IN Brief:
- Only 28% of surveyed food and drink businesses reported having a clear roadmap for integrating AI.
- Legacy IT was identified as a barrier by 38%, while 35% cited shortages of in-house expertise and skills.
- The underlying research surveyed more than 800 C-suite leaders across ten countries and 12 industries.
Argon & Co has found that only 28% of food and drink businesses in its survey have a clear roadmap for integrating artificial intelligence into their processes, compared with 39% across the industries examined. Legacy systems, shortages of suitable expertise, and weak data foundations continue to restrict deployment as demand becomes harder to forecast.
The sector analysis draws on the consultancy’s Operations Outlook 2026 research. More than 800 C-suite leaders were surveyed in October 2025 across the UK, France, Germany, the United States, Canada, Mexico, Australia, India, China, and the United Arab Emirates. The methodology states that respondents were split evenly across 12 industries, including food and beverage, although it does not disclose the exact number in each sector.
Among the food and drink respondents, 38% identified legacy IT as a major challenge to implementing AI, compared with 28% across all industries. A further 35% pointed to a lack of in-house expertise and skills, while 31% selected data quality, availability, or governance as the leading barrier.
Those constraints reinforce one another. A forecasting model trained on inconsistent product codes, incomplete yield records, unrecorded substitutions, or inaccurate inventory may produce a precise-looking result without a dependable operational basis. Recruiting data specialists cannot correct source information that remains fragmented across spreadsheets, enterprise systems, factory software, laboratory records, and retailer portals.
Food and drink demand is also becoming more difficult to interpret. Social-media trends can create sharp short-term spikes, while household budgets, reformulation, retailer promotions, direct-to-consumer sales, foodservice, online channels, and the growing use of GLP-1 medicines can alter what consumers buy and how frequently they purchase it.
Historical forecasts struggle when those influences overlap. A product can move from a predictable low-volume line to a national shortage before a conventional model has enough sales history to recognise the change. Equally, a viral surge may disappear before additional ingredients, packaging, or production capacity can be secured.
AI-assisted demand sensing can combine orders, promotions, search activity, external events, sales signals, and inventory positions. Its usefulness depends on the timeliness and consistency of those inputs, and on whether the information is available at the product, customer, and location level needed for an actual planning decision.
More data does not automatically create a better forecast. Timestamps, pack sizes, customer identifiers, units of measure, and product hierarchies must agree across the systems feeding the model. Where they do not, the project spends more time reconciling data than improving the production or purchasing decision it was intended to support.
Inventory management offers another practical application. A model can identify patterns in ingredient consumption, shelf life, supplier lead times, demand variation, and production loss, helping planners distinguish necessary safety stock from material that is accumulating without a credible route to use.
The recommendation still needs operational interpretation. Allergen controls, minimum order quantities, customer specifications, substitution limits, storage capacity, and supplier contracts may prevent the business from acting on an apparently efficient answer. AI can expose options, but it does not remove the rules governing the factory.
A controlled programme begins with a defined operating problem and a measurable baseline. Forecast error on a high-waste line, repeated failure on a critical asset, excessive giveaway, or poor schedule adherence can be assessed before and after a pilot. A broad instruction to introduce AI is harder to govern and easier to declare successful without changing production performance.
Ownership must also extend beyond the data team. Production, engineering, planning, procurement, technical, and commercial managers determine whether the model’s output is credible and what action follows. A warning that reaches nobody with authority to respond becomes another dashboard reviewed after the decision has passed.
Food safety and product integrity place firm boundaries around automation. A system may identify a cheaper ingredient, a longer campaign, or a shorter cleaning interval, but it cannot bypass allergen controls, validated hygiene procedures, legal composition, customer specifications, or hazard-management plans. Decisions with safety consequences need defined human approval and an auditable record.
Legacy equipment complicates integration because many plants operate reliable assets installed across several investment cycles. Production signals may remain in line controls, maintenance information in another application, quality results in laboratory files, and forecasts in the enterprise system. Connecting those sources securely can cost more than developing the initial model.
Wholesale replacement is not the only route. Companies can improve master data, agree common definitions, capture a limited set of dependable signals, and connect only the assets needed for the selected use case. A narrow system using trusted data is more useful than a broad platform populated with values that cannot be reconciled.
The full Operations Outlook report found that 51% of surveyed leaders were using AI and automation to drive efficiencies and reduce costs, while 36% named AI as a top five-year strategic priority. Executive interest is therefore ahead of operational readiness in much of the food sector.
The gap will close only when projects produce familiar manufacturing results: lower forecast error, less waste, fewer shortages, improved service, higher yield, reduced downtime, or quicker controlled decisions. Until those measures improve, an AI programme remains a technology deployment rather than an operating improvement.
