AI in the food industry: where does a manufacturer start? Quality, labelling, planning and waste

Where does a food manufacturer start with AI?

Not with a strategy programme and not with a new machine. Start with one work process that has three characteristics. It involves a lot of repetitive manual work. There is a standard that says what is right, such as a specification, a label text or a CCP limit. And the data already exists, on paper, in Excel or in a system. The sharper that standard, the better you can measure afterwards whether AI delivers anything.

In practice, the barrier is rarely the technology. According to provisional figures from CBS (Statistics Netherlands) for 2025, 16 percent of companies in manufacturing use AI. Of the companies that have considered AI but do not use it, 73 percent cite lack of experience as the main reason. For many companies, knowledge is therefore the first hurdle. That is why training is part of the implementation itself, for the people who run the process every day.

The rest of this article covers five processes. They are not equally suitable for every company. Choose the process where your manual work, your risk or your frustration is greatest right now.

How does AI help with quality control and labelling?

Labelling is a practical place to start, because the standard is fixed. Regulation (EU) No 1169/2011 requires information on fourteen allergens, and the NVWA (Netherlands Food and Consumer Product Safety Authority) publishes that list on its website. Article 21 of that Regulation requires the name of the allergen to stand out in the list of ingredients through its font, style or background colour. A wrong or mixed-up label therefore means that legally required information is missing or incorrect.

AI can help in two places here. At the desk, a language model compares a new or amended label text with the recipe and the specification, and flags differences for the employee who releases the label. On the line, a camera checks whether the right label is on the pack and whether the best-before and lot codes are legible. A camera does not see cross-contamination, a foreign object in the product or microbiology. Vision therefore does not replace metal detection or other control measures.

In visual inspection, a single accuracy percentage says little, because rejects are rare. Ask for the number of false rejects and false accepts, separately and for each defect category. How to validate that on your own line is explained in our article on computer vision on the packaging line.

• Step 1: choose one check, for example label release after a recipe change or the label check on one line.
• Step 2: write out the standard, with examples of right and wrong, agreed between quality and production.
• Step 3: record the baseline measurement: how many checks, how much time, how many deviations and how many complaints about this type of error.
• Step 4: let AI flag, not release. The employee assesses every flag and the release remains a human signature.
• Step 5: build re-verification into your change management, for every new label design, packaging material or new recipe.

What can AI do in production planning?

With short shelf lives, a planning error soon costs product. AI can produce a demand forecast for each item or customer based on your order history, and work through planning scenarios for changes in staffing, raw material supply or a rush order. The planner gets a proposal and the choice remains theirs. A model knows nothing of an agreement with a customer that is recorded nowhere, or of a line that has just had a breakdown.

In workforce planning, there is one line to watch. Annex III to the AI Act lists as high-risk the AI systems that allocate tasks based on individual behaviour or personal characteristics, or that monitor and evaluate the performance and behaviour of employees. A model that forecasts production volumes is not covered. A model that decides who gets which shift based on someone's performance may well be. Have that classification assessed in advance.

• Step 1: gather at least the order history, the promotions calendar and the master data for each item, and clean them up first.
• Step 2: let the model run alongside the planner for a number of weeks without its proposal being followed.
• Step 3: afterwards, compare both forecasts with actual sales, by item group.
• Step 4: only use the proposal once it demonstrably helps, and record who may overrule it and why.

How does AI help with maintenance and downtime?

In food, unplanned downtime in the middle of a production run costs more than lost production alone. Shelf life, cleaning and delivery commitments are all caught up in it. Predictive maintenance uses sensor data, such as vibration, temperature or power consumption, to spot sooner that a machine is heading for failure. That way, maintenance more often falls within a planned stop or a shift change.

This only works if the data exists and has been kept for long enough, with breakdowns recorded. If that is missing, the first step is the recording itself, not the model. The maintenance decision stays with your engineering department. For equipment around a critical control point, that decision also belongs in your HACCP system. What it takes and what you can realistically expect is explained in our article on predictive maintenance on sensor data.

How does AI make waste visible?

Waste often only becomes visible in the monthly figures. By then the cause is hard to trace. AI can link losses, rejects and leftover batches to line, shift, recipe, raw material batch or changeover time. A pattern then stands out while something can still be done about it. The model points to a correlation. Your people look for the cause and decide what changes.

Waste is also acquiring a European framework. The revised Waste Framework Directive, Directive (EU) 2025/1892, entered into force on 16 October 2025. Member States must reduce food waste in processing and manufacturing by 10 percent by 2030, compared with the annual average for 2021 to 2023 inclusive. That target applies to the Member State, not directly to your company. Member States have twenty months for transposition. How the Netherlands will translate the target for companies was not known at the time of writing.

• Step 1: agree what you count as waste: rejects, losses at changeover, overproduction, returns or leftover batches.
• Step 2: record by batch and by line, with a fixed reason from a short list instead of free text.
• Step 3: after a few weeks, let AI look for patterns, and discuss the outcome in the existing daily stand-up or weekly meeting.
• Step 4: measure the effect of each intervention against the baseline measurement, cause by cause.

Can AI prepare HACCP records?

Yes, provided a human continues to check and sign. Article 5 of Regulation (EC) No 852/2004 requires documents and records commensurate with the nature and size of the business, to demonstrate that the HACCP measures are actually being applied. Among the seven principles, the NVWA also lists: check regularly whether your HACCP approach is effective, and record how you comply with the HACCP rules.

AI can put measurements and checks that are currently retyped into fixed fields, flag missing records, make a trend towards a critical limit visible sooner and prepare an audit file in draft. The quality manager checks, completes and signs. At the CCP itself, the human decides. Why that boundary exists and how to set up the audit trail is explained in HACCP and AI. For traceability and the requirements of the scheme, see our article on FSSC 22000 and quality data.

• For every AI-supported step, record what the system proposed and on what basis.
• Record who checked it, who signed and when.
• Make sure a record can always be traced back to the source measurement.

What stays with the quality manager?

AI shifts work, not responsibility. The HACCP obligation lies with the food business operator, regardless of which tool organises the data. In practice, the following stays with the quality manager and the QA team:

• The decision at every critical control point, and the release or blocking of a batch.
• The release of labels and specifications, even when AI has prepared the comparison.
• The choice of what status an AI check is given in the food safety system: CCP, operational prerequisite programme or verification step.
• The operating point of a vision system: how many false rejects you accept in order to prevent missed defects.
• Re-verification after every change in product, packaging, lighting or software.
• The discussion with the auditor and the customer, and the reasoning behind every choice above.

What does AI cost in a food company?

An honest overview starts with what is not known. There is no reliable public price list for vision systems, sensors and planning software. The costs depend on your lines, your systems and the vendor. Ask for a quote with a trial on your own product and your own data. The amounts below are published figures, all excluding VAT:

• General AI tool for office work: according to Microsoft's Dutch pricing page on 11 October 2026, Microsoft 365 Copilot Business costs €18.20 per user per month on an annual subscription. It is an add-on to an eligible Microsoft 365 Business subscription, which you pay for separately. Anyone who signs up between 1 July and 31 December 2026 pays a promotional price of €15.60 for the first year.
• AIBA Practice Scan: free. A few questions and a report with the route that fits.
• AIBA Practice Day: from €2,500 per group. One work process at your own location, up to twelve participants.
• AIBA Ongoing guidance: from €1,750 per month.
• AIBA Online Learning Track: from €500 per person, six self-study modules.
• AIBA EU AI Act track: from €7,500, a separate compliance track.
• SLIM subsidy (in Dutch), the Dutch subsidy scheme for learning and development in SMEs: 60 percent of the eligible costs, up to €24,999 per SME application. SLIM covers a company review with a training plan and the development of a learning method, not training costs (Article 2.15, point (h)) and not costs arising from legal obligations (point (g)). The application windows for 2027 have not yet been published.
• Sector fund: the Stichting Ontwikkelingsfonds Levensmiddelenindustrie (SOL, the development fund for the Dutch food industry) mentions on its website an ESF+ (European Social Fund Plus) scheme with a 40 percent subsidy on training costs for employees up to and including mbo level 4 (Dutch senior secondary vocational education). Whether it is open, and until when, the site does not say. Check this with SOL before you count on it.

Which rules and developments are in play up to 2028?

The AI Act, Regulation (EU) 2024/1689, has been largely applicable since 2 August 2026. Article 4 on AI literacy has applied since 2 February 2025. Since the amendment by Regulation (EU) 2026/1744, in force from 27 July 2026, it has asked for measures that support the development of the AI literacy of staff. You do not have to guarantee a specific level. The requirements for high-risk systems under Annex III apply from 2 December 2027, and those for AI systems under Annex I from 2 August 2028.

A camera that checks product quality is not on the Annex III list. The same amendment added to Article 6 that AI used solely for non-safety-related quality control is not a safety component. AI whose failure would endanger health and safety does count as a safety component. For an allergen check, that question therefore deserves a serious assessment. AI in personnel decisions, by contrast, can fall under Annex III. Assess each application and record the conclusion in writing. How to approach that inventory is explained in our article on the EU AI Act for operations.

The labour market is changing too. On 3 July 2026, the Dutch minister Vijlbrief announced that from 1 March 2028 the meat sector will no longer be allowed to work with agency workers, with a transition period until 1 June 2028. The ban applies to slaughtering, processing, chilling and freezing meat. Companies with fewer than 20 people are exempt. This is a proposed measure. According to the announcement, it first went to online public consultation and will then be laid before the Tweede Kamer (Dutch House of Representatives) in draft. For companies in that chain, this changes how staff are deployed. AI does not solve that. It can, however, lighten the load of record-keeping and administration, so that permanent employees have more time left for the job itself.

How do you tackle the first process?

The sequence is the same for each of the five processes. It is deliberately small, because a pilot that runs outside the quality system often gets switched off again at the first audit.

• Choose one process and describe in two sentences what currently goes wrong and for whom.
• Appoint an owner in the line organisation, by name: the planner, the QA manager or the team leader who runs the process.
• Record the baseline measurement: time per task, number of errors or corrections and the time that checking takes.
• Train everyone who runs the process, including the second shift and the stand-ins, for example with in-company AI training, and record who has been trained in what.
• Write the AI step into your quality system: work instruction, verification, records and who decides and signs.
• Evaluate against the baseline measurement at a point agreed in advance: continue, adjust or stop.

How AIBA approaches this

On a Practice Day, we build one work process from your own factory, at your own location, with the people who run that process every day. The team learns it while it is being built. We deliberately choose a process for which the data is good enough, and set up the audit trail from the first step. We record the working agreements in the language of your standard: HACCP, FSSC 22000, BRC or IFS. It helps if your quality manager is there. What that looks like in a factory, from scan to a working workflow with an audit trail, is set out on our page on AI implementation in food manufacturing (in Dutch).