Silknet AI Chatbot Case Study: How We Automated 30% of Customer Chats in Georgian
In October 2026, ContainAI launched an AI customer service chatbot for Silknet, one of Georgia's leading telecom operators. The bot runs on Silknet's website and mobile app and talks to customers in Georgian, English and Russian. In its first, partial month it handled 17,000 customer chats and more than 48,000 messages, and resolved 30% of conversations without a human agent.
This case study explains what we built, why a "straightforward" project took three months, and how we keep a multilingual AI chatbot accurate in production.
Results at a glance
- Customer chats handled: 17,000
- Messages processed: 48,000+
- Average response time: 2.6 seconds
- Conversations resolved by AI: 30% (about 5,100)
- Channels: website and mobile app
- Languages: Georgian, English, Russian
- Kickoff to launch: about 3 months
The challenge: AI customer service in Georgian, English and Russian
Telecom customer service runs on volume. In Georgia most of that volume arrives in Georgian, and customers also write in English and Russian. Silknet needed an AI chatbot that could answer every customer correctly in their own language, around the clock, on both web and mobile.
Answering simple questions is the easy part. The hard part is giving the right answer when it depends on the customer's situation, and never inventing one.
Why a "straightforward" project took three months
On paper, the project looked simple: connect an LLM to Silknet's knowledge base and launch. In practice, the work was in the edge cases.
- Branching logic in the knowledge base. Many answers follow a decision tree: the correct reply depends on several conditions. We designed the knowledge base to carry that branching logic itself, managed through an admin panel, so the chatbot follows it reliably.
- A lean system prompt. Every rule pushed into the system prompt can interfere with the others. We kept the prompt focused and moved case-specific logic into the knowledge base.
- Cache tuning. We keep a warm cache in an optimal range. Too little caching makes the bot slow and expensive; too much makes it stale and hard to update.
Getting these three to work together is what took the time, and it is what makes the chatbot dependable now.
The solution: how the Silknet AI chatbot works
- LLM core with a carefully engineered system prompt
- Knowledge base as the single source of truth, including edge-case branching logic
- Admin panel for managing and updating the knowledge base
- Warm cache for fast, cost-efficient responses
- Fallback models: if an LLM provider goes down or slows down under high demand, the chatbot switches to another model
- Reserved infrastructure to keep the service available under load and during outages
- Three languages: Georgian, English and Russian
- Two channels: Silknet's website and mobile app
How we prevent AI chatbot hallucinations
Accuracy matters more than speed in customer service. A wrong answer costs more than a slow one. So every response gets a confidence score.
- If the confidence score is low, the chatbot generates the answer again.
- If the two answers differ, a third assessment decides the response.
- That third step is reserved for rare edge cases, and it almost never runs.
Customers get one clear answer. The cross-checking happens before the answer reaches them.
The results
In October 2026, its first partial month live, the Silknet AI chatbot handled 17,000 chats and more than 48,000 messages, with an average response time of 2.6 seconds. It resolved 30% of conversations from start to finish, about 5,100 conversations that did not need a human agent. That leaves Silknet's agents more time for the cases that need a person.
What this means for Georgian businesses
If your customers write to you in Georgian, English or Russian, an AI chatbot can now handle a real share of those conversations, accurately and at scale. Banks, insurers, retailers, e-commerce stores and service companies face the same challenge Silknet did: high chat volume, complex rules, and no room for wrong answers.
Frequently asked questions
Can an AI chatbot work well in Georgian?
Yes. The Silknet chatbot talks to customers in Georgian and resolved 30% of 17,000 chats in its first month.
Can one chatbot handle several languages?
Yes. The Silknet chatbot answers customers in Georgian, English and Russian, on the website and in the mobile app.
How long does it take to launch an AI chatbot?
The Silknet project took about three months from kickoff to launch. Most of that time went into edge cases and knowledge-base logic, not the basic setup.
What happens if the AI provider goes down?
The chatbot switches to fallback models and runs on reserved infrastructure, so the service stays available.
How do you prevent chatbot hallucinations?
Every answer gets a confidence score. Low-confidence answers are generated again and compared. If they differ, a third assessment decides the final response.
Which channels can the chatbot run on?
For Silknet, the chatbot runs on the website and the mobile app.
Planning an AI chatbot for your company?
The ContainAI team builds AI chatbots for Georgian businesses, from knowledge-base design to production monitoring. Contact us at containai.io to talk about your project.