Integrating AI into a CRM, an ERP or 1C means a separate layer on top of the systems you already run: it reads data through an API, processes the text, the document or the employee's question with a model, and writes the result back into a record, a task or a chat. The software itself does not have to change, but the process you are automating is described and simplified before anything is connected.
Key takeaways
Connecting means three things: access to data through an API, processing by a model, and writing the result back where people work. Neither the CRM interface nor the structure of the 1C reference books changes — a new action appears: a button, a field or a message in a chat.
What changes is the process, not the software. An AI layer does not tidy up your data: if half the deals in the CRM have no stage and no owner, an assistant will answer imprecisely on that half and assemble a report with the same gaps. So before connecting anything you look at what is in the system and in what shape.
In a mobile app the layer works the same way, only it lives on the server side, while the screen gains meaning-based search, suggestions and photo analysis. That case is covered separately, in our article on AI features in a mobile app.
Which processes are worth handing to a model at all and which are covered by ordinary rules is set out in our article on business process automation with AI. This one is about the technical side: what to connect through, where to process the data and what the model can see.
The model has a different job in each system, and it is set by the data that sits there. In a CRM that is correspondence and deal history, in an ERP the numbers on orders and stock, in 1C documents and company details.
| System | What AI does | Example question from an employee |
|---|---|---|
| CRM | Sorts incoming enquiries by topic, condenses the history of contact into a short summary, drafts a reply and the next step | "What did we discuss with this customer over the past three months" |
| ERP | Answers questions on orders, stock and purchasing, spots deviations in the figures, drafts a comment on a report | "Why did purchasing on this item grow against last quarter" |
| 1C and accounting | Reads an incoming document, fills in the details and amounts, and checks them against the counterparty record | "Check the details on this invoice against the supplier record" |
| Knowledge base and policies | Answers from your own documents with a reference to the source and says when the materials hold no answer | "What are the return terms in a contract of this type" |
Technically all of this is API calls: the system hands over the data, the AI layer builds a request to the model, and the answer comes back and is written down. If a program has no open API, the data is pulled by a scheduled export; how systems exchange data between themselves, with no AI layer involved, is covered in our article on integrating a CRM with 1C, telephony and messengers.
The choice is driven by the data, not by fashion: if even anonymised text cannot go to an external service, the model goes on your own server; in every other case you start in the cloud.
| Criterion | Cloud model over an API | Your own model on your server |
|---|---|---|
| What it takes to start | An API key and a connection to the system | A server with a graphics card: gpt-oss-120b is released under the Apache 2.0 licence and fits into a single 80GB GPU |
| Which models are available | Including those distributed only as a cloud service, which cannot be installed on your own hardware | Open-weight models only; they usually fall behind cloud ones on long reasoning and rarer languages |
| Where the data ends up | On the provider's servers. At OpenAI, API data is not used for training by default, and abuse monitoring logs are retained for up to 30 days | Inside your own perimeter, nothing leaves it |
| What the price is made of | A charge per request, with no investment in hardware | Buying or renting graphics cards plus maintenance, after which requests cost almost nothing |
| Who it suits | Most tasks: support, documents, drafts, search across a knowledge base | Data that by law or by company policy does not leave the perimeter |
There is a middle option: anonymising on your side, where names, phone numbers and document numbers are replaced with identifiers before the request goes out and put back inside your own system. Only the text of the task then reaches the external service.
The law splits personal data into what must stay in the country and everything else. Article 27-1 of the law on personal data, as amended by law No. ZRU-1125 of 26 March 2026, requires biometric and genetic data of individuals, along with the data of users of telecommunications operators, to be held in databases on the territory of Uzbekistan.
Other personal data may be processed abroad, provided one of the conditions is met: the country is recognised as providing adequate protection, the operator has adopted standard contractual clauses or binding corporate rules, or it follows international data management standards. The list of such countries was approved by Cabinet of Ministers resolution No. 415 of 29 July 2026.
Important The question of which fields go to the cloud is settled before anything is connected, not after the first incident. The list of fields is written down, anonymisation happens on your side, and data covered by article 27-1 is processed inside the country.
The model reaches the system through a service account, so its rights are set separately from an employee's — otherwise the assistant answers from data a person is not supposed to see.
Roles and logs inside the CRM itself are set up on the same principles — that is covered in our article on access rights in a CRM. Decisions are a separate matter: law No. ZRU-1115 of 21 January 2026 prohibits relying exclusively on the conclusions of artificial intelligence systems when taking legally significant decisions that affect a person's rights and freedoms.
The word "train" covers three different things in conversation, and they cost very different amounts.
The practical difference shows up at update time. With a knowledge base a new policy is picked up by refreshing the index; with fine-tuning it needs another round of data preparation. How the base itself is assembled and what to do about contradictions between documents is covered in our article on AI assistants and chatbots.
It is built one scenario at a time: one input, one system, one measurable result.
How long each step takes depends on whether the systems have an API, on the shape the documents are in, and on who inside the company owns the doubtful cases.
The price splits into the connection work and the charge for requests to the model. The first part depends on the number of scenarios and systems, the second on how much text passes through the model each month.
Starting prices for the AI packages are listed on our AI and automation page, and those for exchange between systems on our ERP and integrations page. Your own model moves the cost into hardware: instead of a charge per request you get a server with a graphics card and its upkeep.
Syntra Systems starts with an inventory: which systems are already in place, what they expose through an API, where data is duplicated and in what shape the documents are. From that we assemble a diagram: which scenario connects to which system, where the data is processed and what the person does.
The model never gets access to the whole database: rights are granted by a list of objects, requests and answers go into a log, and doubtful cases have a route to an employee. What goes to an external service is agreed in writing before launch, together with the decision on what is anonymised on your side.
Tip Start with one scenario in one system and a measurable result: the share of answers needing no edits, the time per operation, the number of hand-overs to a human. A project that connects AI to the CRM, the ERP and the mobile app at once usually gets stuck on approving access.
Let’s discuss your project
Tell us what you need, and we will estimate the timeline and cost and suggest a solution.
Data is pulled through a scheduled export: the program saves it to a file, the AI processes it and writes the result back. That won't let you answer a customer the moment they get in touch, but it's enough for document processing and regular reports.
This is done when even anonymised text can't go to the cloud. It takes a server with a powerful graphics card: the open gpt-oss-120b model, which is licensed for self-hosting, fits on a single 80GB card. Open models lag behind cloud ones on long reasoning and rare languages, and hardware costs replace the per-request fee.
At OpenAI, data a company sends through the API isn't used to train models by default, and abuse-monitoring logs are kept for up to 30 days. Other providers set their own terms, so before connecting anything you read the contract and the data processing policy, not the marketing page.
Low-confidence cases are automatically handed to an employee — the scenario is built with a confidence threshold for exactly that, and every request and response is logged, so a mistake can be found and reviewed. Law No. ZRU-1115 doesn't allow decisions touching a person's rights and freedoms to rest on an AI's output alone — an employee has to confirm them.
A systems integration moves data from one program to another by set rules: an order from the CRM lands in 1C in the same fields. An AI layer does not move data, it processes it: turning an email into a structured request, a scan into fields, a conversation into text. The two jobs often run together, but they are different work.
Sources
