In short
Bitrix24 ships its own AI layer. On bitrix24.com it is called CoPilot: it transcribes calls, fills CRM fields, scores calls against a script, summarizes chats and writes text for emails, tasks and posts. The Russian-language edition calls the same layer BitrixGPT, its own model run on Bitrix24 servers in Russia, and has added two agents on top of it: one that carries out employee requests inside the portal, and one that answers customers in open channels from an uploaded knowledge base.
That built-in layer is enough when the answer already lives in Bitrix24 or in a few documents you can upload. A custom agent earns its cost when customers write free text and expect answers that depend on data Bitrix24 does not hold: stock in your ERP, a customer's balance, a negotiated price, a discount rule kept in a spreadsheet. It also matters when you need to choose the model, test answer quality on your own conversations, and keep a log of every decision.
The example below follows one company: a commercial kitchen equipment supplier in Dubai with a showroom, a projects team and a spare-parts desk. It runs Bitrix24 Cloud with open channels for WhatsApp, Instagram and the website chat, keeps stock and invoices in its ERP, and holds dealer price lists in Google Sheets. The company is made up; the problems are not.
What Bitrix24 CoPilot does today
CRM and calls
CoPilot transcribes calls, writes a short summary into the deal timeline and fills deal fields such as budget and timing. It scores calls against your sales script and builds repeat-sales suggestions. According to the Bitrix24 help center, CoPilot on bitrix24.com uses OpenAI's gpt-4o for chats, text, call summaries, call analysis and CRM field updates, DALL-E 3 for images, and separate speech services for call and voice-message transcription.
For the kitchen supplier this fixed a real gap. Sales engineers spent the day on site visits and logged calls hours later, if at all. Now the deal card fills itself and the engineer only corrects it.
CoPilot in chats
Every employee can open a CoPilot chat, pick one of 30+ roles such as sales manager or lawyer, or mention CoPilot in any chat to summarize a thread or draft a reply. It is a writing and thinking aid inside the portal.
AI agents
The Russian-language help center describes two newer pieces. The BitrixGPT agent takes actions on request: it creates and edits deals, tasks and meetings while respecting the employee's access rights. A separate AI agent for open channels answers customers in messengers, social networks and the site widget from documents you upload, moves data from the chat into CRM fields, and hands the conversation to an operator with its history. The help center notes two limits: the open-channel agent is not available on every plan, and it cannot share a channel with another chatbot.
Rollout differs by region and plan, so check your own portal before you plan around these agents.
Limits
Bitrix24 caps AI requests at 600 per hour for the whole portal on every plan. Outside the Russian zone, each plan includes a monthly request package, and paid plans can buy extra capacity in boosts of 1,000 requests a month. A team of ten sales reps with a hundred calls a day uses several requests per call: transcription, summary and field filling each count.
Where the built-in AI stops
The supplier switched on everything above and still had trouble in the customer-facing channels. These are the messages that broke it.
"Do you have the 10-tray combi oven in stock, and what is our dealer price?" Stock lives in the ERP, and dealer prices depend on quarterly volume in a spreadsheet. The open-channel agent answers from uploaded documents. It does not see this morning's stock or the dealer's tier.
"Can you send our statement? I think we overpaid on the last invoice." That needs account data from the ERP and a check that the person writing actually works for that customer.
"Our contract says 60-day terms, why is this invoice prepaid?" The answer depends on a PDF contract, a credit decision and payment history. The agent needs the company's rules and permission to say "passing you to your account manager" instead of guessing.
A voice note in Arabic followed by a text in English. Mixed languages are normal in Gulf B2B sales. Whether a model handles your customers' language well is something you measure on your own conversations, not something you assume.
So a custom agent is worth it when:
- answers need data outside Bitrix24: ERP, inventory, spreadsheets, a delivery service;
- the agent has to act in more than one system, for example check credit and then create a deal;
- your discount, credit and exception rules change often;
- customers write in several languages or mix them;
- someone needs to see why the agent answered the way it did, and quality has to be tested on a fixed set of real conversations before each change;
- you want a specific model provider, such as OpenAI or Anthropic, under terms your legal team has reviewed.
If none of these apply, start with the built-in agent. It is cheaper and it runs the same day.
How a custom AI agent connects to Bitrix24
A custom agent does not replace the portal. Operators keep working in the same open channels and deals; the agent joins the message flow as one more participant. Everything it needs is in the public Bitrix24 REST API.
| What the agent needs | How Bitrix24 provides it |
|---|---|
| Receive customer messages from an open channel | A chatbot of type openline registered through the chatbot API |
| Hand the conversation to a person | imopenlines.bot.session.operator in the open channels API |
| Read and create deals | crm.item.*, which Bitrix24 now recommends over the older crm.deal.* methods |
| Leave a note for the rep | A timeline comment, crm.timeline.comment.add |
| Plug your own model into CoPilot | ai.engine.register, which adds your service as a selectable AI engine |
Between Bitrix24, the ERP and the model sits a small service of your own. It holds the credentials, checks who is writing, pulls stock and balances, builds the context and logs each decision. The model never touches a database directly and only sees the actions it is allowed to call.
For internal work there is a lighter option: Bitrix24 runs an official MCP server that lets ChatGPT, Cursor and similar tools read CRM data and tasks once an administrator enables it. That is enough for an ad hoc pipeline review. An agent that talks to customers still needs its own service with rules and a log.
The supplier's first agent did three things. It answered stock and price questions from ERP and spreadsheet data, it took orders and created deals with line items, and it passed everything about credit, disputes and overdue invoices to a person with a short summary. CoPilot stayed on calls and internal chats.
Built-in AI or a custom agent
| Task | Built-in AI | Custom agent |
|---|---|---|
| Call transcription, field filling, script scoring | Yes | Not needed |
| Drafting and summarizing in chats | Yes | Not needed |
| Customer FAQ from documents | Yes, the open-channel agent | When you need your own quality tests and logs |
| Stock, price or balance from the ERP in a reply | No | Yes |
| Company-specific credit and discount rules | Partly, through instructions | Yes |
| Several languages, voice notes, mixed messages | Test on your conversations | Model chosen by tests |
| Actions across several systems | No | Yes |
Most companies end up with a mix: calls and internal work stay on CoPilot, and customer conversations that need operational data go through a custom agent. The broader pattern is in our article on AI CRM integration.
Running a pilot
- Export 300 to 500 open-channel conversations from the last two months and tag the questions and the data each answer needed.
- Turn on the built-in agent on one channel with a knowledge base. Count how many conversations it closes and how many go to an operator because data was missing.
- If more than a third fail on missing data, list where that data lives. That list is the scope of your custom agent.
- Build a test set from real conversations with correct answers and run it before every prompt or model change. We explain why in why AI projects need evals.
- Put the custom agent on one channel for two or three weeks and compare the share of conversations closed without a person, and the error count.
The Almaty Marathon support agent works this way: it closes about nine conversations in ten on its own and sends complaints, refunds and VIP requests straight to a person. If you need an agent for Bitrix24 open channels or a sales team, see how we build AI agents and AI for sales.
FAQ
Yes. On bitrix24.com it is CoPilot: call transcription and analysis, CRM field filling, summaries and text in chats and tasks. The Russian-language edition calls it BitrixGPT and adds an action-taking agent and an open-channel agent for customers.
The help center lists OpenAI's gpt-4o for text, chats and call analysis, DALL-E 3 for images, and separate speech services for transcription. Providers can change, and the CoPilot chat settings show which one is in use.
Yes, if WhatsApp is connected to an open channel. Both the built-in agent and a custom agent work through the open channel rather than the messenger itself. The messenger's own rules, such as WhatsApp's customer service window, still apply.
Yes. You can register your own service as a CoPilot engine with ai.engine.register, or build a separate agent that works through an open-channel chatbot and the REST API. The second route is the one to take when the agent needs data from other systems.
Bitrix24's on-premise pricing lists the CoPilot assistant in its editions. A custom agent connects to on-premise the same way, through the REST API.
Sources
- Bitrix24 Help Center, View the AI model used in a chat, updated October 6, 2026. Models behind CoPilot on bitrix24.com.
- Bitrix24 Help Center (Russian edition), AI models in Bitrix24, 2026. BitrixGPT as the default model in the Russian-language edition.
- Bitrix24 Help Center (Russian edition), AI agent for open channels, updated July 22, 2026. What the customer-facing agent does and its limits.
- Bitrix24 Help Center (Russian edition), AI request limits by zone, updated July 22, 2026. Request packages and boosts outside the Russian zone.
- Bitrix24 REST API documentation, Chatbots, 2026. Open-channel bots and handoff methods.
- Bitrix24 Help Center (Russian edition), Bitrix24 MCP server, updated October 16, 2026. Connecting outside AI tools to CRM and tasks.