AI adoption in food businesses is accelerating, but readiness to deploy these systems safely and compliantly lags behind. A new guide from legal and regulatory experts warns that food companies cannot outsource their existing safety, traceability, and recall responsibilities to AI vendors, even as they experiment with tools for demand forecasting, inventory management, and camera-based shelf monitoring.
According to the report, many organizations introduce AI features without a clear strategy, layering new capabilities onto existing systems. Paul Armstrong, Director in the Commercial team at Walker Morris, noted that while workplace assistants and food-specific AI applications are gaining traction, the transition from pilot to production requires careful planning.
“Once people start to use and deploy AI in their business, adoption can often go through the roof really quickly,” Armstrong said. “But there’s often a disconnect around the governance work that you need to do internally to make sure that adoption isn’t uncontrolled and going to create problems.”
Start with the Problem, Not the Platform
Armstrong recommends beginning with a clearly defined operational problem before selecting or developing an AI solution. Common targets include excess short-dated stock, unreliable demand forecasts, time-consuming supplier document checks, slow traceability investigations, and repetitive assurance tasks.
“They’ve typically developed a really clear idea as to what they want to use the AI for before they’ve started,” Armstrong explained. “They’ve looked at their business and their processes and the potential bottlenecks and problems they have. And they’ve been really laser focused and said: ‘Based on what we know about AI and what it can do, we think this and this might be areas where AI could be deployed to improve the situation. And this is what improving it might look like.'”
Defining success metrics upfront helps distinguish genuine operational improvements from impressive but superficial demonstrations. A forecasting tool might be measured against waste reduction and manual intervention levels, while a traceability system could be evaluated on the speed and accuracy of batch identification. Inspection systems require testing for false negatives and edge cases alongside overall accuracy rates.
Building Pilots That Survive Production Conditions
Food operations present unique challenges that limited pilots often overlook. Variations in sites, products, seasons, suppliers, document formats, and working conditions can significantly impact AI performance once deployed at scale.
The FSA Science Council’s June 2026 report examined AI applications in manufactured-food risk assessment, certification data packs, pathology detection in abattoirs, and port document checks. It found that many systems adapted from other industries may lack development or validation for food-safety environments. The report recommends testing with real operational data and understanding each tool’s intended purpose and limitations.
Food-assurance data introduces additional complexity, often spread across scanned reports, certificates, sensor outputs, specifications, supplier portals, and handwritten logs. The Science Council emphasizes data quality, provenance, common terminology, and alignment with food-sector record-keeping practices as essential for reliable and auditable AI systems.
Armstrong cautioned against assuming AI features will operate seamlessly within familiar software platforms. Configuration, integration, testing, and user preparation still require significant investment of time and resources.
Related Post: UK food deficit hits record high amid security warnings
These economy-wide findings carry particular relevance for food manufacturers managing controlled processes, fragmented data, and multi-site operations.
A production-minded pilot must test the entire operating process, including data quality, systems integration, security, human review protocols, usage costs, supplier support, and downtime arrangements. Demand models should be challenged with seasonal peaks, promotions, disrupted supply chains, and incomplete data rather than assessed solely during ordinary trading conditions.
The pilot should conclude with a genuine go/no-go decision based on evidence that the system delivers meaningful benefits, operates within agreed tolerances, and remains safe and supportable under real-world conditions. Even technically successful systems struggle to scale when end-users are treated as an afterthought.
Keeping Human Judgment at the Center
The FSA Science Council emphasizes that food businesses must maintain clear lines of responsibility when using AI. Safety-assurance process owners should remain explicitly identified and competent, even when AI supports analysis or pattern recognition. AI can assist in interrogating large volumes of records, but final conclusions must rest with trained personnel.
Where personal data or automated decisions are involved, the ICO’s AI audit framework recommends standardised review procedures and agreed tolerances. Records of human challenges and overrides must be kept, along with provisions for manual or hybrid fallback when performance drops below acceptable levels. The degree of scrutiny should match the potential impact of an error.
Sampling or exception-based checks may suffice for lower-risk applications. Outputs affecting allergens, product release, authenticity, withdrawals, or recalls require closer review. Depending on the system, a competent person may need to examine every case to ensure compliance and safety.
Tracking AI Use Across the Business
Organisations need to know where AI is being used, who owns each application, and which controls apply. A central register should cover both dedicated AI products and functionality embedded within ERP, CRM, forecasting, and other established platforms. It should record each system’s purpose, owner, users, data, risk level, approved controls, and deployment status.
For higher-risk food applications, evidence should connect the complete decision chain. This includes risk assessment, validation results, accepted performance thresholds, human-review and escalation procedures, training records, supplier due diligence, system changes, and ongoing monitoring. Logs must distinguish AI-generated outputs from final human decisions, recording any challenges or overrides.
The Science Council identifies false-positive and false-negative rates, human override frequency, challenges, and drift alerts as relevant indicators. One-off validation is insufficient, since data, models, and operating conditions change over time. Internal policy must also address shadow AI, where employees use external tools on personal devices, creating visibility gaps for the organisation.