Customs Service Provider Cuts Document Processing Time by 67% with
a Gemini-Powered AI Agent
Challenge
Manual data entry slowing down document processing
Solution
AI-powered Telegram agent for fast and accurate document processing
We’ve implemented an AI agent using a multimodal Gemini model that communicates with clients via Telegram. Unlike OCR systems, Gemini understands the structure and content of documents. It correctly interprets tables and financial data, distinguishing total sums from intermediate values.
The agent also answers common questions and requests clarification when necessary.
Process
How we built a Telegram bot powered by Gemini
Manual document processing took hours of managers' time every day. To speed up submissions, we automated the workflow in 5 steps. Here’s a brief overview.
Stfalcon BAs mapped the document workflow
First, our Business Analysts (BAs) analyzed how managers processed incoming files: where documents arrived, who reviewed them, and how data moved into internal systems.
Most clients sent documents, like photos of IDs, vehicle documents with VINs, technical specs, and invoices via Telegram. Managers manually reviewed each file, extracted key information, clarified handwritten details, and re-entered everything into the CRM.
Mapping this process revealed exactly what we needed to automate and gave us the roadmap
Our backend team built the document bot in Telegram
Next, our backend development team started building the Telegram bot. Most clients already submitted documents there, so we integrated directly into their existing communication channel.
We developed a conversational workflow using Telegram’s Bot API, combined with a backend service that handles message routing, file uploads, and interaction logic. This way, we got a system ready for AI integration.
Our AI team integrated Gemini AI for document recognition
To process incoming files, our AI development team integrated Gemini, an AI model that works with photos, PDFs, and spreadsheets. We built a pipeline that sends documents to Gemini's API with extraction prompts and receives structured data back in a standard JSON format. Before submitting to the CRM, the bot returns the extracted information to the client for confirmation.
We connected the agent to the CRM
At this stage, we connected the AI agent with our client's CRM via API. Now, extracted data automatically moves to the manager’s pipeline. It includes all files (photos or PDFs) provided and a link to the Telegram conversation. This way, every client arrives in the CRM complete and ready for further human-led processing.
Stfalcon launched the agent in 1 month and kept refining it
The system was built and deployed in 1 month. Post-launch, we monitored interactions for 2 weeks and iteratively improved the prompts, dialog flow, document handling, and edge case responses.
Result
AI agent that triples daily document processing
The AI agent cut per-client handling from 30 minutes to 10, reducing document processing time by 67%. While the number of daily requests remained stable at 50–60, managers now spend significantly less time on manual data entry and document clarification. Seeing these results, the client plans to expand automation to additional operational processes together with Stfalcon.
- Automated extraction and CRM filling
- 10 minutes per client
- Agent asks for verification automatically
- High-accuracy structured data via Gemini
Ready to automate your document workflows?
We can help you build an AI agent that saves your team hours every single day
