---
title: "AI in procurement: what works today and where a person decides"
description: "What a language model does reliably in purchasing today, what belongs to plain code, and which steps stay with a person. With a table by task."
url: https://www.cureman.de/en/articles/ai-in-procurement
language: en
updated: 2026-11-17
alternate_de: https://www.cureman.de/de/artikel/ki-im-einkauf.md
---

# AI in procurement: what works today and where a person decides

> In procurement a language model is reliable at four things. It reads documents, it researches on the web and names its sources, it sorts text into classes, and it drafts messages. It is not reliable at arithmetic or at stating a price, so every figure belongs to plain code with a visible calculation. A person decides whatever has an effect outside the organisation, which means sending a message, confirming an offer a model read, changing something at a vendor and awarding. A setup that keeps these three roles apart can be checked. One that mixes them cannot.

AI in procurement works today where the task is reading, searching, sorting or writing, and where something checks the result afterwards. It does not work as a source of figures, and it should not act towards suppliers or employees without a person. The useful question for any task is which of three roles it belongs to: a model, plain code, or a person.

## Three roles

A language model is good with language. It turns a scanned invoice into fields, a supplier's free-text reply into an offer with items and prices, a web search into a list of candidates. Its weakness is the same everywhere: it produces what is plausible, and a plausible number is not a computed one.

Plain code is good with rules. Given the same inputs it returns the same result and can show how it got there: quantity times price, this invoice against the last one, this offer against the agreed price.

A person is needed where something leaves the organisation or commits it. A request to a supplier, a question to an employee, a removed seat and an award are decisions, and somebody has to own them.

## Task by task

| Task | Model | Code | Person |
| --- | --- | --- | --- |
| Read an invoice that arrives as PDF or image | Reads supplier, items, quantities, prices | Checks that lines, tax and total add up | Reviews what the arithmetic flags |
| Find duplicates, price increases, prices above an agreement | Nothing | Compares invoices with each other and with agreements | Decides whether to complain |
| Sort invoice lines into categories | Proposes a category from a fixed list | Adds up by category, supplier and item | Corrects a wrong category |
| Find alternative suppliers | Searches the web and names the page behind each candidate | Drops candidates without a source | Shortlists whom to ask |
| Write a request for quotation | Drafts the text around the items | Fills in items, quantities and deadline from the event | Approves and sends |
| Read a supplier's reply | Extracts the offer | Compares offers line by line | Confirms the offer before it counts |
| Ask employees about a licence | Drafts the question for the situation | Selects the seats by the inactivity rule | Approves the campaign |
| Fill in a form in a vendor portal | Operates the page step by step | Limits the hosts and records every step | Confirms before the submit |

## What works today

**Reading documents.** Invoices, offers and order confirmations arrive as PDFs, pictures and mail text. A model reads them into line items faster than anyone types them. What makes this safe is the check afterwards: if the lines do not add up to the total, the reading or the invoice is wrong, and either is worth a look. Structured e-invoices need no model at all, because their fields are already data.

**Research with sources.** Asking a model which suppliers sell an item gives a usable list when every name comes with the page it was found on. A candidate without a source is dropped. A candidate with one is still only a candidate until a person has looked.

**Sorting.** Is this line software, packaging or freight? Is this reply an offer, a question or a refusal? A model chooses from a list of labels. A label is not a number, so a wrong one is corrected and costs little.

**Drafting.** The request, the reminder, the counter-offer and the regret notice follow patterns. A model writes the text, and the figures in it are placeholders filled in from the data. How to structure the first of these is covered in [writing a request for quotation](https://www.cureman.de/en/articles/write-a-request-for-quotation).

**Operating websites.** Where a vendor offers no interface, an agent can click through the portal. This is the newest of the five and the one that needs the tightest rules: see [browser agents for vendor portals](https://www.cureman.de/en/articles/browser-agents-for-vendor-portals).

## What does not work

- **Asking the model for the saving.** It will answer. The answer has no calculation behind it and cannot be defended in front of finance. The reasons are set out in [why AI must not invent figures](https://www.cureman.de/en/articles/why-ai-must-not-invent-figures).
- **Benchmarks as findings.** A statement that companies of some size usually overpay is not a finding about your invoices. A saving exists when somebody quoted it or billed it.
- **Autonomous sending.** A message with a wrong price or an invented competing offer damages a supplier relationship in one mail. Nothing in the drafting step prevents that. The approval does.
- **Treating what the model read as fact.** A misread decimal point in an offer steers the whole comparison. Until a person confirmed it, a value read by a model is labelled as such.

## A worked example

The amounts here are invented and round. A company buys 1,000 cartons a year at 5.00 each. A model reads three replies to a request for quotation and extracts unit prices of 4.60, 4.80 and 5.10. A person opens the three replies and confirms that the prices were read correctly. Code then computes the comparison: 1,000 × (5.00 − 4.60) = 400.00 a year against today's price for the lowest offer. The model drafted the request and read the replies. It did not compute the 400.00, and it did not choose the supplier.

## Where a person decides

Five points stay with a person whatever the model can do: sending a message to a supplier or an employee, confirming an offer a model read, verifying a price that came from research, approving a change at a vendor, and awarding. How to design these approvals so that they are quick and still a real check is the subject of [human approval for AI agents](https://www.cureman.de/en/articles/human-approval-for-ai-agents). Putting everything that waits into one place is covered in [the procurement decision list](https://www.cureman.de/en/articles/procurement-decision-list). A company that already has its own assistant can give it the same tools, as described in [MCP in procurement](https://www.cureman.de/en/articles/mcp-in-procurement).

## How Cureman does it

Cureman divides the work along these three roles. A model reads invoices and supplier replies, researches prices and suppliers, labels requests and drafts text. What it read is stored as data, checked by arithmetic and labelled as read by the model until a person has reviewed it, and an offer read from a reply counts only after a person confirmed it. Every figure shown is computed by deterministic code with a visible calculation, and a draft that states a figure of its own is discarded. Researched prices are labelled as unverified until a person verifies them. Connectors to vendors only read. Nothing is sent to a supplier or an employee without a person approving it. The single exception is a reminder that repeats an approved request. The autopilot prepares the next step of a finding and never approves, sends or awards. An agent that works in a website enters only approved values and by default stops before a submit. The rules are listed under [principles](https://www.cureman.de/en/principles), and [how it works](https://www.cureman.de/en/how-it-works) shows the routine from data to a finding that is done.

## Questions and answers

### What can AI do in procurement today?

A language model can read invoices and supplier replies into structured fields, look for suppliers and prices on the web with the page each result rests on, sort invoice lines and replies into classes, and draft requests, reminders and questions to employees. All of that is preparation. The result is checked by code or by a person before it counts.

### What should AI not do in procurement?

It should not produce a figure that is shown as fact, and it should not act towards the outside on its own. A model that is asked for a saving returns a plausible number, not a computed one. A message, an award or a change at a vendor commits the organisation, so a person approves it.

### Can AI negotiate with suppliers by itself?

It can draft each round from a position that code computed, for example the lowest comparable offer and a target price. Sending the round is a decision of the buyer, because the message speaks for the organisation and a claim about other offers must be true. A draft that adds a figure or a claim of its own should be replaced by a fixed template.

### Does AI replace the buyer?

No. It takes over reading, searching, comparing and writing the first version, which is the larger part of the hours. The buyer keeps the choices, which are whom to ask, what to accept and when to stop. In a small team this means more categories get looked at, not that fewer people decide.

### Where should a company start with AI in procurement?

With documents it already has. Invoices show what is bought, from whom and at what price, and reading them needs no change at a supplier. The first findings come from comparing invoices with each other and with agreed prices, and the first approval a person gives is usually a question to a supplier about one of them.

## Read next

- [Human approval for AI agents in procurement](https://www.cureman.de/en/articles/human-approval-for-ai-agents): What an AI in procurement may do alone, what needs a person's approval, and how to design the approval point so that it is a real check.
- [Why AI must not invent figures in procurement](https://www.cureman.de/en/articles/why-ai-must-not-invent-figures): A language model writes plausible numbers, not computed ones. How to keep model-made figures out of savings reports, supplier messages and invoice data.
- [Browser agents for vendor portals: five rules](https://www.cureman.de/en/articles/browser-agents-for-vendor-portals): How to let an AI agent work in a vendor's or supplier's website. Approved values, allowed hosts, hidden logins, a stop before submit and a full record.
