Semantic Intelligence: Why AI Agents Need Always-On Data Governance

Watch the full session · 1 hr 4 min

Independent analyst Mike Ferguson argued that fragmented metadata and siloed semantic layers are holding back agentic AI, and that governance must become unified and always on. Collate's leaders then showed agents that catalog, classify and document data on a shared knowledge graph.

Data governance has become too complex for people to run by hand, and AI agents will need shared meaning to work reliably across the enterprise. That was the argument at "From Data Governance to Semantic Intelligence," a Data Science Connect webinar sponsored by Collate. Mike Ferguson, independent analyst, CEO of Intelligent Business Strategies and chairman of Big Data LDN, set out the industry view. Collate CMO Steve Woolledge and co-founder and CTO Harsha Chintalapani then showed how Collate's platform approaches the problem.

Why are agentic AI initiatives stalling on data?

Ferguson said data and AI now underpin nearly every business transformation, and that the focus on AI agents hasn't replaced anything. Enterprises still need data products, BI and machine learning models alongside agents, all aligned to business goals, and all of it depends on data that is high quality, secure and compliant.

The obstacles, in his telling, are complexity and fragmentation. Data sits in SaaS applications, multiple clouds, on-premises systems and the edge. Best-of-breed governance tools each keep their own metadata repository with no standard way to share it. Administrators try to apply policies system by system, which leads to policy drift and firefighting.

AI adds new silos. Ferguson described RAG pipelines and vector databases "turning up like popcorn," more than 50 agent-building tools, and vendors shipping pre-built agents with their own knowledge graphs and semantic layers. "There's no guarantee that everything's consistent across all of this," he said.

What lessons from self-service data apply to AI?

Ferguson warned against repeating the self-service data prep era, when tools were handed to citizen data engineers who then struggled to find the right data. It became a "wild west" of uncoordinated pipelines, he said, and he quoted one of his clients from that time: "Everyone's blindly integrating data with no attempt to share what they create."

What organizations want instead, he said, are trusted, high-quality, secure, compliant and reusable data products, published in a marketplace with data contracts that carry the policies governing their use.

What is agentic data governance?

Ferguson said data governance has grown beyond data quality to include privacy, security, usage, sharing and retention, and that it needs to be unified, active and always on. "We are way beyond the point where humans can handle this," he said. "The level of complexity is too great just for people to do it all."

His requirements list included automatic discovery, sensitive data detection and classification, data quality profiling, a business glossary, a knowledge graph, policies, workflows and metrics. Enforcement agents would apply policies across the estate, while triage agents decide which incidents matter most and recommend actions for people to approve. He sees governance moving from rule-based decisioning toward agent-based decisioning.

"You can't govern the data estate until you understand it," Ferguson said.

What is semantic intelligence?

Woolledge said the biggest consumers of data will be agents rather than people. A person who doesn't understand a dataset knows whom to call. "An AI agent is just going to infer and it may come up with a wrong inference if there's not enough context," he said.

Semantic intelligence, as Collate defines it, combines technical metadata (where data came from, who owns it) with business meaning: shared relationships, common definitions, controlled vocabularies and an ontology that machines can read. Ferguson described the same idea as a new layer in the stack, an enterprise semantic layer and knowledge graph that synchronizes the siloed ones.

How does Collate apply AI agents to metadata?

Chintalapani said the problem traces back to his work at Uber, where each team could give a different definition of a trip, a driver or a passenger. That led to the open-source OpenMetadata project, which Collate builds on. Woolledge said the community has more than 12,000 members and over 400 contributors.

In the demo, Chintalapani connected a data source from more than 120 native connectors. Around eight agents then worked in tandem, he said, collecting metadata, usage and lineage, profiling data, classifying it, assigning tiers, writing missing documentation and attaching data quality tests. Tags and policies set in Collate sync back to source systems such as Snowflake.

He also showed a knowledge graph stored as RDF triples and queryable through SPARQL, glossary terms with custom relationships, and a conversational interface that found a certified customer table and wrote Snowflake-specific SQL against it. New glossary terms suggested by the AI go through an approval workflow with a human reviewer. Collate defines about 700 metadata specifications, available in JSON Schema, RDF and JSON-LD, he said.

Asked how to govern a data estate that keeps growing and shrinking, Ferguson's answer was automation: automatic discovery, mapping, classification and profiling, for unstructured data as well as structured.

Where to hear more on data governance for AI

Data Science Connect's 24 free live webinars in 2027 span four series: the AI × Enterprise Panel Series, AI Demo Days, Agentic AI Showcases and the Plug & Play Series. The AI Enterprise Conference also returns on Tuesday, August 31, 2027, at Pier Sixty in New York City.

Watch: From Data Governance to Semantic Intelligence on YouTube
Event: Webinars · More coverage: Webinars insights