AI translation

Translator Coworkers in Smartcat

Zachary Ralston Updated September 02, 2026 8 min read

Overview

The Translator Coworker (formerly called Translation Agent) is the AI Coworker in Smartcat that translates your content. It's built on top of a large language model (LLM), such as Google Gemini, OpenAI's latest GPT model, or Anthropic's Claude, and you configure it with a custom system prompt that teaches it about your business, brand voice, audience, and terminology. Unlike a generic AI translation engine running on defaults, a configured Translator Coworker produces output that reads like it was written natively for each target market, rather than translated.

Each workspace has its own Translator Coworker, and you can teach it translations skills to create specialized workflows. For example, you can create skills that use different underlying translation models (Gemini, GPT, Claude), or several specialized skills for different content types (Marketing vs. Technical Documentation vs. Legal), because different models and different prompts perform differently depending on content type, target language, and audience.

When to use it

Use your Translator Coworker (rather than the default AI translation engine) when:

You want translation output to consistently match your brand voice, tone, and audience without heavy, manual post-editing.
Different content types you translate (for example marketing copy vs. legal terms) need different voice, tone, or terminology handling.

How it works

No single LLM performs best across every language and content type:

OpenAI's GPT models tend to perform strongly on marketing and brand-voice content, where creative, persuasive phrasing matters.
Anthropic's Claude (Opus) tends to perform strongly on technical, formatting-heavy, and structurally complex content (long documents, nested formatting, dense instructional text).
Google Gemini is frequently competitive across general content and specific languages.
For some less-represented or lower-resource languages, a non-LLM engine such as DeepL can outperform all three LLMs.

Because performance varies by language and content type rather than by vendor reputation, the Translator Coworker lets you identify — empirically, with your own content — which model produces the best results for each language, and lock that in as your default, rather than guessing which model is superior.

The goal is a steady state where your human reviewers need to make as few edits as possible to the AI output, maximizing translation quality while minimizing review cost and turnaround time.

How to set it up

📌 Adding a translation skill to the Translator Coworker requires the Administrator role. If you’re unable to add a skill, speak to an administrator on your account or your Smartcat contact.

In Intelligence Fabric → Smartcat Coworkers → Translator → Translation Skills, create a translation skill for each LLM you want to evaluate (e.g. "Translator — Gemini," "Translator — GPT," "Translator — Claude"). Your Smartcat contact can set these up with you.
Draft the business + voice/tone paragraphs. Provide Smartcat (or write directly, if self-serve) a short paragraph on your business and a short paragraph on your voice, tone, and audience. Smartcat can help draft these from your existing brand guidelines, style guide, or website content if you don't have them written down.
Identify content types with different audiences/voice. Review your content inventory (marketing collateral, product documentation, legal/compliance text, support content, etc.) and flag any category that needs its own skill rather than sharing one prompt.
Select a representative sample document for each content type to use for testing — ideally one that reflects real complexity (formatting, terminology density, tone requirements) rather than a trivial excerpt.
Ask your Chief of Staff to translate the sample with each translation skill variant and prepare the outputs as Variant A / B / C, with model identity withheld from reviewers.
Send the translated variants to your reviewers for evaluation, along with a short feedback template asking for: (a) which variant is best overall, (b) specific issues found in each variant, and (c) suggestions to improve the preferred variant further.
Repeat per target language. Since results can differ by language, run this comparison independently for each of your priority languages, not just once for the account overall.
For weak results on less-represented languages, test a DeepL (or other non-LLM) engine variant against the LLM variants before finalizing.
Finalize defaults. Once you've identified the winning model per language (and per specialized skill, if applicable), Smartcat configures these as the default engines for your projects, and continue tuning prompts based on ongoing reviewer edits.
Iterate — Once a best-fit Coworker/model is chosen for each language and content type, assign it as the default for that language/project. Continue collecting reviewer feedback over time and refining the prompt. This is an ongoing process, not a one-time setup.

💡Translator Coworker is also where you configure AI Translation Profiles — the settings that determine which TM, glossary, and engine apply per language/project. Translator Coworkers and AI Translation Profiles work together under Translator; see the companion AI Translation Profile guide for how profiles map engines to languages.

FAQ

What's the difference between a Translator Coworker and a regular AI translation engine?

A regular AI translation engine translates using its default behavior. A Translator Coworker is that same underlying model (Gemini, GPT, or Claude) configured with a custom prompt specific to your business — including your industry context, voice/tone, audience, glossary, and formatting rules — so the output is tailored to you rather than generic.

What kind of reviewer feedback is most useful?

Not just "Variant A is better." The most valuable feedback explains why a variant is better or worse, and what specific changes would improve the preferred variant further — that's what drives ongoing prompt refinement, not just a one-time pick.

How often should we re-test or re-tune the Coworker prompts?

This is an ongoing process rather than a one-time setup. As your content, terminology, or brand voice evolves — or as new model versions are released — periodically re-validate that your chosen Coworker/model per language is still the best performer.

Does using a custom Translator Coworker cost more or slow down translation?

No — configuring a custom prompt on top of an existing model is a standard part of AI translation setup and doesn't add meaningful extra time. It typically reduces review time and cost by producing output that needs fewer edits.

Who sets this up — the user or Smartcat?

Your Smartcat contact typically builds and configures the initial Translator Coworkers and prompts with you, based on the business, voice, and tone information you provide. From there, you review sample translations and give feedback, and Smartcat iterates on the prompt.

Does the Translator Coworker use AI Translation Profiles too?

Yes — they work together, but control different things. The Translator Coworker is the configurable engine that actually does the translating (the model + your tailored Translation Skill/prompt). The AI Translation Profile is what determines when and how that engine gets applied on a project: which Translation Memories and Glossaries are linked, which engine is assigned per language pair, and whether a fallback engine kicks in if the primary one fails. In short — the Coworker is the "who," the Profile is the "when/with what." See the companion "AI Translation Profile" guide for the full configuration walkthrough.

Do I need to pick just one model?

No. You can maintain several Translator Coworkers, each built on a different LLM, and even assign different ones to different languages or content types based on which performs best. Many customers end up with a "best model per language" mapping rather than one universal choice.

Why would different content need different Coworkers?

If your marketing copy and your technical documentation have different audiences, voice, and tone requirements, a single shared prompt will compromise one or the other. Creating specialized skills upfront (e.g. a "Marketing Coworker" and a "Technical Coworker") lets each be tuned precisely for its own content type.

How do we decide which model is best without bias?

We recommend blind testing: translate the same sample document with each candidate Coworker, label the outputs neutrally (Variant A / B / C) without revealing which model produced which, and have your reviewers evaluate them. This avoids reviewers favoring a model because of its brand reputation rather than its actual output quality.

What if none of the LLM models perform well for a language?

This is more common with less-represented/lower-resource languages. In that case, test a non-LLM engine such as DeepL alongside the LLM variants — for some languages, a dedicated AI translation engine still outperforms general-purpose LLMs.

Was this article helpful?

Still need help?

Our support team responds within one business day.

Open a support case