We have implemented a Telegram bot with AI in a federal company

How we implemented an AI-powered Telegram Bot in a federal company: from MVP to automation of expertise

My name is Alexey, and I lead the artificial intelligence division in a large federal-scale commercial organization. The company is an industry leader, with a well-developed infrastructure and a high level of automation.

Today I want to share the story of how we implemented a Telegram bot service with elements of artificial intelligence to solve a specific business problem. This is a story of how we started with an MVP, gradually added AI, and how it helped employees save time and improve the accuracy of operations.

Diving into company processes

When I first started working, my task was to find out where inside the company there was potential for implementing artificial intelligence technologies. I began by deeply immersing myself in processes: studying the structure, talking to department heads, analyzing existing systems.

Special attention was given to identifying "pain points"—processes that were labor-intensive but had the potential for automation or enhancement with algorithms.

The problem: labor-intensive document creation process

One of these processes was the following: employees needed to photograph an object, transfer the photo through several systems (personal email → corporate email → remote desktop → program → creation of a rejection document), create a document in a specialized internal program and send it for approval.
Now: TG-Bot → Rejection document.

This entire operation took 40 to 60 minutes. Not only was it time-consuming, but it also increased the likelihood of errors.

Initial project plan

Initially, we planned to start by creating several artificial intelligence models capable of determining key parameters of the object to be rejected. It was also intended to set the model’s sensitivity threshold so that it could gain the trust of experts.

After that, we planned to develop a Telegram bot in which users could test the model and provide feedback. Only after that was a full-fledged service supposed to be created.

But reality made its own adjustments—we decided to change the approach and start with what would bring results already at the early stages.

Changing the approach: from MVP to AI

We decided to implement an MVP (minimum viable product)—a simple Telegram bot that would allow employees to:

• Take a photo;

• Automatically create a document;

• Send it for approval.

Technically, it was Python + libraries + CNN, integrated with the company's internal API. The result? The task completion time dropped from 40–60 minutes to less than 5 minutes.

This was the first success, which won us user support and allowed us to move forward.

Version 2: Upload quality support

The next stage was improving photo quality, since the accuracy of subsequent analysis depended on it.

We added hints to the bot: how to take photos correctly — the right angle, lighting, and object position. This significantly reduced the number of errors at the approval stage and increased the efficiency of the entire process chain.

Version 3: AI Integration and Reducing the Load on Experts

We have now launched the third version of the service, which completely changes the approach to the task:

• The user can upload photos in any format.

• The system automatically sorts the images and sends them to the appropriate neural network models.

• Each parameter is checked by artificial intelligence algorithms against six criteria.

• Inspectors only need to confirm or adjust the AI's conclusions.

This has significantly reduced the load on experts, moving from full manual validation to monitoring the models' results.

Model Training and Adaptation Process

We collected the data for training the models ourselves, labeling it within the team. We used pre-trained YOLO models, which we fine-tuned for our specific tasks.

The process was iterative: we ran experiments, compared metrics, and chose the most accurate options. Integration with internal systems happened via REST API, which provided both security and flexibility.

Results and Impact

• Reduced operation time from 40–60 min to <5 min.

• Increased approval accuracy thanks to better data quality.

• Lowered expert workload due to AI support.

• Increased trust in technology within the company.

Lessons and Conclusions

• Start with the most pressing problem, even if it’s not directly related to AI.

• Actively collect feedback from end users.

• AI implementation isn’t just about technology — it’s also about change management.

• You can start with an MVP and gradually increase complexity.

Challenges We Faced

The path wasn’t easy. One of the biggest challenges was limited resources: financial, computational, and human. It was especially difficult to explain to the internal customer what was happening "under the hood" of development and to deliver results quickly, without the luxury of lengthy R&D.

These challenges deserve a separate article, which I’ll definitely prepare in the future.

Outlook

We are currently working on expanding the bot’s functionality: we plan to add voice descriptions of situations, improve the interface, and roll out the service to other departments and regions.

Moreover, this project became a launching point for the development of the entire AI direction within the company. Now we can talk about scaling solutions and new automation opportunities.

If you found this story interesting — subscribe, there will be even more cases from the world of AI implementation in large companies!