AI Plugin for Bitrix24: Automate CRM with Neural Networks

Your CRM manager spends 25 minutes filling out a deal card after each call? Commercial proposals are generated from templates, and conversion suffers due to generic emails? This is a typical picture in sales departments using Bitrix24, where routine eats up time for real negotiations.

We are a team of AI/ML engineers with over 5 years of experience integrating neural networks into CRM. Our approach: not just attaching an "AI feature," but embedding it into workflows so that the manager feels no friction. Result: operational cost reduction by 80% and conversion increase by 12%.

Why Embed an AI Plugin Directly into Bitrix24?

Ready-made AI assistants (SalesGPT, Gong) do not integrate with Russian CRM. Bitrix24 is a flexible platform: through REST API, Placement, and Events, you can add any logic. The challenge: the complexity of developing the AI layer—from LLM selection to latency optimization. We address this with four modules:

How We Implement the AI Plugin

Each plugin is built modularly. Stack: Python + Anthropic/OpenAI API, deployment via Docker + Bitrix24 Placement. Webhook handler code:

# webhook_handler.py — processing Bitrix24 events
from flask import Flask, request, jsonify
from anthropic import Anthropic
import requests

app = Flask(__name__)
client = Anthropic()

BITRIX_URL = "https://your-domain.bitrix24.ru/rest"
BITRIX_TOKEN = "your-webhook-token"

def bitrix_api(method: str, params: dict) -> dict:
    """Call Bitrix24 REST API"""
    response = requests.post(
        f"{BITRIX_URL}/{BITRIX_TOKEN}/{method}/",
        json=params,
    )
    return response.json().get("result", {})

@app.route("/webhook/call-ended", methods=["POST"])
def on_call_ended():
    """Handler for call end event"""
    data = request.json
    call_id = data.get("data", {}).get("CALL_ID")
    crm_entity_id = data.get("data", {}).get("CRM_ENTITY_ID")

    # Get call transcript
    call_info = bitrix_api("voximplant.statistic.get", {"CALL_ID": call_id})
    transcript = call_info.get("TRANSCRIPT", "")

    if not transcript:
        return jsonify({"status": "no transcript"})

    # AI summarization
    summary = summarize_call(transcript)

    # Record in CRM as activity
    bitrix_api("crm.activity.add", {
        "fields": {
            "OWNER_TYPE_ID": 2,  # 2 = Contact, 3 = Company
            "OWNER_ID": crm_entity_id,
            "TYPE_ID": 6,  # Call
            "SUBJECT": "Call summarization (AI)",
            "DESCRIPTION": summary["text"],
            "DESCRIPTION_TYPE": 1,
        }
    })

    # Extract key data and update deal fields
    if crm_entity_id:
        updates = extract_crm_fields(transcript)
        if updates:
            bitrix_api("crm.deal.update", {
                "id": crm_entity_id,
                "fields": updates,
            })

    return jsonify({"status": "ok"})

def summarize_call(transcript: str) -> dict:
    """Summarize call transcript"""
    response = client.messages.create(
        model="claude-haiku-4-5",
        max_tokens=1024,
        system="""Summarize sales negotiations.
Format:
- Brief summary (2-3 sentences)
- Key agreements
- Next steps
- Client objections""",
        messages=[{
            "role": "user",
            "content": f"Transcript:\n{transcript}"
        }]
    )
    return {"text": response.content[0].text}

def extract_crm_fields(transcript: str) -> dict:
    """Extract data for updating CRM fields"""
    import json

    response = client.messages.create(
        model="claude-haiku-4-5",
        max_tokens=512,
        messages=[{
            "role": "user",
            "content": f"""Extract data from conversation for CRM.
Return JSON: {{
  "TITLE": "deal title if mentioned",
  "OPPORTUNITY": number (budget if mentioned),
  "COMMENTS": "important notes"
}}
If field not mentioned — do not include it.

Conversation: {transcript[:2000]}"""
        }]
    )

    text = response.content[0].text
    try:
        return json.loads(text[text.find("{"):text.rfind("}") + 1])
    except Exception:
        return {}

UI Placement — embedding into deal card

// placement.js — embedded widget in CRM card
BX24.init(function() {
    // Button "AI Analysis" in deal card
    BX24.placement.bind('CRM_DEAL_DETAIL_TAB', {
        title: 'AI Assistant',
        onClick: function() {
            showAIPanel();
        }
    });
});

async function generateCommercialProposal(dealId) {
    // Get deal data
    const deal = await BX24.callMethod('crm.deal.get', { id: dealId });

    // Request proposal generation
    const response = await fetch('/ai/generate-proposal', {
        method: 'POST',
        body: JSON.stringify({ deal: deal.result }),
        headers: { 'Content-Type': 'application/json' }
    });

    const result = await response.json();

    // Insert proposal into description field
    await BX24.callMethod('crm.deal.update', {
        id: dealId,
        fields: { COMMENTS: result.proposal }
    });
}

Commercial Proposal Generation

@app.route("/ai/generate-proposal", methods=["POST"])
def generate_proposal():
    deal = request.json.get("deal", {})

    response = client.messages.create(
        model="claude-sonnet-4-5",
        max_tokens=2048,
        system="""You are a B2B sales manager.
Create professional commercial proposals based on deal data.""",
        messages=[{
            "role": "user",
            "content": f"""Create a proposal for the client.
Deal data:
- Title: {deal.get('TITLE')}
- Client: {deal.get('COMPANY_ID')}
- Amount: {deal.get('OPPORTUNITY')} {deal.get('CURRENCY_ID')}
- Comments: {deal.get('COMMENTS', '')}

Proposal structure: greeting, understanding of problem, proposed solution, benefits, cost, next step."""
        }]
    )

    return jsonify({"proposal": response.content[0].text})

Which LLM Models Are Optimal for Different Tasks?

Model selection is critical for balancing speed and quality. For call summarization we use Claude Haiku — it is 3x faster than Sonnet and 5x cheaper, while key field extraction quality drops by less than 5%. For proposal generation, where depth and personalization matter, we use Claude Sonnet or GPT-4o. Internal tests showed that with the same API budget, the Haiku + Sonnet combination processes 40% more calls than using one expensive model.

Practical Case: Sales Department of 15 Managers (from our practice)

Client — an IT equipment distributor. Each manager spent 30–40 minutes after a call filling CRM and composing summaries. Conversion of sent proposals was 18% (too template-like).

We implemented three modules: call summarization, automatic deal field updates, and proposal generation. After two weeks of use:

• Post-call processing time: 30 min → 5 min (83% operational cost reduction)
• Conversion of sent proposals: 18% → 30% (gain due to personalization)
• Errors in filling fields (budget, date) — 12 per day → 0
• Managers could handle 3 more negotiations per day

Head of Sales Department: "We expected time savings, but didn't expect such conversion growth. Now proposals truly engage clients."

What's Included in the Work

• Audit of current business processes in Bitrix24
• Designing AI module architecture (LLM selection, prompt tuning, RAG schemes)
• Developing webhooks and Placement widgets
• Integration with telephony (VoxImplant, Mango Office)
• Configuring proposal generation for your product line
• Testing on real data (at least 50 transactions)
• Publication in Marketplace or installation on your server
• Operation documentation and administrator training

Work Process

1. Analytics (1-2 days): study deal schemas, fields, events. Identify AI insertion points.
2. Design (2-3 days): select models, draw architecture, agree on scenarios.
3. Development (3-7 days): write code, embed widgets, configure webhooks.
4. Testing (2-3 days): run on historical data, fix errors, optimize latency to p99 < 2s.
5. Deployment and training (1-2 days): roll out to production, train administrators, hand over documentation.

Indicative Timeline

Final cost is calculated individually for your process. Leave a request — we will assess the project for free and offer the optimal solution.

We guarantee quality: certified Bitrix24 engineers, over 5 years of experience, more than 20 successful integrations. Our solutions undergo code review and load testing. Contact us to get a consultation and see a demo on your data.

LLM Development: Fine-Tuning, RAG, Agents, and Production Deployment

Using GPT‑4 or Claude 3.5 Sonnet through a public API is not a solution — it's just a tool. When the requirement is to "make it like ChatGPT, but on our data," there is a real engineering challenge behind it: from prompt engineering to training a 70B model on your own infrastructure. End-to-end LLM solution development is a complex stack, and we have been doing it for over 5 years. During this time, we have completed over 20 projects in generative AI: from RAG systems for legal departments to custom support agents. Where exactly your task falls depends on data, latency requirements, budget, and how critical confidentiality is.

A typical situation: the client has already tried ChatGPT, but results are unstable — sometimes accurate, sometimes hallucinating. Or they need integration into a corporate portal while complying with security policies. Let's break down each layer of the stack in detail — from RAG to production deployment.

Why Do RAG Systems Break and How to Fix It?

RAG (Retrieval-Augmented Generation) looks simple: find relevant documents, put them in context, get an answer. In practice, it fails in several places.

Chunking without overlap. Classic mistake: chunk_size=512, overlap=0. If the answer lies across two chunks, retrieval won't find either with sufficient confidence. Solution: overlap 15–25% of chunk_size, or better yet, sentence-aware splitting with spaCy or NLTK instead of naive character splitting.

Poor embedder. text-embedding-ada-002 is good for general use, but on legal or medical texts, specialized models like E5-large-v2, BGE-M3, or fine-tuned sentence-transformers on domain data outperform it. Recall@5 differences can be 15–25%.

No re-ranking. Vector search optimizes for speed, not relevance. A cross-encoder re-ranker (ms-marco-MiniLM-L-6-v2, bge-reranker-large) after initial retrieval improves top-3 accuracy with acceptable latency (+50–150ms). This is often more impactful than improving the embedding model.

Hybrid search. Dense vectors alone work poorly on exact queries: names, SKUs, codes. BM25 (sparse) finds exact matches but misses semantics. Hybrid via RRF (Reciprocal Rank Fusion) is the optimal compromise. Qdrant, Weaviate, and pgvector 0.7+ support hybrid search natively.

When to Fine-Tune Instead of Prompt Engineering?

Prompt engineering solves ~70% of LLM adaptation tasks for a domain. The remaining 30% require fine-tuning. Three indicators: the model ignores a specific output format even with detailed prompting; the task requires deep knowledge of specialized vocabulary (medicine, law); you need to significantly reduce token costs by replacing a large model with a smaller specialized one.

LoRA and QLoRA are the standard for SFT. LoRA adds trainable low-rank matrices to attention layers. A typical configuration for Llama-3 8B: r=64, lora_alpha=128, target_modules=["q_proj","v_proj","k_proj","o_proj"] yields ~0.8% trainable parameters, training on one A100 40GB. QLoRA adds 4-bit quantization (NF4) and allows fine-tuning 70B models on two A100 40GB, though speed drops by half compared to bf16.

DPO instead of RLHF. Direct Preference Optimization requires only (chosen, rejected) pairs, not scalar reward signals. DPOTrainer from the trl library (Hugging Face) implements it in a few dozen lines.

Common mistake. A dataset of 500 examples, 5 epochs, validation loss 0.8 — seems fine. But on test, the model degrades on general instructions. Cause: catastrophic forgetting. Solution: add 10–20% general instruction-following examples (Alpaca, FLAN) to the training set to preserve original capabilities.

How to Choose a Base Model: 8B or 70B?

For most tasks, fine-tuning an 8B model is sufficient. 70B is needed when deep reasoning is required or the 8B baseline does not reach the required quality even after fine-tuning. Inference cost for Llama-3 8B via vLLM on A100 is efficient; the exact cost depends on volume.

What Does PagedAttention Bring to Production?

vLLM is the first choice for serving open-source models. PagedAttention is the key technical innovation: KV-cache is managed like virtual memory in an OS, without fragmentation. This yields 2–4x higher throughput compared to naive HuggingFace Transformers inference. The vLLM documentation confirms that continuous batching and PagedAttention are the standard for high-load LLM services.

Typical numbers on A100 80GB for Llama-3 8B (bf16): 400–600 req/s, P50 latency 200–400ms, P99 latency 600–900ms at concurrency 64. For 70B on two A100 with tensor parallelism: 80–120 req/s, P99 latency 1.5–2.5s. AWQ or GPTQ quantization reduces memory consumption by 2x with quality loss within 1–3%.

Multi-Agent Systems

Agents are LLMs with access to tools: search, code execution, API calls, database interaction. Common patterns:

• ReAct (Reason + Act): the model reasons → chooses a tool → observes the result → reasons again. LangChain and LlamaIndex implement it out of the box.
• Multi-agent orchestration: multiple specialized agents with a coordinator on top. Example: coordinator → researcher (search + summarization) → coder (code generation and execution) → critic (verification). Tools: AutoGen (Microsoft), CrewAI, custom implementation on LangGraph.

In production, agent systems are non-deterministic. Essential: guardrails, step limits, logging of each step, human-in-the-loop for critical actions.

How We Work: Stages, Timeline, Deliverables

What Is Included

We deliver:

• Technical documentation (model card, configs, deployment instructions)
• Access to infrastructure (code repository, trained weights)
• 1 month of post-deployment support (consultations, bug fixes)
• Customer team training (2–3 sessions on system operation)

Timeline: basic RAG prototype — 1–2 weeks. Fine-tuning with customer data — 3–6 weeks (including data preparation). Production system with monitoring and retraining — 2–4 months. Cost is calculated individually based on data volume, model complexity, and infrastructure requirements.

We guarantee the quality of the final model with performance benchmarks and ongoing monitoring. Our engineers have hands‑on experience with dozens of production LLM systems.

Want to evaluate your project? Leave a request — we will prepare a preliminary summary within 1–2 business days. Or get a consultation on choosing the approach: RAG, fine-tuning, or hybrid — we will tell you what works best for you. Contact us to discuss your LLM development needs. Schedule a free consultation today.