How to integrate neural networks with Bitrix24
A manager spends hours listening to call recordings, a marketer waits for a designer for a simple banner, leads are distributed manually. We integrate neural networks with Bitrix24 via REST API to automate these tasks.
Our ten years of experience in Bitrix development and certified specialists guarantee stable operation of the integration. Over four years, we have completed more than 50 neural network integration projects — from transcription to complex ML pipelines.
How is the integration architecture structured?
The scheme is standard: Bitrix24 REST API → middleware server → neural network APIs (OpenAI, Replicate, proprietary models). Middleware receives an event from Bitrix24, determines the task type, selects a model, forms a request, receives the result, and returns it to Bitrix24. One middleware works with a dozen models — routing by task type.
How to automate call transcription using Whisper?
The most requested scenario after text generation. The call ends, the recording enters Bitrix24. Whisper turns audio into text, and GPT turns it into a structured summary.
Technical chain:
- Event
ONVOXIMPLANTCALLENDis sent to middleware. - Middleware downloads the call recording via
voximplant.statistic.get. - Audio is sent to Whisper API (parameters:
model=whisper-1,language=ru,response_format=verbose_json). - Transcription is passed to GPT with the prompt: “Highlight key agreements, client questions, next steps. Format: JSON”.
- The result is written to the call comment or to a deal custom field via
crm.timeline.comment.add.
Time savings for the manager are obvious: instead of listening to a 20-minute call — reading a summary in 30 seconds. The cost of each Whisper and GPT request is minimal, while savings on manual labor can be significant each month. Payback occurs within the first weeks.
How to generate images directly from Bitrix24?
A marketer needs a banner for a newsletter or a visual for a product card. Instead of waiting for a designer — a request to a neural network from the Bitrix24 interface.
Implementation:
- Via a chatbot. The marketer writes to the bot: “Generate a banner for the -20% promotion on the winter collection, minimalism style, format 1200×628”. The bot sends the prompt to DALL-E 3 or Stable Diffusion via Replicate API. It receives the image URL, uploads it to Bitrix24 Disk via
disk.file.uploadtofolder, and sends a preview to the chat. - Via a business process. When a marketing activity is created in CRM, visual variants are automatically generated based on campaign parameters.
DALL-E 3 works well with conceptual images. For photorealistic images and style control, we use Stable Diffusion XL — it is more flexible in configuration, supports ControlNet and LoRA models.
According to tests, SDXL generates images 2 times faster than DALL-E at similar quality, and request cost is 40% lower.
Why do lead classification and sentiment analysis increase conversion?
Text models solve tasks that are not limited to generation:
- Lead classification by topic. The model determines the request category: “website development”, “support”, “integration”. Based on the category, middleware via
crm.lead.updatesets the direction and assigns a responsible person. - Sentiment analysis. Emails and messages pass through a sentiment analysis model. For negative sentiment, the request receives increased priority, and the manager is notified. We use a fine-tuned model based on BERT or a prompt to GPT.
- Entity extraction (NER). Structured data is extracted from the email text: company name, tax ID, deadlines, budget. Lead fields are filled automatically via
crm.lead.update. Manual card filling is eliminated.
Example: an incoming email enters CRM. Middleware extracts the text, sends it to GPT-4o-mini with the prompt “Classify the request by categories: development, support, integration, other. Return JSON with category and confidence”.
The result is written to the custom field UF_CRM_CATEGORY. If confidence is below 0.7, the lead is sent for manual review.
Which models to choose for different tasks?
| Task | Model | API |
|---|---|---|
| Call transcription | Whisper | OpenAI |
| Conversation summary | GPT-4o-mini | OpenAI |
| Lead qualification | GPT-4o-mini | OpenAI |
| Image generation | DALL-E 3 | OpenAI |
| Photorealistic images | SDXL | Replicate |
| Sentiment classification | BERT fine-tuned | Proprietary server |
| Entity extraction | GPT-4o-mini / spaCy | OpenAI / own server |
For complex scenarios, we build ML pipelines: incoming email → text extraction → classification → sentiment detection → qualification → routing. Each step is a separate API call with error handling and fallback.
Implementation timelines
| Scale | What is included | Timeline |
|---|---|---|
| One scenario | Call transcription or image generation | 3–5 days |
| Complex | 2–3 scenarios, classification + transcription + generation | 1–2 weeks |
| ML pipeline | Full processing chain, custom models, monitoring | 3–5 weeks |
What is included in the work
- Middleware server with request routing by task type and model
- Configuration of Bitrix24 REST API for receiving events and writing results
- Integration of selected neural networks (Whisper, GPT, DALL-E, SDXL, custom)
- Writing prompts and processing algorithms for each scenario
- Dashboard for monitoring token usage and logging
- Integration documentation and employee training
- Warranty support after implementation
We prepare a project estimate after reviewing the tasks. Contact us to discuss your tasks. Request a consultation — we will select optimal models and prepare a turnkey integration. Get a timeline and cost estimate within one business day.