AI agent in CRM: how Bitrix24 and amoCRM automate the work with the customer's card
When does CRM already know how to use AI, and when does it need a separate agent
When they say "let's implement AI in CRM," they often overlook the main thing: Bitrix24 and amoCRM already have built-in AI, and it's enough for many tasks. He summarizes calls, helps fill out fields, generates texts, analyzes correspondence, sets tasks and moves transactions according to clear rules. If the whole process lives inside the CRM and the data is already there, this is the fastest and cheapest option.
But as soon as the process goes beyond the boundaries of CRM, the built-in functionality becomes insufficient. The client writes to the messenger before creating the transaction, the price must be checked with 1C, the availability must be checked with the warehouse, the status must be checked with the internal service, and personal data cannot be transferred to the external contour. In such scenarios, you need not just an "AI in CRM", but an agent who can work on top of CRM and other systems.
A simple rule:
Briefly about the approach: not instead of CRM, but on top of the process
An AI agent inside CRM is not a replacement for Bitrix24, amoCRM, or sales managers. With the right architecture, it becomes a working layer between communication channels, CRM, accounting system, warehouse, telephony, and the company's internal regulations. He receives the event, understands the context, accesses the allowed tools and returns the result to the client's card or transaction.
What Bitrix24 can already do
BitrixGPT
A separate block includes sales scripts and speech analytics. The system can evaluate script compliance, register objections, and help the sales manager see not only the outcome of the transaction, but also the quality of the dialogue. For teams with dozens of calls per day, even a simple summary saves hours of listening.
BitrixGPT Follow-up
According to the models, Bitrix24 uses its own BitrixGPT and BitrixAudio for speech recognition, as well as YandexGPT and GigaChat. Other models can be connected via applications from the Market. This is an important detail: in some scenarios, a company does not need an abstract "neural network", but a specific model due to quality, cost, data processing policy, or security requirements.
According to the cost for the Russian market for September 2026, the open materials indicate cloud tariffs of 1,000, 3,000, 6,000 and 13,600 rubles per month, the boxed version starting from 19,500 rubles. For BitrixGPT to work in a box, a platform of at least 24.700.0 is required. Since 04/15/2026, the cost for boxed versions has changed; in Russia, a 50% discount is valid until 04/15/2027. Before publishing and calculating the project, the tariffs need to be checked: for CRM platforms, prices and the composition of functions change several times a year.
What amoCRM can already do
amoAI"Professional" and higher
amoAI analyzes the correspondence and the history of the transaction, after which it can perform actions in CRM: set tasks, change the responsible person, move the transaction through the funnel, launch Salesbot, put tags and fill in fields. This is well suited for situations where communication has already entered amoCRM and further actions need to be put in order.
The setup is built around four blocks: the data source, what exactly to analyze, what action to perform, and under what condition. This approach is understandable to the sales manager and the CRM administrator: you don't have to write the code, but assemble the rule in the interface. Plus, there is a test mode — you can run the script on existing transactions and see the result before engaging in combat.
The models used are OpenAI and YandexGPT. Speech analytics is not normally included in the basic suite and is usually connected via widgets: Voice AI, Speech2Text, NOVA and other solutions. For companies with high demands on personal data, it is important to look at exactly where the transcription of calls takes place: some solutions offer local processing, and this can be a decisive factor.
Where is the built-in AI enough, and where is not
The main mistake when choosing a solution is trying to buy the "smartest AI" instead of decomposing the process into specific tasks. The built-in AI in CRM is good where the event, data, and action are all within the same system. A separate agent is needed where external sources, non-standard logic, multiple systems, and data processing loop requirements appear.
| Task | The built-in AI is enough | We need a separate agent |
|---|---|---|
| Summary of a call that has already been recorded in CRM | ✓ | |
| Filling in fields by correspondence within CRM | ✓ | |
| Generating a letter or a commercial offer based on a template | ✓ | |
| Assessment of compliance with the sales script | ✓ | |
| The first response to the client in the messenger at night, before the transaction is created | ✓ | |
| Qualification of the application with verification of data in external systems: 1C, warehouse, registers | ✓ | |
| Outgoing call with a full conversation | ✓ | |
| Working in channels that are not available in CRM | ✓ | |
| Actions based on rules that are not available in the settings interface | ✓ | |
| A single agent on top of CRM and other systems | ✓ | |
| Placing data processing in its own contour | ✓ |
For example, if a manager needs to receive a summary of a conversation that is already recorded in a transaction card, an integrated AI would be a natural choice. If the client wrote in Telegram at night, the agent must clarify the need, check the availability of goods in stock, create a lead, assign a task to the manager and attach a resume — the process already goes beyond the CRM alone.
The same applies to companies with complex sales logic. You can set up many rules in the CRM interface, but not all of them. Sometimes it is necessary to take into account marginality, region, balances, legal status of the client, payment history and logistics restrictions. An embedded tool can be part of the process, but it does not always become the center of decision-making.
How the agent connects to CRM
a set of integrations and access rights
In Bitrix24, incoming webhooks are available for integration: the agent can call REST API methods, for example, create a lead, find a deal, add a comment, or set a task. Outgoing webhooks work in the opposite direction: CRM notifies the agent of an event — a new request has appeared, the stage has changed, and a responsible person has been appointed.
fetch
In amoCRM, the basic bundle is built through the REST API, event webhooks, Salesbot for chat scripts, and widgets for front-end scripts. This is enough for the agent to receive the transaction event, analyze the correspondence, access the external system and return the result to the card.
The rule of thumb for both systems is:
Secure architecture is usually based on the principle of minimum necessary rights. If the agent only needs to read the correspondence and add a comment, he does not need the rights to delete transactions or change financial fields. If it can change the stage, then only in pre-described cases. Irreversible actions such as deleting data, changing the transaction amount, or transferring to a stage that affects premiums are best left to the person or require confirmation.
Working with the client's card: what is really automated
The customer card seems to be a simple entity: name, phone number, company, source, need, comment. But it is she who determines the quality of reporting, the reaction speed and the ability of the sales department to work with the customer's story. If the card is empty or filled out randomly, even a strong model will draw conclusions in the sand.
filling in the fields from the explicitly stated
The second scenario is data normalization. The phone number is reduced to a single format, the INN is checked against the checksum, the city is compared with the directory, the company name is cleared of unnecessary characters. It is important to understand: this is often done not by the model, but by regular code and reference books. The neural network helps to extract data from the text, but it is better to check and normalize them using deterministic rules.
The summary of the dialogue works well: the agent writes down in the field or comment what the client wanted, what objections were raised, what was agreed upon and what next step is needed. The manager doesn't have to read all the correspondence, and the manager gets into the context faster after the weekend or vacation.
You need to be more careful about classifying and evaluating the probability of a deal. Classification by topic, stage, or source works if the categories are described in advance and there are examples. Estimating the probability of a deal can be a useful clue, but it should not immediately be turned into a field that affects reporting. The model can pick up on the tone of the conversation, but it doesn't know your seasonality, marginality, customer characteristics, or actual payment history.
The assignment of tasks and next steps is also automated, but requires escalation rules. If the client requests an urgent payment, the task is assigned to the manager. If the budget is named above the threshold, notify the manager. If there is a legal issue in the correspondence, transfer it to the responsible specialist. The clearer the rules, the less magic there is and the higher the confidence in the outcome.
What you should not give to the agent without confirmation: changing the transaction amount, deleting data, massively changing the stages affecting premiums, and any actions that are difficult to roll back. The agent should speed up the work, and not become an invisible participant in the reporting, which then cannot be verified.
The main reason for failures is not AI, but data.
The most frustrating discovery in CRM AI projects is that the problem is often not with the model. The model may be strong enough, the integration may be working, and the scenario may be reasonable, but the result will still disappoint if the CRM data is chaotic.
60% of AI projects will be abandoned by 2026 due to lack of AI-ready data.
70% of respondents say that bad data undermines the results of AI
If half of the transactions have empty fields, the agent filling out the cards can make the statistics more beautiful, but not necessarily more accurate. If one client has three cards, the agent risks responding based on a random story. If the funnel stages do not correspond to the actual sales process, automatic stage transitions will not improve management, but will break reporting.
A good example: a company wants an agent to evaluate the likelihood of a deal. But there is no single "need" field in CRM, the sources of requests are recorded differently, and the stages are used as reminders to the manager, rather than as a reflection of the client's path. In such a situation, the agent will simulate analytics. First, you need to agree on the data rules, and only then connect forecasts.
What is happening in the world
Large vendors are actively moving the market towards agent-based scenarios: CRM becomes not only a transaction base, but also an operating environment where AI suggests actions, performs routine and takes over part of communications.
Salesforce claims annual revenue of about $1 billion for Agentforce, processing of about 4 million requests, and an average payback for customers in about 8 months. It is important to read these numbers correctly.: these are the statements of the company itself, and not an independent universal result for any business.
HubSpot reports that more than 55% of customers use Breeze features on Pro and Enterprise tariffs. This shows the direction of the market: embedded AI tools are becoming the norm for CRM. But at the same time, there is a growing demand for solutions that run on top of multiple systems and are tied not to the user's license, but to the result of the process.
There is also a shift in payment. The market is gradually moving away from a simple "pay per user" model to paying for a result, action, or volume of processed transactions. This is logical for a business: the head does not buy a model or a beautiful button, but a reduction in reaction time, an increase in data quality, a reduction in manual routine and a more manageable sales process.
What to do next
If you already have Bitrix24 or amoCRM, start with the process. Where are the applications lost? Which fields do managers not fill in? What actions are repeated every day? Where is the client waiting too long for a response? What data do I have to check in external systems?
If the answers are inside the CRM — calls are already recorded, correspondence is connected, transactions are conducted carefully, and resumes, texts, hints and field filling are needed — the built-in functions may be enough. This is a normal result. Sometimes the best AI project is not the development of a new agent, but the competent inclusion of what has already been paid for in the tariff.
If the process goes outside — messengers before the transaction is created, verification in 1C, warehouse, registries, special qualification rules, requirements for its own data processing circuit — then a separate agent becomes justified. It does not replace CRM, but connects it to the actual sales process.
A practical next step:in what range will it cost
Strong automation starts with an honest fork in the road: sometimes you really don't need to buy anything. And sometimes it is a separate AI agent that becomes the layer that turns CRM from an archive of transactions into a living system of actions.
Автор
Антон Амосов
