Beyond the hype cycle: GenAI use cases in supply chain planning


The rapid recent advancement of GenAI is captivating the imagination of supply chain professionals, but a noticeable gap remains: while consumers have quickly adopted these technologies, businesses are lagging behind. Why is this?

By Ralf W. Seifert, Alberto Fabregat and Srinath Goud Vanga

One explanation is that the value proposition for consumers is clear, and the service is mostly offered for free. For businesses, on the other hand, value-creating use cases are less clear and have often been over-hyped to the point of inflated expectations. Another explanation is a lack of understanding about what the technology is and what it can do, as is often the case with emerging technologies.

If we look at our planning and decision processes, 80% of enterprise knowledge at global companies is still “tribal” – accumulated experiences, skills, and know-how – and siloed. That is why, if you ask questions like, “Why did we miss the forecast for product X in Market Y last month? What commercial actions can help increase demand for product X in Q3 to be 10% higher, and can the supply chain support that? And at what incremental cost?”, answers are not instantly forthcoming.

With this article, we want to shed some light on how to accelerate the digitization of expertise and tribal knowledge of key functions and processes in customer-facing, planning, supply chain, commercial, and product innovation domains. We believe that future enterprises will compete against each other based on the quality of the digital knowledge models driving their processes.

So, what are the technology’s fundamental capabilities and limitations? What are the current practical applications of GenAI already in beta testing with marquee clients of leading software planning companies such as o9 Solutions?

The strategic deployment of agents

One of the hallmarks of GenAI is its ability to perform tasks. It achieves this through “agents” – AI-driven models or systems capable of performing tasks, interacting, or generating content based on given inputs. They are ultra-efficient digital personal assistants who can juggle multiple tasks at high speed, from data analysis to report writing.

These AI agents learn your preferences and needs, adapting to serve you better. While they excel at handling the heavy lifting of data and tasks, they still rely on human direction to set goals and make key decisions.

Atomic agents handle specific, standalone tasks, while composite agents manage a series of tasks, integrating their outputs to achieve a more complex, overarching goal. This teamwork between the focused expertise of atomic agents and the holistic oversight of composite agents is what makes sophisticated AI systems so powerful and versatile.

Atomic agents

Atomic agents are the foundational elements of composite agents. They perform basic tasks, such as data retrieval or editing, but with contextual understanding and conversational user interaction.

For example, when asked, “Give me my forecast,” an atomic agent would sequentially identify the specific forecast needed (sales, marketing, etc.), the relevant time, region, and product set, locate the report containing this information, and summarize the findings. By training large language models (LLMs) in business practices, these agents can autonomously navigate these steps.

If additional context is needed, they can ask clarifying questions or infer from previous conversations, translate requirements into actions or code to retrieve information, and interpret report data to provide a synthesized response. Examples include generating report summaries, editing data, and more.

Composite agents

Composite agents can connect various data retrieval and synthesis sequences to, for example, analyze and report on the forecast changes for a product across markets, identify top performance variations, and assess risks and opportunities. A composite agent running a supply chain impact simulation for a 20% demand increase would first clarify market specifics, then create a scenario, run the supply chain analysis, and provide a summary with potential constraints and opportunities.

Preparing your organization

Adopting GenAI into a business requires a strategic approach, with a deep understanding of the technology and strategic investment in data and product strategies. Here are three fundamental steps to prepare your organization:

• Digitize tribal knowledge
Document and digitize your existing business practices and insights. GenAI can assist in this process by indexing unstructured documents, using language models to identify key information, and automating the validation process to detect inconsistencies or gaps. This digital transformation lays the groundwork for creating a knowledge base that AI can effectively utilize.

• Translate digital knowledge into AI-compatible formats
Next, convert this digital information into a format that can be understood and utilized by LLMs, commonly known as vector embeddings. These are stored in specialized vector databases, optimized for textual document retrieval. This step is critical in ensuring that the GenAI model can accurately retrieve and utilize your business-specific knowledge.

• Invest in suitable platforms and co-develop GenAI algorithms
Publicly available language models may not be tailored to your specific business needs, as they are not trained on your proprietary data. Train AI on your unique data sets, refine algorithms to suit your operational needs, and ensure that the AI solutions integrate seamlessly with your existing systems and workflows.

Expected value and risks

GenAI offers significant opportunities for productivity gains (30-50%) and converting tribal knowledge into digital formats (80% to 100%). However, there are several risks and recommendations we should be aware of if we want to succeed in using them.

The probabilistic nature of LLMs necessitates careful planning and guardrails for effective deployment in settings where accuracy is paramount. AI outputs may not be factually reliable, a phenomenon commonly referred to as “hallucinations”. Enterprises intending to deploy generative AI for planning must carefully address this tendency, implementing secondary checks to validate the AI’s results, a crucial approach for mitigation.

Bias and privacy are paramount concerns. Many AI/ML systems, including generative AI, can inherit biases present in the data they’re trained on. Organizations building their own generative AI solutions, or even those fine-tuning existing models, must be vigilant to ensure any biases in the training data don’t become ingrained in the final product.

An approach to this is to manage the hierarchies or levels of knowledge. This means that when faced with a question you will define the sequence from which to generate the answer. Finally, data privacy must be a key pillar of any generative AI implementation. Ensuring you tailor your organization’s LLMs and that these work within your private environment instead of in the open space is key.

Overall, generative AI promises groundbreaking capabilities in planning and decision-making. If we can address potential inaccuracies, mitigate biases, and safeguard sensitive data, we will be able to safely and reliably harness its power.

This article was first published by IMD on March 14, 2024.