Knowledge Graphs: An Underutilized (and Safe) Agentic AI Unlock

I just returned from a whirlwind tour across the States, where, in addition to consuming my body weight in chips and queso, I spoke with nearly 150 supply chain and operations leaders about a little-known enabler for AI agents – knowledge graphs. I asked every group the same two questions: (1) Who has heard of knowledge graphs? and (2) Who has tasked someone on their team with creating one? The results were remarkably consistent. In every instance, 30-40% of the group had heard of the tool, and (drum roll) not one hand went up in any session for question two. Are ops leaders missing a trick?

Knowledge graphs connect networks of entities through meaning and relationships. Critically, they unify structured and unstructured data – linking spreadsheets, systems, documents, and conversations – to reflect how knowledge actually exists in the business, forming a live, contextual map of your enterprise.

Imagine a piece of data as a coin with notches on it representing attributes of that data. For example, “Paris” would have notches for “city,” “France,” “Olympics,” “Louvre,” etc. A knowledge graph links data based on these notches, connecting Paris and Tokyo as recent hosts of the Olympics and Paris with the Mona Lisa through the Louvre. As the figure below shows, this mental model can move from discrete data points through to fully contextualized insights (aka a knowledge graph).

Knowledge Graphs + Agentic AI = Context-Rich Autonomy

Knowledge graphs may not elicit quite the same excitement as agentic AI (interest in the latter rose 9x in the last two years, and the chart below shows an even more meteoric rise since 2020), but the link between them is crucial.

AI agents require robust data (think LLMs) to power their “brains,” decision-making, and outcome-driven autonomous operability. Knowledge graphs link data, providing context for AI agents and allowing them to reason better, apply logic across domains, and identify key information like causality.

We’ve seen operations and supply chain folks studying use cases to learn more about agents in supply chain, especially during this early adoption phase. It’s unsurprising, and it’s the exact approach we’ve taken at Zero100. We’ve been building an Agentic AI Hub that now has 122 agent use case examples – across every supply chain function and every major industry – alongside the 1,000+ general and genAI cases already present.

I’ve included some of my favorite examples below, ranging from the discrete “task” level through to “systems” of agents working together autonomously. All of them are possible because they have been built on top of LLMs trained, in part, on a semantic/knowledge graph approach to context development.

Single Task/Functional Quick Wins: 

• UCB and Contract Renewals: Routine contract renewal automation 
  → 50% reduction in source-to-contract time via a self-serve platform
• Johnson & Johnson and Invoice Processing: Agent-led extraction, validation, discrepancy flagging, and AP handoff
  → $1B saved over three years, with higher accuracy and throughput

Multi-Step Processes Within a Function:

• Walmart and Supplier Negotiation Assistant: Agent automation of long-tail supplier negotiation from initiation to signed contract
  → 75% of suppliers prefer bots over humans, delivering 3% average savings
• Alibaba and Contract Lifecycle Coordination: Conversational agent drafts contracts and tracks changes across stakeholders
  → Time saved in supplier discovery and evaluation

End-to-End Orchestration/System of Agents:

• Walmart and Trend-to-Product: System of agents augment the product development lifecycle in sourcing, from demand sensing to prototype
  → Increased speed to market (+257%)

As you progress from task-level to multi-agent systems, the middle layer of the tech stack – and knowledge graphs – are instrumental in working between foundational technology/data repositories and the choreography UI of prompt interfaces and autonomous agents.

Ingesting data, connecting APIs, and powering agent-to-agent connectivity are all strengthened by teaching your tech and letting your tech teach you about what data connects where and why. Knowledge graphs will map and show these connections, making them more accurate and more granular. In other words, they enable better AI agents.

Give Your Agents a Map

Over the last year, 3% of companies have been investing in new jobs that manage knowledge graphs in the supply chain, while 2% are prioritizing new jobs in managing data schemas. Leading companies like Volkswagen and Amazon are among them. The underinvestment in this essential layer of agentic AI is conspicuous, but we predict these data points to grow to >10% in the next six months.

Building a knowledge graph is a low-risk, low-stakes effort and an opportunity to get ahead of your data needs for the future. Not sure where to start? Task a keen team member or a small tiger team from your planning org to map inventory via SKU attributes, then build from there. It will feel a lot like a metadata tagging exercise from yesteryear.

Context is the way forward. Invest in it.