MMOOZZGG.com

How to build an AI agent

AI agent for manufacturing: which plant tasks to start with

5 min read9 sources

An AI agent for manufacturing takes over the office work of a plant: customer requests, purchasing, reports from 1C, a Russian accounting and ERP system, and answers from technical documents. By 2032, Russian industry needs to attract more than 1.5 million people, so plants keep people for the shop floor and give office routine to agents.

On this page · 5 sections

A developer is needed

Where an agent works at a plant

An agent has the easiest time in the plant office: the work there is emails, spreadsheets and documents. AI also works on the shop floor, but those are other systems, with cameras and sensors; more on them below.

Five tasks plants usually start with: replying to a request and drafting a commercial quote from the price list, checking supplier offers, telling customers the order status from 1C, a morning report for the director, and search in technical documents.

Two of them are covered in examples with numbers: the agent for requests and quotes and the morning report for the plant director. We described checking supplier offers on the purchasing agent page.

Start with one task. In its guide to agents, Anthropic advises starting with a simple solution and adding complexity only when it is not enough.

Agents and quality control: different tasks

Quality control on the line is done by computer vision systems: cameras and models that find defects in images. This is not an agent that writes messages and works with documents, but a separate class of AI with its own equipment.

Example: Severstal reported on 17 February 2026 that digital quality control with AI covers about 60% of the marketable output of the Cherepovets Steel Mill. The system recommends a decision on whether the rolled steel is fit, and an employee makes the decision.

Robotization in Russia is slow. According to the Ministry of Industry and Trade, the industrial robot fleet grew by 10.9% in 2025, to 21,857 units, while density stayed at 29 robots per 10,000 workers.

An office agent needs no equipment, and it is launched in weeks, not years. So it often pays for a plant to start with it and run shop floor projects separately, with engineers and their own budget. How AI works on the shop floor is covered on the page on AI in industry.

Answers from technical documents

A plant has thousands of pages of documents: process sheets, equipment manuals, standards, internal rules. A foreman or process engineer searches them for an answer by hand or calls a colleague who remembers.

An agent with a knowledge base answers questions such as what is the tolerance for this part or what is the line start-up procedure from your documents and shows the page the answer came from. It works on the RAG principle: first it finds fragments in the documents, then it answers from them.

A mistake in such a reference is costly, so the agent answers only with a link to a document and passes questions with no answer in the base to a process engineer. In the first weeks, a person checks the answers on a set of test questions.

Drawings and process know-how are often trade secrets. For them the agent works on an open model inside the plant network: Ollama runs models on your server and is distributed under the MIT license.

Data from 1C and accounting systems

Order status for a customer, a shipment summary, raw material stock: this data sits in 1C and other accounting systems. For the agent to fetch it by itself, you need a developer and a connection to the system.

The connection goes through an API, that is, an official channel for data exchange between programs, or through MCP. This is a protocol that Anthropic opened on 25 November 2024 to connect AI to data and services in one standard way.

The agent gets rights only to read and to make drafts. It must not change orders, prices or production plans; people decide those. How the connection to 1C works is described on the integration page.

We launch such an agent as a pilot on one process: 14 days, from $1,499, with the metric written into the contract.

Where a plant should start

Start by counting: where in the plant office do most hours go to reading and moving data? Usually it is requests and quotes, purchasing and reports for management.

Choose one task and one metric: the response time to a request, the number of requests without a reply, the hours spent on the morning report. Measure it before the launch, or in a month you will have nothing to compare with.

Plants lack people. According to Denis Manturov at a board meeting of the Ministry of Industry and Trade on 7 July 2026, industry needs to attract more than 1.5 million people by 2032. An office agent does not replace a worker on the shop floor, but it frees the hours of those already hired.

A typical agent for requests is built from open blocks. If you are asked to pay for it as a unique development, ask what the price consists of. All our solutions for plants are collected on the page for manufacturing companies.

Questions and answers

Can an agent run at a plant with no internet?

Yes. Open models are installed on a server inside the plant network, for example through Ollama, which has an MIT license. Qwen 3 models come under the Apache 2.0 license and DeepSeek R1 under MIT, so you can run them in-house. A server for the model costs money, so it is calculated before the start.

How much does an agent pilot in manufacturing cost?

Our AI agent pilot on one process costs from $1,499 and takes 14 days. We write the metric, for example the response time to a request, into the contract before the start. If the agent needs its own server, we calculate its cost separately.

Sources and fact-checking

Checked by the editors on 16 October 2026.

  1. PRIME, 7 July 2026 (1prime.ru)
  2. Kommersant, 25 August 2026 (kommersant.ru)
  3. Severstal, 17 February 2026 (severstal.com)
  4. Anthropic, 19 December 2024 (anthropic.com)
  5. Alpine Linux packages, 2026 (pkgs.alpinelinux.org)
  6. Anthropic, 25 November 2024 (anthropic.com)
  7. Simon Willison, 20 January 2025 (simonwillison.net)
  8. Simon Willison, 29 April 2025 (simonwillison.net)
  9. arXiv 2005.11401, 22 May 2020 (arxiv.org)

Found a mistake? Tell us

Publications

Contact us