Automation unloaded a translation team — Legitum

A small translation agency reached out to us because their time was simply lost on routine and manual work processes. The team wanted to stop drowning in files and small tasks and focus more on what they do best — translating. 

Challenge 

Before automation, most processes had to be handled manually. Mundane tasks en masse took too much time and ultimately limited growth. 

• For each client’s order, folders were created and files were passed to translators manually. Eventually, the abundance of folders and files to distribute across folders was a constant risk of errors and confusion.
  
• Managers prepared documents for translation themselves: they extracted text from files of different formats, adjusted formatting and unified the structure.
  
• The growing flow of clients made it harder to control statuses and quality, and scaling this process without adding people seemed almost impossible.
  

As a result, business growth hit a human‑resource ceiling. Overloading a human manager meant compromising speed and quality. 

Solution 

We built an end‑to‑end n8n-based solution that redesigned the entire document path from incoming file to finished translation. 

The core of the solution is made up of 3 major compenets: 

• n8n for process orchestration on the client side
  
• Dropbox as the main storage that the agency has already been using
  
• Google Gemini and other LLMs for text recognition, structuring and translation
  

The automation works in clear steps: 

1. File normalization. All incoming documents are converted to a single PDF format, duplicates are removed and orders are brought to a clean version before they are sent to the translator.
   
2. Classification and renaming. The LLM detects the language and type of document (for example, a diploma or certificate) and automatically assigns a standardized file name according to the scheme: “Source language_order number_document type_target languages_date”.
   
3. Advanced OCR (Optical Character Recognition) with structure preservation. The system extracts text so that tables and two-column layouts remain intact and unreadable fragments are marked UNREADABLE to quickly guide the translator the unclear segment.
   
4. Translation with strict rules. If there are several target languages, translation runs in parallel. The LLM prompts use explicit commands to preserve legal accuracy: “do not omit anything”, “do not abbreviate” and “do not rephrase”.
   
5. Document assembly using templates. The model selects the required template from the library, inserts the translation, converts text tables into full Word tables and brings everything to a professional layout. At the output, there is a clean source text and a finished translation that can be sent to the client immediately — naturally, after a final human review. 

To avoid repeated processing, the system maintains a table of processed orders and checks for the presence of special subfolders (EXTRACTION+TRADUCTIONS). If an order has already gone through the process, it is not launched again, even if the files are renamed. 

The automation runs every 10 minutes, scanning work folders. It can be triggered manually through a Telegram bot without waiting for the next cycle. 

Result

After launching the automation, the team’s workload noticeably decreased. 

• The time to process a single document was made considerably shorter. The entire cycle of extracting raw text to returning a complete translation now only takes a few minutes, only a fraction of the original time.
  
• A unified structure, templates and clear document-handling rules reduced errors in final documents.
  
• The agency was able to increase its capacity without growing its team and also streamline operations for effective future scaling.
  

Learnings

What our team took away from this was, the most intense work begins after automations are deployed.  

• Real‑world cases helped refine processing and translation rules to match the agency’s day‑to‑day practice. While we theorized on potential weaknesses of the automation, it was only after first orders were processed that we discovered parts that require extra attention and fine tuning.
  
• For complex and atypical documents, it is convenient to use more powerful models and manual review only where it truly matters. We were so focused on building the entire automation on a single LLM that we nearly missed the potential of using several models at once.
  
• Gradually, the team formed a transparent process: from incoming email to completed translation, with clear analytics and room for further scaling. This is to say that automation and streamlining don’t happen overnight. No matter how much testing you’ve put into it, the frictionless flow is a result of continued real-life application and feedback.
  

Importantly, automation did not replace people but became a tool that simplifies routine and lets focus on the most valuable part of the job — high‑quality translation and careful human review.