Four AI terms every food buyer should know

Between mid-September and the end of October, AI tools moved into several working steps of the food trade. Freight forwarders opened their booking systems to them, a card network began a pilot in which they pay supplier invoices, a large US grocery retailer set out how they can shop for its customers, a software maker for food businesses gave them supplier invoices to check, and the EU wrote them into the way customs will look for risky consignments. Four terms explain what is happening in each case: agent, MCP, price reliability and approval.

An AI agent is software allowed to take steps on its own

An AI agent is software that a person or a company allows to carry out steps on its own, such as searching, comparing, filling in a form or placing an order, instead of only answering a question. Since September, agents have been able to request a freight rate and book a container through the systems of two forwarders and a transport software company, and since 20 October a pilot has let them pay supplier invoices through the suppliers' payment portals.

The difference from an ordinary AI assistant lies in what the software is allowed to do. An assistant that summarises a price report leaves every decision to the reader, while an agent that books, pays or orders creates a commitment, which is why each case the column has followed comes with rules on what the agent may do.

MCP is the common connection between an AI tool and a company's system

MCP, the Model Context Protocol, is an open standard that lets an AI tool connect to another company's software and ask it for data or actions in a fixed format. It was first published in November 2024 and since December 2025 has been looked after by a foundation set up under the non-profit Linux Foundation. Because the standard is shared, a company builds one connection and every AI tool that supports MCP can use it.

In food trade the standard now carries freight bookings, benchmark prices for food commodities and the beta of a grocery retailer's gateway for shopping agents. In a food industry ERP launched on 28 October, the MCP connection lets AI tools read the business's data but not change it, with write access planned for later. The difference between reading and writing is the first thing to check in any MCP connection.

Price reliability depends on knowing which kind of price an agent returns

A question about "the price of Spanish extra virgin olive oil today" can be answered by four different figures: a benchmark assessed by a price reporting agency, an average of completed sales such as the one Spain's transaction-price system publishes, the European Commission's weekly national average, or one seller's offer. Each is built from different inputs over a different period, so on a given day they need not match, and a benchmark can also be corrected after publication.

The same question runs through the other cases. A freight cost from an agent is only useful with the rate type and the date, a US retailer states that agents see the same prices as people on its website, and an invoice agent that can schedule price changes alters the price the system expects on every later invoice. In each case the figure an agent returns needs its type, date and source before it goes into a negotiation or a contract.

Approval decides which step stays with a person

Approval answers who decides before an agent acts, and the cases followed so far place it in different spots. In the invoice payment pilot, agents pay only invoices a person has already approved. At the US retailer, the shopper approves every purchase by default or sets a spending limit, and in order intake tools the distributor chooses which customers' orders go through without a person.

Public authorities follow the same split. Under the EU's new customs code, a risk signal produced by software can lead to a result reached automatically or after human assessment, while the measures stay with customs authorities. The Commission's food fraud tool builds the trace map and leaves the decision to national inspectors. For payments, EU law counts a payment as authorised only if the payer has given consent in the form agreed with the bank, and it does not yet say who answers for a mistake made by an agent.

The four terms already appear at several stages of the food chain

Logistics: agents book freight through MCP, and the approval point differs from one booking system to another.
Selling and order intake: two in three EU food makers took no electronic orders in 2024, and tools that read orders sent by email, text or voicemail keep a person on the orders they flag.
Buying and payment: agents pay approved invoices in a pilot, and an invoice agent proposes corrections that need approval by default.
Retail: a grocery retailer offers per-purchase approval or spending limits to shopping agents, while another large online retailer blocks them.
Customs and food safety: EU customs risk analysis may use AI systems from 21 September 2027, and EU food inspectors already trace fraud with an AI tool.

Frequently asked questions

What is an AI agent in food trade?

It is software that a buyer, seller or service provider allows to take steps on its own, such as booking freight, placing an order, checking an invoice or paying one. The steps it may take are set by the company that uses it and by the system it connects to.

What does MCP change for a food company?

It lets one connection serve every AI tool that supports the standard, so a supplier, a forwarder or a price agency does not have to build a separate integration for each tool. A company should check whether the connection only reads data or can also change it, and who grants that access.

Can an AI agent pay a supplier without a person's approval?

In the pilot launched on 20 October, the agent pays only invoices the buyer has already approved and makes no payment decisions. EU payment rules count a payment as authorised only with the payer's consent in the agreed form, and the regulation that is to replace much of them was still awaiting final adoption on 20 October.

Do AI agents see the same prices as people?

One US grocery retailer states that the prices an agent sees on its website are the same as those a person sees, a claim that has not been checked independently. Outside a single shop, an agent can return an assessed benchmark, an average of completed sales, a public average or a seller's offer, and these need not match.

Will EU customs use AI to check food imports?

The new Union Customs Code, which applies from 21 September 2027, defines risk analysis as data processing that may use AI systems as defined in the EU's AI Act. The importer remains responsible for the data and for compliance with food and plant health rules, and a notified breach counts towards its risk profile.

How many EU food companies use AI today?

In 2025, 12% of EU food and beverage processors with 10 or more employees used at least one AI technology, according to the European Commission's Joint Research Centre. Among all agri-food firms, 2% used AI for accounting and 4% for business administration.