Oriented to data, not only by data
Databases are no longer just backing services. With AI agents as connective tissue, modeling moves up to the layer of meaning.
In the traditional view of software versus analytics, a database is just a backing service for an application. BI used data warehouse modeling, cubes, dimensional modeling and ETL to integrate data from many of those services and build some analytical representation of the business. The two worlds were almost completely separate.
Then came the revolution of everything in a company becoming software, from operations to product, and the wish to integrate and understand all of it. Cloud arrived, then big data and the data science hype. We invested heavily in a new infrastructure that changed the paradigm of software and analytics, both in where each space ends and in the tools we use to handle the flow of data.
Two traditions that barely talked
Traditional knowledge modeling uses thesauri and ontologies: ways of organizing meaning, names and concepts. “Technical” modeling stayed in the practical world of tables, relational and analytical databases, and ETL. Translating the granular layer (raw data, tables) into an abstract representation (concepts, names, relations) was always the great bottleneck.
And then LLMs turn out to be remarkably good at drawing that map in both directions: from formal relations written as a graph to SQL over the base tables, and from tables, schemas and descriptions to formal graphs of relations. Juan Sequeda’s work reaches a similar conclusion: enriching the context with taxonomy and ontology gives far more accurate answers than throwing in loose metadata.
The new challenge
With LLMs and agents, the challenge becomes integrating all these systems not only technically but symbolically. Databases stop being mere backing services stitched together by an isolated analytical world. Processes become oriented to data, not only driven by data: what matters most is that data is understood, and means something, to agents, people and other systems.
AI agents act as connective tissue in this network of systems and processes. And in doing so they push the need to model software and data up to an even more abstract layer of the organization.
The market noticed
In 2025 large enterprise software companies bought metadata governance specialists: Salesforce bought Informatica, ServiceNow bought data.world. In a world that wants AI in every layer, the message is clear: for AI to understand your organization’s context, storing is not enough. You have to organize and catalog the information. It’s the road to neurosymbolic AI.
AI won’t fix what organizations don’t understand. If humans can’t reliably pass information across teams, hierarchies and time, it’s naive to assume AI will make sense of that chaos. Poorly defined semantics, handed to autonomous systems, lead to mistakes, inefficiency and even harm. The future isn’t just AI-powered. It’s knowledge-powered.
What I don’t know yet
Metadata, knowledge and intelligence seem to be converging on some semantic, or symbolic, layer that weaves together the “low-level” information assets. I’m intrigued by how this gets built for real. It seems reasonable and right, yet it’s hard to grasp, and the tools and methods for broad adoption are still missing.
People and processes are a major limit on the state of all things data. But the sheer amount of complexity won’t be managed by people and processes alone either. I think about what operating systems did for personal computing, and wonder what kind of system could do that for organizations.
Difficult to see. Always in motion is the future.