How AI forecasting can boost resilience in volatile food supply chains

Artificial intelligence (AI) could help agri-food companies forecast demand more accurately, reduce waste and build more resilient supply chains, but the technology will deliver limited value unless firms treat it as an organisational transformation rather than a software upgrade, according to a systematic review by researchers from the Agricultural University of Athens.

The study, titled AI-Driven Demand Planning: A Systematic Review of Adoption, Barriers and Strategic Implications and published in Administrative Sciences, reviewed 37 peer-reviewed studies from 2015 to 2025 and found that AI-driven demand planning can improve forecasting accuracy while also exposing persistent gaps in data governance, workforce skills, change management and technology access, especially among small and medium-sized agri-food enterprises.

AI models improve forecasting accuracy in volatile food supply chains

Demand planning has become a strategic pressure point for agri-food organisations as they face climate instability, geopolitical shocks, supply chain disruption, shifting consumer preferences and rising sustainability demands. The sector is especially vulnerable because agricultural products are perishable, production is seasonal, supply depends on biological cycles and weather, and profit margins are often thin.

Traditional forecasting tools, including moving averages, exponential smoothing and conventional time-series models, often struggle to capture the non-linear patterns that shape agri-food demand. A small error in forecasting can lead to excess inventory, spoilage, stockouts or missed sales. In a sector where products may lose value quickly, inaccurate demand planning is not only a financial problem but also an environmental one.

The review found that AI and machine learning models are increasingly being used to address these weaknesses. Advanced tools such as long short-term memory networks, gated recurrent units, random forests, gradient boosting, hybrid optimisation systems and deep learning models can process complex datasets and identify patterns that traditional methods often miss.

Across the case studies, the review found that advanced AI tools can reduce mean absolute percentage error by 20% to 40% compared with traditional statistical methods. The improvement has direct consequences for inventory control, logistics performance, food waste reduction and supply chain governance. The authors frame this shift as part of a broader move toward Supply Chain 4.0, where predictive and prescriptive analytics replace reactive planning.

The clearest evidence emerged from research on core forecasting models. These included research comparing machine learning and deep learning models with traditional methods, as well as studies testing hybrid architectures that combine multiple techniques. The review found that no single algorithm is universally superior. Model choice depends on demand stability, product type, data availability and organisational readiness.

Long short-term memory models performed well in contexts involving price volatility and complex temporal dependencies. Tree-based ensemble models, including random forests and gradient boosting, showed strength in some retail and stable-demand settings. Hybrid models, including systems that combine decomposition methods and intelligent optimisation, were increasingly used for volatile agricultural commodities.

The review also highlights the importance of external data. Weather conditions, seasonality, consumer trends, promotions, search activity, regional economic signals and market disruptions can improve predictive accuracy when properly integrated. In agri-food planning, demand is rarely shaped by sales history alone. AI systems perform better when they can combine internal company data with wider signals from the market and environment.

However, the authors warn that higher accuracy does not automatically lead to operational success. More complex models can be harder to explain, maintain and integrate into existing planning systems. For managers, the technical performance of an algorithm must be balanced against interpretability, cost, compatibility with legacy systems and the ability of staff to act on its recommendations.

Better forecasts can reduce inventory costs and food waste

The review found that improved forecasting accuracy can translate into measurable operational benefits, especially in inventory optimisation and waste reduction. Several studies linked AI-enabled planning to better replenishment decisions, improved stock control and lower waste in perishable supply chains.

The review notes that inadequate demand management is a major contributor to food losses before consumption, especially for products with short shelf lives such as fresh fish, dairy, vegetables and other perishable goods. AI forecasting can help firms align procurement, production, storage and distribution more closely with expected demand.

Some studies reviewed by the authors showed that AI-enabled planning can support a 15% to 20% reduction in inventory costs. This matters for agri-food firms that operate with narrow margins and high exposure to waste. Better forecasts can reduce overstocking, lower holding costs, improve product availability and limit unnecessary disposal.

The operational value of AI was especially clear in perishable inventory systems. Deep reinforcement learning, decision support systems, neural network controls and ensemble models were used to manage products with fixed shelf lives, uncertain demand and changing storage conditions. In these settings, AI can help determine when to replenish, how much to stock and how to respond to demand shifts before losses occur.

The review also found that AI-driven forecasting can help reduce the bullwhip effect, where small changes in consumer demand become amplified as they move upstream through the supply chain. This distortion can lead to excessive inventory, production instability and poor coordination between suppliers, distributors and retailers. Advanced forecasting can improve visibility and reduce these systemic fluctuations.

The authors identify four main areas of evidence across the reviewed literature. The first focuses on core forecasting models and technical accuracy while the second examines inventory and waste management applications. The third addresses strategic impacts, including resilience and supply chain governance and the fourth includes methodological reviews that map the broader field.

Together, these strands show that AI demand planning is moving beyond a narrow technical function. It is becoming a tool for organisational resilience. In the context of climate shocks, pandemics, geopolitical disruption and agricultural crises, firms need systems that can sense risk earlier, adjust operations faster and support more stable decision-making.

The review uses the Technology-Organisation-Environment framework to explain why adoption depends on more than algorithm quality. The technological dimension includes model complexity, data requirements, system compatibility and explainability. The organisational dimension includes firm size, leadership, data governance, staff capability and change management. The environmental dimension includes market volatility, regulation, sustainability pressure and external supply chain shocks.

The framework helps explain why technically superior systems often fail to deliver results in practice. A company may adopt an advanced forecasting model but lack clean data, skilled staff, executive support or planning routines that make the model useful. AI can improve demand planning only when it is embedded into procurement, replenishment, inventory management and strategic decision processes.

SMEs need data governance, skills and change management to adopt AI

Large agri-food corporations are more likely to have the capital, data infrastructure and analytics teams needed to deploy AI planning systems. SMEs often face fragmented records, limited digital infrastructure, high technology costs and shortages of staff who can interpret and manage AI outputs.

Data governance is one of the biggest barriers. Many agri-food firms still depend on incomplete, inconsistent or siloed data. Sales records, supplier data, weather inputs, inventory records and logistics information may sit in separate systems or manual spreadsheets. In such cases, even advanced algorithms can produce unreliable forecasts.

The study also points to the black-box problem. Deep learning models may be accurate but difficult for managers to interpret. When supply chain professionals do not understand why a model recommends a particular action, they may override it or ignore it. The review therefore highlights the need for explainable AI, which can make predictions more transparent and strengthen managerial trust.

Workforce skills are another major barrier. AI adoption requires staff who can work between operations, analytics and management. Companies need personnel who can understand data quality, interpret model outputs, connect forecasts to business decisions and explain results to non-technical teams. Without this bridge, AI systems may remain trapped in pilot projects rather than becoming part of daily operations.

The authors also point to the growing need for change management. AI-driven demand planning alters routines, responsibilities and decision authority. It can shift firms from reactive planning to predictive decision-making, but that transition requires leadership, training, communication and clear integration into business processes. Technology adoption alone is not enough.

For SMEs, the review points to phased adoption rather than sudden transformation. Firms can begin with cleaner centralised data, basic forecasting improvements and simple machine learning models before moving toward real-time data streams, advanced deep learning systems and explainable AI. This maturity roadmap allows smaller firms to build capability gradually.

The review also suggests that SMEs can reduce barriers by working through cooperatives, producer groups, sector platforms or shared data ecosystems. Such arrangements may allow smaller firms to access tools, training and data resources they could not sustain individually. This is especially relevant in sectors such as dairy, wine, fresh produce and other agri-food systems where collective infrastructure can support shared forecasting capacity.

Implications and limitations

Public support for digital infrastructure, training, data-sharing standards and SME-focused AI adoption could help close the gap between large and small firms. Since demand planning affects food waste, supply chain resilience and food security, the benefits of adoption are not limited to individual companies.

Managers should treat AI demand planning as a strategic capability. It can reduce waste, improve inventory governance, strengthen resilience and support sustainability goals, but only if it is aligned with organisational strategy. The review argues that firms must integrate AI into human resource development, innovation management, supply chain governance and sustainable business planning.

The authors also acknowledge limits in the evidence base. The review relied primarily on Scopus-indexed studies, excluded non-English literature and focused on peer-reviewed research, which may underrepresent failed or inconclusive AI implementations. Given the pace of development, existing technical benchmarks may not remain current for long.