Is Your Agentic AI Built on Sand or Bedrock?

Moving From Big Data to ‘Deep Data’

Put on your database admin hat for a moment. Like any admin worth their salt, you should be asking, “Where’s the data model for all of this?” Unfortunately, when it comes to GenAI, there’s hardly any to speak of. At least not with any sense of a standard or consistency. This is especially true when, as mentioned, the underlying agentic data is being generated and managed by agents, with little or no transparency.

The upshot is that data and knowledge do not, anymore, exist as separate components. They are rapidly merging into a single architecture. As KM’ers, we can no longer leave data management solely up to the admins. Rather, we need to work closely with them on creating data architectures that are contextually and semantically rich enough to be reliably actionable for use by autonomous and semi-autonomous agents.

There’s an Upside to This and a Key Role for KM

Clearing a path through the data jungle isn’t easy. But it need not be complicated. Here are a few key steps to get started.

Many fall under the rapidly emerging practice of adding a semantic layer to your enterprise architecture. This includes investing in building and maintaining machine-readable layered ontologies. The lower layers deal with domain-specific, esoteric knowledge. The more generalized middle and upper ontologies facilitate knowledge flows across different disciplines and domains. Applying industry and, where possible, open standards, is critical. The world’s largest open source ontology, SUMO (ontologyportal.org), and the recently released NIEMOpen ontology (niemopen.org), which was designed with the semantic layer in mind right from the get-go, are major steps in the right direction.

Other practices incorporate tried-and-true, proven data and text analytics, including sentiment analysis and entity-association extraction for building knowledge-rich graph databases for rendering as knowledge graphs. There's also extractive AI for trend-spotting, causal reasoning, provenance chains, de-biasing, and ethics.

In a nutshell, we’re getting back to good old data and information governance. Does your organization have a chief data officer, and if so, how often do you interact? You’ll need their support in making your organization’s data catalogs more AI-friendly, especially as AI-based auto-classification increasingly becomes mainstream.

All of these efforts help ensure semantic consistency across agents, thereby making agentic data more actionable. This is sorely lacking in LLMs built mainly on conventional wisdom, or, worse yet, generating output based mainly upon positioning of tokens in a text stream. A well-designed semantic layer will go a long way in helping to extract weak signals from the cacophony of noise present in organizational and social discourse.

Learning From Successes as Well as Failures

Be sure to look into, adapt, and incorporate what others are doing successfully. Instacart has built an impressive array of mobile intelligent agents to assist customers and suppliers at major grocers such as Sprouts Farmers Market and Kroger. One product is its agentic analytics tool, which is layered onto a retailer’s existing data stack to provide instant, actionable insights (instacart.com/company/retailer-platform/ai-solutions). Another example is ServiceNow, which reports that by combining its NowAssist agent platform with Microsoft Copilot, AI agents work together as seamlessly as human colleagues, sharing context and coordi- nating complex activities in real-time (devblogs.microsoft.com/semantic-kernel/customer-case-study-pushing-the-boundaries-of-multi-agent-ai-collaboration-with-servicenow-and-microsoft-semantic-kernel).

Finally, don’t overlook our KM community’s oft-forgotten stepchild, tacit knowledge, and its role in incorporating human sensemaking into the mix. LLMs respond based upon their training dataset, which can have blind spots and drift out-of-date over time. A human-in-the-loop should always be present, especially for those occasions when something just doesn’t seem right. This is an essential “gate” in separating signal from noise.

So what will it be? The shifting sands of the endless barrage of noise swirling around the digital universe? Or the stable, rock-solid foundation of a semantically rich data architecture? The choice is obvious. The real question is, as KM’ers, are we ready, willing, and able to assume this critical role?