Blue Yonder had its ICON event for customers and partners in the middle of May. Blue Yonder, is one of the largest providers of supply chain software. The company has spent the better part of four years and $2.5 billion to rebuild its technology stack from the ground up.
Unfortunately, there are several supply chain events in May, and I missed ICON this year. Nevertheless, Blue Yonder recently provided an update on key announcements for analysts that missed the show.
Demystifying Agents
Blue Yonder is still committed to an agentic framework. There has been a trend among enterprise companies to break their code into smaller components. These components, which are now called “agents,” have entry and exit points. Data goes in one end; and an output flows out the other. That output is routed to other agents in situations where getting an answer involves several steps. Or it is routed to a user when the answer is complete.
The term “orchestration” is now being widely used. An agent framework is flexible. For example, agents inside what used to be a warehouse management system can be combined with agents in transportation management to provide optimized execution across logistics.
Thus, an orchestrated workflow allows agents to be combined like Lego blocks to build more holistic, industry specific, or even customer specific solutions. Orchestration requires a unified data model, ideally supported with a knowledge graph with a semantic understanding of how data is related. Blue Yonder has partnered with Snowflake for this.
Some pundits refer to “agentic AI.” If we are referring to AI based on Large Language Models, that is a small part of what Blue Yonder is providing. And that is a good thing! LLM AI (or generative AI, if you prefer) is not the best solution for all problems.
LLM solutions from Anthropic and OpenAI are incredibly powerful. But as Duncan Angove, the CEO at Blue Yonder, pointed out at the conference, “supply chain is not a generic reasoning problem. It’s a deeply operational environment with hard constraints, real-time execution, physical consequences, extreme scale, and thousands of interconnected decisions happening across warehousing, transport networks, suppliers, factories, stores!”
A warehouse management system generates massive numbers of transactions. Traditional transaction processing works well here. Transportation management and supply planning get their best solves from optimization engines – the optimization engine becomes an agent in the agent framework. Demand planning works best when built on classic machine learning. LLM agents can provide value, but they are not the most important tool in the supply chain toolbox.
Blue Yonder’s Approach to Agents
An agentic framework can get complex. Gundip Singh, the chief product officer at Blue Yonder, explained that agents can end up working with many subagents. “If you take an inventory ops agent, for example, it’s actually leveraging dozens of sub agents. Each agent itself could have multiple skills. When I refer to a skill, you should think of it as a family of sub-agents that are working together to deliver a certain outcome.”
Agents are empowered to accomplish a task. In order to do that, the agent may have access to interfaces that access data or access their sub agents. Agents usually contain their own logic to act upon what it has learned.
Agents can help automate and accelerate planning and execution. If there is an exception, a business user might spend hours following the breadcrumb to a root cause, chaining out to understand the downstream impacts, and then looking at possible resolution paths to mitigate the problem. Agents can solve exceptions much faster.
Wayne Usie, the chief strategy officer at the company, explained that the agentic framework also simplifies the user interface. Users are not flipping through multiple screens and applications as they try and solve problems. “The application is the interface, and the interface is the application,” he said.
Blue Yonders’ Agent Clusters
Blue Yonder announced five ops agents last year – inventory ops, logistics ops, warehouse ops, space ops, and networks ops. “In each and every case,” Mr. Singh said, “we’ve gone built upon the skills. “Each of these agents now has new skills.”
For instance, last year’s warehouse ops broke slotting out from the WMS. Slotting has been further enhanced. Further, this ops work orchestration skills have been expanded. “So now when there is a disruption in the in the warehouse, either due to resources suddenly being unavailable or hot orders dropping in, then the instead of someone manually having to come in and decide how the hot order should be prioritized, that is something that an agent will do as a automatically for the warehouse operators,” Mr. Singh explained.
Similarly, they have significantly enhanced the network ops agent. This agent pulls in risk alert data from providers of real-time risk solutions. Now, this agent can use that information to derive an imputed category bill of material and a category bill of distribution. These BOMs can identify what product or product families are being impacted. It can then show which nodes in the extended supply chain will be constrained. This has always been the vision behind real-time risk alerts, but it has been difficult to execute efficiently in practice.
The inventory ops agent now has skills that go beyond setting inventory targets based upon a demand plan to understand whether the demand can be fulfilled based on production and supply constraints. This, Mr. Singh asserted, takes Blue Yonder into “insights-driven planning” built upon “a full sense of what the risks and opportunities might be.”
Some customers might want to plug in agents, including LLM agents, they have built into the Blue Yonder framework. The company has built a solution that helps companies evaluate how well those agents are working or if the agents are hallucinating or in other ways generating results they shouldn’t.
Building generative AI agents can get expensive depending upon the amount of data required, the quality of the data, and the way in which the model is trained. Blue Yonder partnered with NVIDIA to develop an AI “Model Training Factory” to make this less onerous and expensive.
Customer Validation
At last year’s ICON, the company announced the new agents and explained the framework. At this year’s ICON, it wanted their customers to speak about their journey, their adoption, and their success with these solutions.
Sainsbury’s, the United Kingdom’s second largest grocer, was one of several companies that provided validation. Simon Roberts, CEO of Sainsbury’s, said that Sainsbury’s ambitious goal is to make good food joyful, accessible, and affordable. To accomplish this, the company invested in technology built to reduce waste, improve accuracy and availability, and simplify operations for team members.
For the most part, the CEO praised the solution without quantifying the outcomes. He said the company had:
- Improved availability while reducing stock through more accurate forecasting and replenishment decisions.
- A more resilient supply chain with embedded simulation to de-risk key trading events like Christmas.
- A single forecasting and replenishment platform across SKUs on one end-to-end platform.
- Digitization of category management to improve decision quality, productivity, and alignment between commercial and supply chain execution.
The one place Mr. Roberts did get specific was around reducing waste at scale through accurate ordering and allocation. This drove 150,000 fewer disposals and 800,000 fewer markdowns during their peak season in 2025.
