Today’s deep supply chains hide dependencies that don’t show up in any single company’s records. A supplier’s supplier can carry a risk — product contamination, unsustainable sourcing, a factory sitting in a disaster-prone region — that may never really gain a buyer’s attention until an actual product defect or failure becomes evident. 

Buyers do encourage suppliers to provide visibility into their own networks. That approach mostly fails, because supply chains are too complex to map by hand and because suppliers have their own reasons why they’d prefer to keep their own supplier information private.

Researchers at Cambridge’s Institute for Manufacturing propose a different fix: don’t wait for disclosure, infer the hidden relationships directly. Their method combines two different kinds of graphs — a knowledge graph, which stores what’s actually known about the supply chain as explicit facts and relationships, and a graph neural network, which learns patterns from that graph and predicts links nobody has stated outright. 

Put together, the system doesn’t just store what companies choose to reveal. It reasons its way to relationships that were never reported at all.

Consider a practical example. A retail buyer wants to know whether a product contains a banned allergen that could be hidden in the many tiers of agri-food supply chains and almost never gets surfaced through normal reporting. Instead of relying on that upstream supplier to admit the connection, the model looks at everything it does know about the network. Sometimes the buyer can infer the missing link on its own.

So the shift that’s occurring is from a supply chain that’s only as visible as its least cooperative supplier, to one in which the system makes it possible to infer key elements of information.

Mathematical proofs and explainability

Explicitness and explainability go hand in hand, whether for humans or agents. Nicholas Thompson, CEO of The Atlantic, posted a short video on LinkedIn on August 21. He paraphrased an idea from mathematician Terence Tao: a proof isn’t really a proof if not even the person who wrote it can explain how it works.

This observation applies directly to today’s AI systems. Most AI today can’t explain itself. It takes two kinds of AI, blended together, to make explainability possible and make it scalable. 

Two kinds of AI, one mostly overlooked until recently

AI has always come in two varieties:

Symbolic AI dominated the field from the 1950s through the 1980s. It works with explicit, deterministic facts and rules: if this, then that. A symbolic system follows documented logic, so you can always trace why it reached an answer. 

One key problem with the symbolic methods that predominated for decades was architectural, not conceptual — symbolic system semantics and rules got buried inside application code, hidden from anyone but the original programmers, and locked into one narrow use at a time. Symbolic systems, whether we called them “AI” or not, never scaled well because it was hard to reuse the rules. Web semantics and graph databases together resolve this architectural dilemma.

Probabilistic AI — neural networks and the deep learning that followed — has also existed since the late 1950s. N-dimensional neural nets became dominant once computers had enough processing power, storage, and network capacity. 

Neural nets provide the means for today’s frontier models. They learn and evolve by being trained on examples. Cat videos, for instance.

Instead of following rules, these systems learn patterns from massive amounts of data. That’s the technology behind ChatGPT and every other large language model on the market today. It’s remarkably capable, but not trustworthy as long as it remains a black box: even the engineers who build these models often can’t explain exactly why the model produced a given answer.

Since the early 2010s, the major AI vendors have placed most of their bets on the probabilistic side — bigger statistical models, more data, more brute force compute. Those bets produced real progress. They also inherited the black-box problem at a massive scale.

That split between fact & rule-based and pattern-based AI is one reason today’s systems struggle to explain themselves. But it’s not the only reason. A second, less discussed split — this one between the research world and everyday business practice — has shaped technology for 60 years, and it explains why the problem resolution has taken this long to arrive.

Two tracks that never connected — until now

On one track, AI researchers worked on how to represent knowledge itself: how to capture facts, concepts, and the relationships between them in a form a computer could reason with. On a separate track, businesses built the systems that actually ran their operations: databases for structured data, and entirely different systems for documents and unstructured content. These two tracks rarely talked to each other.

The result: for 60 years, businesses ran two disconnected systems — one for structured data, one for documents and content — while the research needed to unify them stayed mostly in academic papers. Since 2012, knowledge graphs have evolved to become the first practical technology to close that gap. They give businesses a single, shared way to represent facts, documents, and the rules connecting them.

That’s the missing piece that makes hybrid AI possible. Once a business has one shared structure for its knowledge, it finally has somewhere to keep the symbolic side of AI — the explicit facts and rules — as a knowledge and instance data foundation, that complements the probabilistic side, instead of keeping them apart.

How both kinds of AI have begun to work together

Neurosymbolic AI, or NeSy for short, pairs the pattern-recognition strength of neural networks with the transparency of symbolic rules. Instead of leaving business descriptions, facts and rules buried in application code or underexplained in tabular databases, companies pull those rules out and store them, along with the facts they act on, in a knowledge graph — essentially a structured map of an organization’s data and the relationships between its pieces. Once those rules live in one reusable place instead of scattered across every application, they stop being redundant, and they become inspectable.

That inspectability and lack of duplication leads to more effective data reuse. Andreas Blumauer, SVP of Growth at Graphwise, points to a shift underway in 2026: enterprises are moving away from just chasing raw model power, and toward context that’s articulated for a specific purpose.

Why this matters for a business

Key cost saving and data compliance enabling benefits emerge from managing data and rules using the shared semantics of standard knowledge graphs:

• Less duplicate code. When rules live in one shared knowledge graph instead of being copy-pasted into every application, companies can eliminate a large share of redundant code — Semantic Arts has cited figures as high as 85 percent.
• Answers you can audit. Because the rules are explicit and stored separately from the model, a person can trace why the system did what it did. That traceability is what regulators and auditors actually want to see.
• Domain models for disambiguation and context sharing. Adding a model of one domain allows it to dovetail with others and expand the overall sharing footprint of the graph.
• Ownership of your own data. Companies that build this layer themselves keep control of their data and logic, rather than depending entirely on a model or application provider.
• Compliance with visibility and control. AI and data-management rules are tightening worldwide. A system that can explain its own reasoning has an easier time meeting those requirements than a pure black box does.

That last point on ownership also explains some of the resistance to this approach. If companies build and control their own knowledge graphs and rules, all-in-one AI vendors lose some of the leverage that comes from being the only entity that understands how a system works. That’s a business reason, not just a technical one, for why blending the two approaches has been slower to take hold than it should have.

The gap this resistance leaves behind isn’t just theoretical. It shows up in incidents businesses actually have to clean up after.

What happens when this layer is missing

The risk isn’t hypothetical. A recent breach involving OpenAI and Hugging Face is a case study in what goes wrong when AI agents operate without a grounded, explicit layer of facts and rules to check their actions against. Read the full case study here.

More visibility is becoming possible in neural nets as well. Pascal Hitzler and colleagues are now using neurosymbolic methods to explain what’s happening inside neural networks at the level of individual neurons — work that begins to make the black box more transparent.

Put the business case and the research together, and the direction is clear.

The bigger hybrid AI picture

Pure neural scaling isn’t going away, but it’s no longer the whole story. The organizations getting the most out of AI right now are the ones combining probabilistic pattern-matching with symbolic structure — treating explainability and auditability as requirements, not afterthoughts. That combination is what makes AI systems trustworthy enough to run a business on.

For more information:

Andreas Blumauer, “The 2026 Enterprise AI Horizon: From Models to Meaning and the Shift from Power to Purpose,” Dataversity, January 2026, https://www.dataversity.net/the-2026-enterprise-ai-horizon-from-models-to-meaning-and-the-shift-from-power-to-purpose/

A. Dalal et al., “Toward a Neurosymbolic Understanding of Hidden Neuron Activations,” Neurosymbolic Artificial Intelligence, 2026, https://doi.org/10.1177/29498732261419302

Aidan Hogan et al., “Knowledge Graphs,” ACM Computing Surveys 54, no. 4 (May 2022): Article 71, https://doi.org/10.1145/3447772

E. E. Kosasih et al., “Towards Knowledge Graph Reasoning for Supply Chain Risk Management Using Graph Neural Networks,” July 18, 2022,  https://doi.org/10.1080/00207543.2022.2100841

Alan Morrison, “How Knowledge Graphs Improve AI,” The GraphRAG Curator, March 18, 2026, https://graphrag.info/2026/03/18/how-knowledge-graphs-improve-ai/

Alan Morrison, “The Hugging Face Breach: A Case Study in Why Agents Need Grounding,” The GraphRAG Curator, August 1, 2026, https://graphrag.info/2026/08/01/the-hugging-face-breach-a-case-study-in-why-agents-need-grounding/

Nicholas Thompson, “Explainability in AI Proofs” (video), LinkedIn, August 21, 2026, https://www.linkedin.com/feed/update/urn:li:activity:7496702735875112960