A leading provider of supply chain planning solutions, o9 Solutions , held its user conference in Chicago last week. They discussed how their platform is powered by “Neuro-Symbolic AI.” It struck me as a marketing term, but an extremely useful one, nonetheless.

o9 has grown quickly. This year alone, they added 40 new customers. They generate over $500 million in annual revenue, up from ongoing subscription revenue in the $200 million range in 2023 . They are recognized as both one of the largest supply chain planning vendors and a leader for their comprehensive planning capabilities.

There is so much ignorance among executives about AI, and in particular about the capabilities of AI based on large language models, that this term – “Neuro-Symbolic” - can help educate the market about what LLMs can do, and what they can’t.

Many executives, and even supply chain practitioners, think LLMs can do everything better. They can’t! LLM AI will not supplant traditional planning tools; it will complement them. And the LLMs are far from being the most important tool in the SCP stack.

What Technologies are Core to Supply Chain Planning?

The supply chain planning market got started when supply chain models were put into in-memory databases in the early 1990s. In-memory relies on system memory rather than disk space. This is much faster. Without this technology, an advanced plan that could account for complex constraints across a supply chain could not be generated quickly enough to be useful. This is still the core technology at the heart of supply planning.

For demand planning, forecasting is based on statistical, genetic, machine learning, and other algorithms.

But in recent years, the key advance has been adding knowledge graphs to SCP platforms. o9 was the first to do this. Other leading SCP vendors now understand the need for a knowledge graph – the heart of the ontology layer - and are working to catch up.

When you think knowledge graph, think “flexibility.” Knowledge graphs support flexibility by using a graph‑structured model that can integrate diverse data sources and be updated without the rigid restructuring required by relational databases that rely on fixed tables and schemas.

The core process in supply chain planning is integrated business planning. This process balances demand (what could be sold) with supply (what the organization can actually make) to create a plan that maximizes profitability and other strategic goals. The process has evolved to integrate financial, marketing, merchandising, and other functions more tightly into planning. In fact, calling o9’s solution “supply chain” planning understates the breadth of its planning capabilities. The knowledge graph’s flexibility has been key to integrating these diverse business functions more tightly.

Another term bandied about in the market is agentic AI. This involves multiple specialized agents that are orchestrated to complete core supply chain tasks. This is also a flexible architecture. These agents are small pieces of software code that work together.

But SCP systems are based on really big pieces of code. You can’t do supply optimization or advanced forecasting with agents. For example, with supply planning, you need a massive engine to create a plan when dozens or hundreds of supply chain policies and tens or hundreds of thousands of constraints spanning an extended supply chain are factored into the optimized plan. SCP technology is, and will remain for the foreseeable future, based on large code bases. The knowledge graph is core to flexibility when you need to integrate huge amounts of enterprise data to create a plan.

Agents will be used in SCP platforms. But they will mainly run at the edge.

My view has been that LLMs, in addition to being prone to misunderstanding what they are being asked to do, are not good at deterministic, mathematical reasoning. I began to question this when I saw that OpenAI had solved a millennium problem , one of the toughest problems in mathematics. I asked Ashwin Rao, the chief technology officer at o9 (as well as an adjunct professor of applied mathematics at Stanford), about this.

“I am very familiar with this,” he said. “So that is not a pure language model.” OpenAI built a model they call LEAN to solve this problem. It mixes an LLM, which is good at language-based thinking (neuro), with math-based thinking (symbolic).

OpenAI is known as one of the leading providers of LLMs. But even OpenAI had to blend neuro and symbolic to solve this problem. What LLM’s can do seems like magic. But they are not the best tool for all problems.

When O9 talks about Neuro-Symbolic AI, that is what they mean. Any SCP vendor that has an optimization engine, forecasting algorithms, and uses LLMs in the user interface also has neuro-symbolic capabilities.

But this term – “Neuro-Symbolic AI” – is very useful for educating the broader market. I hope the industry adopts it. CFOs and CEOs won’t agree to buy traditional SCP solutions if they think pure LLM-based solutions can solve tough planning problems better.

Acuity Brands Turns Deep Enterprise Data into Action

In the next two sections, I’ll look at the powerful, clever solution that Acuity Brands implemented, then look under the hood to examine which technologies mattered most to delivering it. This will help clarify the role of Neuro (LLMs), Symbolic (math-based solutions), and agents in the solution.

Acuity Brands (NYSE: AYI) provides innovative lighting, lighting controls, and building management systems. They operate a highly complex supply chain involving over 1 million finished goods, 2 million components, bills of materials that can go 13 layers deep, and 18,000 vendors. Roughly 350 stocking locations hold inventory. Some of the components, like printed circuit board assemblies, are hard to source in sufficient quantities.

Their business model focuses on growing net sales, converting profits to cash, and managing the balance sheet efficiently. To help achieve this, Acuity implemented o9 for demand/supply planning and inventory planning. They are at a point in their journey where they understand their revenues and the costs associated with achieving them.

In his presentation, Amit Shah, a product manager for supply chain planning at Acuity, explained that there is a “huge amount of data sitting in o9 that we really need to tap into.” Plans can’t always be executed for several reasons: the forecast isn’t quite right; at times, they have to create a plan before the vendor commits to what they will deliver; and even with commitments, vendors don’t always deliver what they said they would. As a result, the company can end up with finished goods inventory that doesn’t really meet its needs. The company may have too much, too little, or the inventory may be in the wrong locations. In short, the company must continuously adjust its inventory levels.

Mr. Shah then ran a live demo using real data. The first capability they built, focused on inventory execution decisions, was a solution designed to find and mitigate inventory issues. Mr. Shah said to the machine, “I have an inventory problem, but I’m not exactly sure.” The LLM responds, “You have different versions of your plan. Which one do you want to use? A current work review, or a constrained plan?” Shah selected the constrained plan.

The data miner navigates the knowledge graph to gather the pertinent data sets. As it does so, the planner can see the chain of thought the miner uses to analyze the data and generate reports. The system says, “Yes, there are inventory problems. Let me show you some inventory exceptions.” This week, you have 85,000 different items that do not meet your safety stock standards. 70,000 of these are in excess. This translates into $43 million in excess inventory. For 16 SKUs, mitigation efforts are already underway. The planning engine also identifies future risks.

Next, the system produces a burn-down report. The report answers the question, ‘How fast are we going to be able to burn that inventory down?’ It breaks the excess inventory down by SKU/location.

Conversely, it looks at shortages. Where do we not have enough inventory? And reports are generated for that problem as well.

This is not a new type of analysis. o9 could already do this. “But historically you would have to go into the platform and know where to look to be able to get this information,” Shah explained. That requires detailed knowledge from an expert planner. Now, this analysis doesn’t require experts. The large language model democratizes the analysis.

Shah then demonstrated using the LLM to refine the analysis. “So here I say, look, we have different types of inventory, and there are some types of inventory that … don’t really follow the same safety stock rules. So, I now want you to limit this to whatever is labeled as an MRP item.” The engine then produces the same analysis and reports but focused on the inventory that matters. This kind of filtering would also have historically required significant effort from an expert planner.

Next, the planner needs to decide how to mitigate the situation. The planner can ask the system to identify SKUs that are overstocked in one location but understocked in another. This offers an opportunity to ship inventory from one location to another, which, of course, comes with a cost. But the cost may be less than inventory holding costs.

The number of SKUs that could be rebalanced can be very large. A planner can then ask the system to prioritize the SKUs that offer the biggest savings opportunities. Then the question is: where is rebalancing practical? Where do we have existing transportation lanes set up with qualified carriers that can move the inventory?

When inventory is short, a different analysis is done. For understocked SKUs, “do I have any incoming POs that I can ask my vendors to expedite? You can imagine,” Shah continued, with 18,000 vendors and 2 million different components, it is very difficult for planners to go through this.” But the system can do it quickly.

Finally, planners get an executable plan that helps keep customers happy while saving money. Finally, a planner can ask the LLM to create a PowerPoint report if they need to explain what they have done to management.

What is Going on Under the Hood?

So, let’s make sense of what has occurred here technologically. This problem is unsolvable without an optimization engine , what o9 calls symbolic AI. Without the knowledge graph , the analysis and reports couldn’t be generated quickly enough to execute . The large language model , what o9 calls Neuro AI, greatly improves the user experience and democratizes analysis, but isn’t necessary for this solution. When companies look at the cost of tokens to build LLMs, they might decide they can get by without a better interface or the ability to summarize what happened in a PowerPoint.

But Acuity is about to embark on a second phase of this project. For continuous improvement, the company needs to know why something happened. “There’s a bunch of questions that planners have to go through for just one SKU,” Mr. Shah explained. “Was the forecast correct? Was the mix correct? Did my suppliers adhere to their lead times? Did extra demand come in that wasn’t really accounted for?” And when you have 2 million components, “this is a very difficult question for any set of planners to answer on a daily basis.” To sort through massive amounts of data and to answer the “why” questions, the LLM is essential .

o9 refers to this as an inventory agent solution. One could argue that agents traverse the knowledge graph to find and retrieve the exact data needed for the analysis. But I think this is an expert system, rather than an agentic one. An expert system has a defined hierarchy of rules it uses to solve a very specific problem. I lean toward the second explanation. Agents are not part of this solution . But in the future, agents will be used on the edge to autonomously execute the decisions in ERP and other systems.