Europe is investing in shared data, digital twins and artificial intelligence for food production. Their practical value depends on whether information survives the journey between organisations and supports decisions that users can check.
A temperature reading, a delivery record and a supplier certificate can describe different parts of the same food shipment. Bringing them together should help a processor understand what arrived, where it came from and how it was handled. Yet a useful answer depends on more than collecting information. Records need compatible meanings, reliable links to the physical product and clear rules about who can use them.
European research is addressing these problems through shared information models, digital representations of supply chains and facilities for testing artificial intelligence. Taken together, these initiatives suggest a practical way to judge digital food innovation: follow one operational decision from the original measurement to the person who acts on it. Every transfer along that route needs to preserve enough context for the decision to remain defensible.
Start with the people supplying the data
A Joint Research Centre study published in 2025 provides a useful corrective to assumptions about fully automated agriculture. Its survey covered 1,444 respondents in nine EU Member States, with data collected between June and October 2024. General IT tools were widely used, but more expensive technologies designed specifically for crop or livestock production had lower adoption. High costs and limited skills remained barriers.
The report also found that farm records were often collected manually or using basic digital tools. Farmers were selective about sharing, with concerns about privacy, security and control. These survey findings describe the participating farms; they are not a direct measurement of the performance of every European food supply chain.
For processors requesting more supplier information, the implication is practical. An additional reporting requirement may depend on someone finding an old record or typing the same figure into another portal. A useful digital service should therefore reduce avoidable work at the point of collection, while making corrections possible when a record is wrong.
Give information a common meaning
The EU-funded DEMETER project developed an Agricultural Information Model to support exchange between systems and organisations. Its final report describes a common model built on existing standards, alongside tools for connecting services and handling data securely. The project ran from 2019 to 2023.
This addresses semantic interoperability: the ability of different systems to understand the same information consistently. For example, a quantity needs a unit, and a measurement needs a time and an identifiable subject. Moving a number successfully between databases is of limited use if its meaning changes along the way.
DEMETER also developed tools for comparing farm performance and worked on data-sharing agreements. These are useful building blocks, although a functioning exchange mechanism does not establish that every participating business achieved savings. Buyers assessing a new system could test how it handles an unfamiliar supplier, a changed product identifier or a correction to historical data. Those situations reveal more than a demonstration using carefully prepared records.
Digital twins need boundaries and baselines
A digital twin represents a physical product, process or supply chain in software, using data to examine its condition or possible changes. The BBTWINS project applied this approach to meat production in Spain and fruit production in Greece. Its final report describes models for activities including ham drying, fruit storage and logistics, combined in a platform for decision support.
The project, which ended in May 2025, reports a 28% reduction in mileage in its real-world logistics work. This is a project-reported result from particular applications, rather than a saving that another operator can assume. Transferring it requires knowing the original routes, loads and constraints, and which operational changes produced the improvement.
A useful trial would compare the proposed system with the operator’s existing planning method over equivalent workloads. It should record both the predicted improvement and the outcome after staff act on the recommendation. If the model suggests a shorter route that cannot meet a delivery window, the prediction has not yet become a business benefit.
Traceability depends on the original evidence
Connecting Food’s DIGI-TRUSTY project offers another application. A CORDIS account published in December 2025 describes a platform that gathers supplier certificates, questionnaires and other records, converts them into a standard format and links them to a digital representation of the supply chain. It reported more than 4,000 suppliers in 23 countries using the platform.
The account describes blockchain records of who submitted information and when. That can help establish a record’s history. The separate question is whether the submitted information accurately describes the ingredient or batch. A durable record of an incorrect entry remains incorrect until it is identified and addressed.
For a food business, the useful questions concern evidence behind the record: how a certificate connects to a delivery, how mixed batches are represented, and whether an expired document is flagged. Supplier participation demonstrates reach. Verification needs an examination of the information itself and the process used to resolve discrepancies.
Test artificial intelligence under operating conditions
Europe’s agrifoodTEF network provides facilities and expertise for testing AI and robotics in agriculture and food. Its services include trials on farms and in facilities, checks on robustness and interoperability, user testing, and business assessments. The network describes a process in which a test plan is developed for the particular technology and intended use.
That specificity matters. A visual inspection tool could be tested against different lighting, product varieties and line speeds, with errors separated by their consequences. Rejecting an acceptable item and accepting a defective one impose different costs. An overall accuracy figure can hide the distinction, so evaluation should report the outcomes that matter to the operator.
Human checking also remains part of institutional experimentation. In a March 2026 EFSA discussion, executive director Nikolaus Kriz described the use of AutoCAT to help identify relevant research, while retaining human validation. This is an example of AI supporting evidence review within a defined task, rather than taking responsibility for a complete food-safety assessment.
Make sharing workable for smaller suppliers
The Commission’s Common European Data Spaces initiative combines technical infrastructure with arrangements for access and use. Its stated principles include transparent access rules, privacy protection and the ability to share data either without charge or for compensation. Agriculture is among the sectors covered.
The governance arrangements deserve as much attention as the software. A supplier should be able to understand the requested data, the purpose of sharing and the practical route for correcting errors. Interoperability also needs to work when a business changes software provider. Otherwise, a new reporting platform could create another dependency instead of simplifying an existing process.
For researchers, a useful next study would measure the time required to supply and reconcile data, including work carried out by smaller firms. For buyers, a useful acceptance test would follow a batch across several independently operated systems and deliberately introduce a correction. The result would show whether the chain can recover an accurate account without rebuilding every record manually.
Procurement can make those expectations concrete. A buyer could ask a provider to demonstrate an export of the underlying records, explain the support available after installation and identify the ongoing work required from suppliers. Those questions also improve cost comparisons: the subscription price alone will not describe a service that requires extensive manual reconciliation. Agreeing who maintains shared identifiers and who investigates conflicting records would help prevent routine problems from being passed between organisations without resolution.
Digital food systems will be judged through these ordinary tasks. A recommendation needs a traceable basis; a supplier needs a manageable way to contribute; and the person making the final decision needs to understand what the system has established. Better connections between those responsibilities would make the growing volume of food data more useful.
