Which tasks an LLM handles in a company
An LLM, a large language model, reads and writes text in ordinary language. For a company this means a simple thing: the model reads what an employee used to read and prepares a draft reply, document or summary.
Most often the model is put on five flows: incoming requests, support questions, emails, documents and reports for the manager. They have one thing in common: a lot of similar text and a clear rule for what counts as a good answer. More on how the model works is on a separate page.
Email. The Deutsche Telekom service sorts about 5,000 customer emails a week with a GPT model. 80% of emails fall into one of 60 categories by themselves and go to the right queue, saving about 2 million euros a year.
Reports. Axon’s Draft One service writes a draft police report from body camera footage, and the officer checks and signs it. The Microsoft blog cites an 82% cut in time spent on reports.
The model does not replace the accounting system and does not make decisions for the manager. It takes figures from the CRM and 1C and passes disputed cases to a person.
Such tasks are already common. According to the Economy chapter of the Stanford AI Index 2026, 70% of organizations use generative AI in at least one function.
Examples: what already works in companies
Below are open projects from our market cases section. They are not our clients: the figures come from publications by the companies themselves and their contractors, and each case page links to the source.
Support. Air India’s AI.g agent, a generative assistant, answers about 40,000 requests a day, and only 3% of questions go to a person. Since launch, the agent has held more than 13 million conversations and knows more than 1,300 different questions.
A knowledge base for staff. Verizon loaded into an assistant on the Gemini model almost 15,000 internal documents for a support team of 28,000 people. Agents find answers faster, and sales through the service team have grown by almost 40% since July 2024.
What these projects share: one process, their own knowledge base and a number before and after. A pilot on any model starts the same way, no matter which vendor stands behind it.
A model, an agent or a department: what to deploy
A model on its own answers questions in a chat window. Staff use it as a helper, but nothing changes in the CRM and 1C, and the result is hard to count in money.
An AI agent is a model plus tools and rules. It reads a request, checks stock in 1C, creates a deal in the CRM and passes a hard case to a person. An agent is put on one process and measured by one metric.
An AI department is 3–5 agents on shared data: sales, support, booking, reports for the manager. We build it in 4–6 weeks, from $4,999, connected to your systems.
The usual order: first one agent on a process with a clear metric, then a department. If there are many processes and it is unclear where to start, you first need a map: it shows where AI will return money and where it is not needed.
Integration with a CRM, 1C and a knowledge base
Integrating an LLM with corporate systems means the model does not live in a separate window. The agent gets access to the CRM, 1C, email and the knowledge base with rights only for its own process, and each of its actions is written to a log.
An agent in 1C reads stock, prices and debts, answers the customer with figures and creates an order or invoice by your rules. It does not change reference data and does not post documents without a check.
A request from a messenger becomes a deal in amoCRM or Bitrix24, with the conversation history in the card. The manager sees the whole conversation and steps in when the customer is ready to buy.
The model finds procedures, the price list and support answers by meaning-based search and answers from them with a link to the document. This approach is called RAG: the model is not retrained, but each question comes with the right fragments of your documents.
How it all works together is shown in our review of a wholesale company on open data: a reply to a request in a minute instead of the next day, stock from 1C, a deal in the CRM. This is a model calculation, not a client report.
Data and 152-FZ: cloud or your own model
If the model works with customer data, 152-FZ, Russia’s personal data law, applies. Under Part 5 of Article 18, recording, accumulating and storing personal data of Russian citizens in databases abroad is not allowed, except in cases the law lists separately.
So the main question is not which model is smarter but where the data goes. A cloud model abroad fits tasks without personal data: website texts, analysis of public documents, draft emails.
Where there are names, phone numbers and order history, the model runs inside your own network, or the flow is split. An anonymized request goes to the cloud, and personal data stays on your server. Read more about the legal requirements and about the hybrid scheme on the pages of the sovereign AI section.
Your own model is already a working option for large companies in Russia too. Severstal built its own 6 digital assistants for procurement on open language models: the speed of getting data and preparing documents grew 10 times.
What it costs and how to count payback
The cost of deployment has three parts: building the agent, the model fee and support. Our pilot of one agent costs from $1,499 for 14 days, and support from $480 a month.
Payback is counted from the process, not from the model. Take the staff hours spent on the process each month, multiply them by the cost of an hour with taxes and compare with support and the model cost. A step-by-step calculation with a formula is on the page about AI agent payback.
The model cost depends on the volume of text. If simple and hard requests come mixed, simple ones can go to a cheap model, and a strong one is kept for difficult cases: this is how model routing works.
Server and model prices change every month, so we do not list them here. Fresh calculations are in the calculators of the sovereign AI section.
We write the success metric into the pilot contract. If it is not reached, you see it in the report in numbers, not in impressions.
Where to start: a map and a pilot
If the process is already clear, start with a pilot: one agent, one process, 14 days, a metric in the contract. If you go further, the cost of the pilot is credited toward the price of the department.
If there are many processes and it is unclear where AI will return money, start with an AI solutions map. In 7 weeks we review the company process by process and count losses, solutions, the effect in money and risks, including 152-FZ. A map for the whole company costs from $93,000.
To choose the first process, answer three questions: where is there the most similar text, where is a mistake cheapest, and where is the result easiest to count. A process that passes all three is the candidate for a pilot.