Why ChatGPT Won't Fix Your Demand Forecasting Problems
Operations executives understand the transformative potential of AI for their supply chain. A number of teams we’ve talked to have tried integrating popular products such as ChatGPT. Every new and legacy planning system now claims to be “Agentic” or “AI-powered” – however, a closer look reveals that all they have done is layer a chat interface powered by LLMs (Large Language Models) over their existing demand forecasting capabilities.
At Omnifold, we believe this approach fundamentally misunderstands what makes AI valuable for supply chain planning, and it's setting customers up for expensive disappointment.
Here’s the one minute version on why:
Read on for more details:
1. Generic AI interfaces lack specialized training on your supply chain
ChatGPT (and LLMs, which power other chat interfaces and agents) is a massive technological breakthrough because it can:
Use its vast, unstructured training data (i.e. all the words on the internet);
Understand the questions you ask it; and
Give answers that are often superior to human answers.
In the supply chain setting, this means that non-technical users can more easily query datasets and generate reports – it’s a powerful tool for these use cases.
However, chat interfaces cannot, for example, autonomously understand that adding a new major retailer will create complex cannibalization effects with existing channels, or that your West Coast seasonal patterns don't apply to East Coast markets. These nuances require AI systems trained specifically on your supply chain dynamics, not general-purpose language models that are embedded into planning systems as AI chat interfaces, or agents.
2. Supply chain planning is ultimately a quantitative prediction problem, not a conversation
LLMs excel at text generation because they are trained with that specific objective in mind – predict the next word in a sentence. They were not designed for the problem of how to integrate all the numerical and contextual data that drives supply chain decisions, into a system that can autonomously predict and optimize outcomes.
You might ask “What about solutions that use Large Language Models, and then customize them for my data and use cases?” This is certainly possible, but is not recommended for business-critical use cases. None of the most significant AI breakthroughs (such as self-driving cars or Nobel Prize-winning biology) were bolted on to existing LLMs – they were purpose-built from the very beginning, using training methods that are specific to the end goal.
3. Chat interfaces are reactive, but supply chains need proactive, adaptive intelligence.
A conversational AI waits for you to ask the right questions. In contrast, an effective AI for supply chain must automatically detect patterns, anomalies, and optimization opportunities as changes occur in the business (ideally, before problems arise at all).
Moreover, “shallow AI” systems can't autonomously improve their forecasting accuracy based on how well their previous predictions performed in your actual operations. Purpose-trained AI for supply chain, by contrast, continuously learns from its forecast accuracy, inventory turns, and operational performance to become more accurate over time—turning every planning cycle into training data for better future predictions.
The takeaway? Be careful to understand claims made by your planning software vendors. Thoughtfully evaluate the capabilities of supply chain AI by asking:
What is the objective function of the AI? How does the use of AI improve the accuracy of my forecasts or optimizations?
Was this AI trained specifically for optimizing my supply chain outcomes? Or is it just an LLM that is prompted or fine-tuned on my data?
Will the AI be able to proactively anticipate changes in my business or market, and adjust my forecasts and decisions autonomously? Or is it purely reactive (accepting or rejecting planner suggestions)?
Learn more about how Omnifold addresses these demand planning challenges here.