Three weeks ago, a software engineer rejected code that an AI agent had submitted to his project. The AI published a hit piece attacking him. Two weeks ago, a Meta AI safety director watched her own AI agent delete her emails in bulk — ignoring her repeated commands to stop. Last week, a Chinese AI agent diverted computing power to secretly mine cryptocurrency, with no explanation offered and no disclosure required by law.

David Krueger, “Rogue AI Is Already Here,” Fortune, March 27, 2026, https://fortune.com/2026/03/27/rogue-ai-agents-autonomous-safety/

David Krueger expresses the nature of the problem well: We’re in a phase where rogue AI threatens to undermine the efforts of enterprises to harness AI’s power before they even get started in earnest.

So what’s the solution enterprises need to embrace, then? As I’ve said before, today’s AI is a black box, and the problem has to be on the input side, in a classic garbage-in, garbage-out scenario. Implementing agentic orchestration on top of this sort of ill-managed AI compounds the risk.

There’s far too much underdescribed, unmodeled, decontextualizated data coming in. The semantics community’s answer is the only viable alternative I’ve encountered.

The GraphRAG Alternative to Agentic Orchestration

I had a chance to moderate Graphwise’s March 25, 2026 webinar “Agentic AI Orchestration: So Where’s the Music?”. (Please click on this link to register and view the recording itself.)

In that webinar, we discussed the shift from monolithic, probabilistic AI to trustworthy, hybrid knowledge-centric AI architectures that allow far more accuracy and reliability. What follows are some of the main insights I took away from the webinar.

How to build trust in agents

You may have encountered the hype in the media about agentic orchestration. Despite claims about multi-agent systems, a Cisco Systems customer survey found that 85% of customers experimented with agents, but only 5% moved them to the production phase, due to a lack of trust and reliability.

The core trust solution Graphwise Growth SVP Andreas Blumauer and Presales VP Gerald Mann discussed that solves the problem Cisco identified centered around the six steps to a Semantic Backbone (SBB), a data-centric framework that grounds AI in explicit domain knowledge to prevent hallucinations.

Unlike a thin semantic layer used for basic business intelligence, an SBB integrates multi-modal data (structured and unstructured) to support complex reasoning.

Gerald and Andreas both used an AI flywheel metaphor as a knowledge-first, iterative approach to embedding AI within an organization.

Core Concepts of the AI Flywheel

Knowledge-First Orientation: Gerald noted that while data was the new oil years ago, knowledge is the new oil today. The flywheel aims to maximize the impact of knowledge that is currently locked in disparate data structures or within human heads.

Iterative Maturity: The flywheel represents an organizational journey through six implementation phases.

Grounding for Impact: Gerald pointed out that attempting to jump directly into multi-agentic systems without a foundation is like going on a ride that won’t have a significant impact; the flywheel requires grounding in data to answer questions accurately and achieve progress.

Structure of the Flywheel Circles

Andreas described two intertwined circles that make up this process:

The Inner Circle: This involves technical roles such as data scientists evaluating AI risks and knowledge scientists continuously improving the knowledge graph. The data application owner then justifies these technical improvements by documenting better answers and business results.

The Outer Circle: This focuses on embedding the technology into organizational workflows. It involves strategic oversight (often by a CTO) to manage the semantic backbone as a long-term strategy, track return on investment (ROI), and establish a level of trust that allows the system to be accepted at scale.

Key Questions the Webinar Answered

The Graphwise webinar as Polina Lyabiahova designed it mainly answered questions from those who’d registered for the event. But I know Andreas, Gerald and I answered other questions to frame the overall topic. Key questions and answers overall included these:

What is the agentic paradigm and how does it change enterprise architecture?

The agentic paradigm shifts architecture from being application-centric, where every application has its own siloed database, to being data-centric. This approach puts knowledge at the center to guide agents that understand context (the who, what, when, where, and why) similarly to how humans do.

What really is new this time around is that the architecture really has to shift. We really have to go from an application centric approach with a silo for each app to putting the knowledge in the middle here so that you’ve got explicit knowledge that’s guiding the agents.

–Alan Morrison

What is a “semantic backbone” (SBB)?

A semantic backbone is an incremental, data-centric framework that provides domain knowledge and a “shared world model” for AI systems. It acts as a highly structured long-term memory for multi-agentic systems, helping to ground them, reduce hallucinations, and lower operational costs.

The semantic backbone provides means to have a highly structured long-term memory. for grounding, reduced hallucination, and orchestration reasoning.

–Andreas Blumauer

How does a semantic backbone differ from a “semantic layer” in Business Intelligence (BI)?

A semantic layer is often a thin wrapper optimized for specific BI tasks and table-based data. A semantic backbone is a deeper, incremental approach designed to integrate multi-modal data, including unstructured documents, which a traditional thin layer cannot effectively handle.

A semantic layer in the BI community is useful for the BI tasks for sure… but for the more sophisticated queries a semantic layer as a thin wrapper by the BI community will fail.

–Andreas

How do humans fit into the agentic AI workflow?

Humans, specifically subject matter experts (SMEs), are critical for vetting information and providing the context that AI lacks. They provide a feedback loop to refine the semantic layer without needing technical skills like Python programming.

The human in the loop is absolutely key to be able to go through and vet some of this information. There’s knowledge that exists in our heads that our current AI applications don’t have the ability to go through and replicate.

–Gerald Mann

What is the difference between a knowledge graph and a context graph?

An enterprise knowledge graph (EKG) contains relatively static business knowledge like entities, facts, and regulations. A context graph is an extension that adds dynamic, procedural knowledge—such as recent interactions and environment state—to provide long-term memory for agents.

A context graph is an extension of a knowledge of an enterprise knowledge graph adding to rather static business knowledge the procedural knowledge.

–Andreas

What are the six steps to building a semantic backbone?

The repeatable methodology includes:

1. Taxonomies: Building controlled, interoperable vocabularies using standards like SKOS.

2. Ontologies: Expanding structures with specific properties, classes, and industry standards.

3. Enterprise Knowledge Graph (EKG): Automating the linking of business entities and facts.

4. GraphRAG: Implementing retrieval-augmented generation as a basis for agents.

5. Context Graph: Adding procedural and dynamic environmental knowledge.

6. Agent Chaining: Sequencing grounded agents to solve complex workflows.

It’s a six-step, How do humans fit into the agentic AI workflow? approach… ending with a graphRAG in place… and then we suggest to add the context graph which is an extension also referring back to the enterprise knowledge graph.

–Andreas

Why is GraphRAG considered more accurate than a standalone Large Language Model (LLM) alone?

LLMs perform mathematical inferencing rather than true reasoning. GraphRAG grounds the LLM in domain knowledge, making it traceable and deterministic. One industrial gas compressor company found GraphRAG to be over twice as accurate as a standalone LLM.

GraphRAG was more than two times as accurate as the LLM directly on its own because it was grounded in the domain knowledge itself.

–Gerald

What is the benefit of owning a semantic backbone internally versus using a platform’s graph?

Owning the SBB prevents vendor lock-in and allows a company to customize the graph to its own vocabulary and specific business rules. This internal ownership provides a competitive edge that generic third-party platforms cannot offer.

If you want to have a competitive edge, the Semantic Backbone should be yours, not theirs.

–Andreas

There is no AI shortcut

In this AI era, there is no alternative to transformed, graph-based data management that desilos and disambiguated with the help of standards-based description logic, rules and other semantic metadata.

The headlines regarding rogue AI agents are a sobering reminder that without a deterministic foundation, autonomous systems remain more of a liability than an asset for the enterprise. As the insights from Gerald and Andreas make clear, the trust gap—that chasm between experimental curiosity and production-ready reliability—cannot be bridged by better prompts alone; it requires a fundamental shift in architecture.

By moving away from application-centric silos and toward a unified semantic backbone, organizations can finally provide agents with the “shared world model” they need to function safely and accurately.

If you are ready to move from the 85% of companies stuck in experimentation to the 5% who are delivering real-world ROI, grounding your strategy in a knowledge-first, iterative flywheel is no longer optional. It is the only way to ensure that when you finally orchestrate your AI agents, the result is harmony rather than digital cacophony.

For more information:

Kietzman, Ted. “The Agent Trust Gap: What Our Research Reveals About Agentic AI Security.” Cisco Blogs (blog). March 23, 2026. https://blogs.cisco.com/security/the-agent-trust-gap-what-our-research-reveals-about-agentic-ai-security.

Graphwise. “Modernizing Your Data Strategy with a Graph Center of Excellence.” White paper. Accessed April 7, 2026. https://graphwise.ai/resources/white-paper/graph-center-excellence.

Graphwise. “What Is a Semantic Backbone?” Graphwise Fundamentals. https://graphwise.ai/fundamentals/what-is-a-semantic-backbone/.