Artificial intelligence (AI) is redefining the boundaries of what's possible across industries, and the supply chain sector in particular is ripe for transformation. At Isazi, we have placed big bets on this transformation through our Hudson product. It is our belief that the integration of foundation models and deep reinforcement learning will not just enhance supply chain forecasting—it will fundamentally reshape how global commerce operates.

1. Foundation Models: The New Cornerstone of Supply Chain Forecasting

Harnessing the Power of Transformers

AI has made significant advances in recent years, and many of these have centred around the transformer architecture. This is a particular type of deep neural network that is very good at sequence modelling. In particular, these models exhibit favourable scaling properties when training them on massive internet-scale datasets. This has been applied with great effect to language modelling, with Large Language Models (LLMs) being created to power technologies like ChatGPT and Claude. These LLMs, often described as "foundation models", have been trained on internet scale data and can be further fine-tuned or prompted to perform a variety of specialised tasks. Since time series data is also sequence data, it is natural to ask the question of whether foundation models can be built specifically for time series data, in particular for retail and supply chain sales and stock data. 

Some progress has been made recently in this space, and we see this accelerating to the point where massive foundational supply chain models will be built that will be fine tuned to specific applications. These models will blend natural language as covariates, and be able to understand market activity, promotions, stock and sales. We have already started on this journey at Isazi, building Hudson’s supply chain forecasting foundation model trained on billions of tokens of time-series data. Our foundation model leverages all this data to predict individual time series, far out-matching the statistical and machine-learning based forecasting tools we used to use before. We foresee this "foundation model" idea as core to the future of supply chains, and are betting heavily on it, adding more data, more covariates, and more context to our foundation model every day.

Beyond Numbers: Integrating Context and Language

What's truly exciting about this new approach is the seamless integration of numerical data with natural language covariates, which is made possible by combining (at the architecture level) language models trained on the whole internet with supply chain foundation models trained on billions of tokens of time series data. These AI models won't just crunch numbers; they will interpret market activities, promotional strategies, and even news articles to provide a holistic view of supply chain dynamics.

2. Overcoming the "Nostradamus Effect": From Static Predictions to Dynamic Strategies

The Limitations of Traditional Forecasting

Accurate demand forecasting is invaluable - we have seen first hand at our customers how even a 20% increase in forecast accuracy is capable of reducing working capital by up to 50% and boosting service levels by as much as 10%. However, there is a limitation to what forecasting alone can achieve. Traditional approaches often create a situation where the demand forecast, whilst accurate, is independent of decision making processes. If the forecast says a company’s sales are declining, it does not provide a way to change this, but merely an accurate proclamation of future doom. This is a phenomenon we have seen play out in the market, and we have come to call it the “Nostradamus effect”. A forecast that is treated as immutable has limited utility, even if it might be extremely accurate.

Empowering Strategic Decision-Making

A well trained foundation model has the potential to break free from the Nostradamus Effect, and increasingly the supply chain forecasting models of the future will enable decision makers to use their predictive power when running scenarios. 

In the Hudson product, we already feed our foundation model pricing information, promotional information (timing, mechanic, depth), and stock information when making forecasts, and we think this is just the beginning. In the future, it will be routine to include news articles, market research, competitor communications, etc. as natural language covariates. Because all of this information is baked into the model, it is possible to use the model to understand the impact of potential promotions, pricing strategies and major market activities before implementing them.

The real power of this kind of model is that it can be used for scenario planning and optimisation. Businesses can shift from asking, "What will we sell?" to more strategic questions like, "How can we achieve 10% growth?" or "What's the optimal scenario for maximising gross profit?" 

This KPI driven usage of the “forecast” model will enable the business of the future to make far ranging decisions both operationally and strategically by leaning on a powerful foundation model that has learned from "internet scale" data in the supply chain domain. We're entering an era of dynamic systems that don't just predict the future but help shape it.

3. The Next Frontier: Deep Reinforcement Learning in Supply Chain Optimization

Beyond Forecasting: Total Supply Chain Optimization

Whilst we believe that a foundational model for forecasting is at the core of the AI supply chain revolution, it is by no means where it stops. Even today in Hudson we feed these forecasts into optimisation models in order to reduce working capital, streamline logistics costs, improve service levels, and maximise operating margin. This already creates hundreds of millions of dollars of value for our clients. 

In the future, we see another massively powerful AI technique being harnessed to fundamentally disrupt this space: deep reinforcement learning. Deep reinforcement learning is a type of AI that learns optimal actions through trial and error, making it particularly suited for complex, dynamic environments like supply chains. Models like AlphaZero have already proven that they are extremely good at game playing, where the requirement is to optimise the next move in a sequence of moves. A model of this type, paired with a foundational model for forecasting, could in theory optimise decision making across the supply chain, from what to order to how to manufacture to how to pack and transport items, and even how to market, brand, price and innovate on new products.

The the AI-Powered Supply Chain

Picture a supply chain where:

• Inventory levels adjust in real-time based on global market trends and local events.

• Production schedules adjust automatically and optimally to balance cost, speed, and sustainability.

• Logistics routes are recalculated instantaneously to account for unexpected disruptions.

• Marketing, pricing and promotion strategies evolve dynamically to maximise profit while maintaining customer satisfaction.

This isn't science fiction—it's the near future of supply chain management, enabled by the technologies described above, and will unlock order of magnitude improvements in supply chain efficiency.

4. Holistic Supply Chain Optimization: Beyond Single-Node Models

Traditional supply chain management often suffers from a siloed approach, where each entity optimises for its own objectives without full visibility into the broader system. This introduces inefficiencies at each connection in the chain. AI may unlock value for each participant in the supply chain independently by following the approaches above, but there are still immense losses due to out of sync planning across the supply chain.

The supply chains of the future need not be hindered by this problem. Indeed, there is no fundamental technological limitation that should constrain the models and AI decision makers to focus on a single entity alone. Quite the opposite - an AI agent that can take into account data from across the supply chain - integrating data from stores, depots, suppliers and shippers - will be able to identify hidden patterns, predict cascading effects and optimise globally rather than locally. Not only will this minimise wastage and improve outcomes globally, but it will cascade benefits to all of the entities in the supply chain. 

This is another space where we see immense potential to create order-of-magnitude improvements for all stakeholders, and we are even now starting to implement pieces of this vision in Hudson (although there is a long way to go). We have successfully helped our customers implement vendor managed inventory as well as collaborative forecasting and planning with their customers. Our goal is much bolder, however: an AI planning agent, fed by data across multiple entities in the supply chain, that delivers insights and optimal outcomes for the entire network and all of its participants.

5. A Vision for the Next Decade: Transforming Industry and Society

Industry-Wide Impact

As we continue to push the boundaries of AI in supply chain management, we can anticipate a cascading effect of efficiency improvements across industries. Early adopters of these technologies will gain significant competitive advantages, driving wider adoption and accelerating innovation.

Broad Societal Benefits

The ultimate beneficiaries of this revolution will be consumers and society at large. As the AI revolution spreads to more and more supply chain companies, competitive forces will drive down prices on a global scale. The overall efficiency gains will also manifest as a more sustainable global supply chain, reducing waste and environmental impact.

We are excited to be at the forefront of this revolution at such a pivotal time in the history not only of AI but also of supply chain management. The next 10 years are going to transform this space significantly, and provided enough of the industry can be transitioned to this exponential technology, we believe it will have a material impact on the cost of living. This is our BHAG at Hudson, our raison d'etre, and we are fully committed to accelerating the advent of this future.

Conclusion: The Dawn of a New Era in Supply Chain Management

The convergence of AI and supply chain management represents one of the most exciting and potentially impactful developments of our time. We're witnessing the early stages of a revolution that will reshape how businesses operate, how consumers access goods, and how global trade functions.

As these technologies continue to evolve and mature, they promise to create supply chains that are more efficient, responsive, and sustainable than ever before.