An AI agent now checks food supplier invoices

On 28 October a software company in Melbourne released a new version of its ERP, the system that runs sales, purchasing, stock, production, warehousing and deliveries in one place, which it builds for food manufacturers, wholesalers and retailers. The release adds three AI tools: an agent that reviews supplier invoices, an assistant that answers questions inside the ERP, and a connection through which outside AI assistants can read the ERP's data. An AI agent here is software that a business allows to carry out steps on its own, such as correcting a document or sending a request to a supplier.

The company's founder described the problem the tools address as the stream of small issues in a food business: "An invoice doesn't match what was ordered, a supplier changes their pricing."

The agent reviews supplier invoices and proposes corrections

According to the company, the agent reviews supplier invoices, investigates discrepancies and recommends what to do about them. Depending on the issue, it can update invoice lines, correct tax details, schedule price changes or request credits from the supplier. The agent is available only to selected customers in a private beta, and the company has not said how many businesses use it or what it has found so far.

The tasks are those an accounts team does when an invoice does not match the order or the delivery: a quantity that differs, a price that is not the agreed one, a tax rate entered wrongly. A scheduled price change goes further than a single invoice, since it alters the price the system expects on the next ones.

Every correction needs approval unless the business allows it to run alone

The company says the agent's actions require approval by default, and that a business can selectively allow automatic approval as it gains confidence in the system. The approval sits with the buyer's own staff, inside its own software, before the corrected invoice moves on.

The same split between the software's step and the person's decision runs through other tools covered in this column. In a payment pilot launched on 20 October, agents pay only supplier invoices that a person has already approved, and in order intake tools the distributor decides which customers' orders go through without a person. The invoice agent fits between the two, checking the invoice before anyone approves the payment.

AI assistants can read the ERP data through MCP but cannot change it yet

The release also adds a connection through MCP, the Model Context Protocol, an open standard that lets an AI tool connect to other software and ask it for data or actions in a fixed format. It is the same standard logistics companies used to open freight booking to AI agents in September and October. Through it, users can reach the ERP data they are authorised to see from supported AI assistants, which the company does not name.

At first the MCP connection is read-only, and write access is planned behind what the company calls a control plane, which manages permissions and the actions that are allowed. The assistant inside the ERP is also read-only in its first release, follows existing user permissions and cannot change operational records. Both are in beta.

Few EU agri-food firms use AI for accounting so far

In 2025, 2% of EU agri-food firms used AI for accounting and 4% for business administration, according to the European Commission's Joint Research Centre, which draws on Eurostat's survey of firms with 10 or more employees. In a policy brief, the JRC described the rise in AI use in the sector as "almost solely concentrated" on tools that analyse and write text, used for marketing, sales and administration.

Four points a buyer or supplier can check

Actions: which corrections software may make on its own, such as a tax detail, and which always need a person, such as a price change or a credit request.
Prices: the price list or contract the invoice is checked against, and who approves a scheduled price change before it applies to the invoices that follow.
Supplier side: what records support a credit request generated by software, so the supplier can match what was ordered, what was delivered and what was invoiced.
Access: which data an outside AI assistant can read through MCP, who grants that permission, and what changes when write access arrives.