What the agent does in procurement
A purchase goes through the same steps: a request from a department, a technical specification (a list of requirements for the goods or work), a request to suppliers, a comparison of bids and a decision. The agent is useful where people read and check documents.
The agent’s standard tasks: collect a request from a plant unit or a department by a template, check that all fields of the specification are filled in, send out the request, put the suppliers’ answers into a table and mark what is missing in them.
SIBUR has an example with numbers. According to a report of 15 September 2026, the agent checks the technical part of bids against the specification, finds incomplete and contradictory data and prepares a follow-up question to the supplier about each empty parameter by itself.
Before automation, such follow-up questions took up to 80% of the technical analysts’ time. The company runs about 8,000 procurement procedures a year with more than 2,000 suppliers, so the gain is visible.
Build it yourself: bid comparison
A simple version is built without a programmer. You need an agent builder, a model and a folder with the purchase documents: the specification, supplier bids in PDF and your contract requirements.
The agent works by the RAG principle: before answering, it searches for the right passages in the uploaded documents and answers from them. The model is not trained; it just gets your files.
A task for the first version: “Compare three bids with the specification, list the differences for each item and write questions to the suppliers”. The buyer checks the result. Such an agent works on copies of files and changes nothing in the accounting systems.
In its guide to agents, Anthropic advises starting with a simple solution and adding complexity only when the simple one is not enough. For procurement this means: first bid comparison, then integrations.
Integration with 1C and trading platforms
The agent starts saving hours when it takes data by itself: stock and purchase history from 1C, the Russian accounting system, requests from the document management system and bids from an electronic trading platform. Here you need a developer.
The connection goes through an API, the official channel for data exchange between programs, or through MCP. Anthropic opened MCP (Model Context Protocol) on 25 November 2024 as one standard way to connect AI to data and services.
Rights are given narrowly: reading reference lists and history, creating a draft request. The agent must not post documents or change prices in contracts. How such connections work is covered on the page about integration with 1C.
The agent first turns scans and attachments in supplier emails into text. How this works and where it makes mistakes is described on the document agent page.
Where the buyer decides
The agent prepares material for the decision, and the buyer decides. The SIBUR report talks about the first check and follow-up questions; it does not describe the choice of supplier as the agent’s work.
The choice of winner, price negotiations and the signature on the contract stay with a person. The agent can misread a table or miss a condition in an appendix, so every number in the comparison comes with a link to the page of the document.
Do not upload supplier bids and their prices to public chatbots. For such documents the agent works on a model deployed on your own server, or in a cloud under a data processing agreement.
If the agent gets the right to write to suppliers, in the first weeks every email goes out only after the buyer approves it. The agent gets the right to send without a check when the log shows that it writes without errors.
Metric and pilot
The agent’s effect in procurement is measured in time: how many hours the first check of one bid takes before and after launch. The second metric: how many follow-up questions go to suppliers on the first day, not a week later.
Measure before launch on 10 purchases from last month. Without it, a month later you have nothing to compare, and the debate about the agent’s value rests on feelings.
A standard bid comparison agent is built from open blocks. If someone charges for it as one-of-a-kind custom work, ask what the price is made of, and pay for a result by a metric, not for hours.
Our AI agent pilot runs for 14 days on one process and costs from $1,499. The metric is written in the contract before the start.