AI Chatbot for Support: RAG, Integrations, and Escalation
Picture this: you're a support operator, and 80% of queries are repetitive: "Where is my order?" or "How to return?" These consume 70% of your time. We build AI chatbots that solve real tasks — not just respond, but modify orders, process returns, check statuses. The difference lies in integration with your systems and RAG.
How We Integrate an AI Bot with CRM
A useless bot answers "Check your personal account" without seeing any data. We connect the bot to your CRM via REST API: it retrieves history, order status, and tickets. Every action is authorized — only a verified user can change an address. We implement OAuth 2.0 or JWT.
Case study: For an online store with 50,000 orders per month, we implemented integration with 1C-Bitrix. The bot checks status by phone number and processes returns without an operator. Containment rate rose from 20% to 65%.
Why RAG Is Key to Accurate Answers
Storing all answers in an LLM prompt is inefficient — context windows are limited and knowledge becomes outdated. RAG adds retrieval from a knowledge base: semantic caching, vector embeddings, ranking. We use Pinecone or Qdrant for embeddings, and models like GPT-4 or LLaMA 3.
Details of the RAG pipeline
- Documents are split into chunks (256–512 tokens).
- Each chunk is indexed using an embedding model (text-embedding-ada-002).
- On a query, cosine similarity is computed, returning top-5 chunks.
- The LLM generates an answer considering retrieved contexts and dialogue history.
Architecture of a Production Chatbot
[Client] → [Omnichannel Interface] → [NLP Engine: Intent + Entity] → [Dialog Manager]
├── RAG: answers from knowledge base
├── Action Engine: integrations with CRM/ERP
└── Escalation: handoff to operator
→ [Response Generator] → [Analytics & Logging]Core: Classification and Slot Filling
Comparing the classic approach (Rasa) with LLM + function calling:
| Parameter | Intent + Slots (Rasa) | LLM + Function Calling |
|---|---|---|
| Intent accuracy | 90–95% | 95–98% |
| Setup complexity | Requires labeled examples | Just describe tools |
| Flexibility | Low for new scenarios | High — LLM understands paraphrases |
LLM with function calling reduces new scenario setup time by 3× compared to Rasa. Example tools configuration:
tools = [
{
"name": "check_order_status",
"description": "Проверить статус заказа по номеру",
"parameters": {
"order_id": {"type": "string"}
}
},
{
"name": "initiate_return",
"description": "Оформить возврат товара",
"parameters": {
"order_id": {"type": "string"},
"reason": {"type": "string"}
}
}
]Escalation to Operator
Triggers for handoff to a human: client explicitly asks for an operator, negative sentiment, bot fails to solve after 3 attempts, VIP client, legal claims. On escalation, the operator receives the full dialogue context — no need to ask the client to repeat.
Metrics and Quality
- Containment rate: target 50–70%.
- CSAT (bot): >4.0/5.0.
- Resolution rate: % of issues actually resolved.
- Escalation rate: balance — no higher than 30%, no lower than 10%.
| Stage | Time | Result |
|---|---|---|
| MVP | 4–6 weeks | Basic intents + CRM integration |
| Full feature | 3–4 months | RAG, omnichannel, all scenarios |
What's Included in the Work
- Audit of business processes and preparation of knowledge base.
- Design of dialogue scenario (up to 50 intents).
- Development of integrations with CRM/ERP/billing.
- Deployment on infrastructure (Kubernetes, AWS, or self-hosted).
- Model training on historical dialogues (LoRA fine-tuning).
- Monitoring and support for 3 months.
Timeline and Value
- MVP: 4–6 weeks.
- Full-featured bot with integrations: 3–4 months.
- Operator cost savings — up to 70%, payback period — 6–12 months.
Omnichannel: Telegram, WhatsApp, Web Widget
The bot is deployed across multiple channels via a unified API. The architecture includes a channel-agnostic Dialog Manager: configure scenarios once, get the bot on Telegram, WhatsApp Business API, and website via web widget. Historical data is stored with a single CRM binding: if a client starts a conversation on Telegram and continues on the website, the operator sees the full chain.
For an omnichannel bot, we use RabbitMQ or Kafka queues: this ensures fault tolerance under peak loads and allows processing up to 500 simultaneous conversations on a single instance. Horizontal scaling is achieved by adding workers without changing the architecture.
We have been in the market, developed 40+ chatbots for e-commerce, fintech, logistics. We guarantee a containment rate of at least 50% after 3 months of operation. Assess your project: we'll send case studies and propose a turnkey architecture. Contact us for a consultation and get a detailed timeline estimate.