $200–$300 to $1.20 per 1,000 words.
Smartcat Help Center. AI Translation Profiles: configuring linguistic assets and translation engines.
Smartcat Help Center. Translation Quality Score (TQS).
Smartcat Help Center. Supported file formats.
Smartcat Help Center. Glossary properties and structure.
Smartcat Help Center. Add placeholders.
Last updated: 29 September 2026