Everyone talks about data as the fuel for AI, but even more important is how you describe that data. Metadata, information that tells you what your data means, where it comes from, and how it’s used, is the real key to building powerful, reliable AI applications and agents.

In this article, Pierre Gerardi and Inge Lemmens dive deeper into why you should invest in a metadata catalog and how to tackle the challenges of building one. Inge, Data Governance Lead at Port of Antwerp-Bruges, shares insights on the obstacles they overcame to create a high-quality catalog used to solve strategic business questions. Pierre explains how a well-designed metadata catalog helps AI agents better understand the business.

What is a metadata?

It’s “data about data.” Think of a spreadsheet: the raw numbers are the data, but metadata tells you what each column means, what units are used, which values are valid, and where the info came from. It’s the context layer, and smart thinking thrives on context.

I am not only talking about some labels you add to certain data points. I am talking about a clear description of each data element, what it actually means, and how it relates to other data elements. In fact, I’m referring to the translation of your entire data landscape onto paper, so that readers can understand and use it effectively.

What is a metadata catalog?

A metadata catalog is a centralized repository that stores and organizes metadata across your organization. It serves as a searchable inventory of all your data assets, providing detailed information about what data exists, where it's located, how it's structured, and how it can be used and most importantly what the data means.

Benefits of a metadata catalog

The big benefit of creating a metadata catalog is that it allows an organization to give clear, consistent, and fast answers to strategic, tactical, and operational questions. Starting from the business questions that matter most, the catalog maps the concepts used to describe the organization and connects them to the underlying physical data.

Next to this, it creates a shared understanding of business concepts across the organization, avoiding conflicting definitions. It also provides greater transparency into data itself, where it lives, who owns it, and how reliable it is.

Give your AI Agent a map of the data landscape

This structured understanding that a metadata brings doesn’t just help humans, it also gives AI agents the guidance they need to navigate the data landscape efficiently.

Imagine you’re building an AI agent tasked with exploring your company’s entire data ecosystem. Without metadata, it’s like walking into a library where none of the books have titles or labels. The agent might open every file, sample a few rows, and still have no idea what it’s looking at. A metadata catalog solves this by providing an understanding of the data landscape, giving AI agents the roadmap they need to navigate efficiently and interpret information correctly.

Here are some concrete examples of metadata that can help your agents to provide more accurate and informed responses:

• Dataset description – clarifies what the dataset represents.

• Relationships – describes how datasets are connected.

• Data freshness & trust – shows how recent and reliable the data is.

• Ownership & governance – specifies who owns the data and access rules.

• Expected data types & values – defines valid inputs (e.g., a “priority” field only allows low, medium, high).

• Domain-specific terminology – explains jargon (e.g., ROI means “Return on Investment” in business vs. “Region of Interest” in medical imaging). 

With a robust metadata catalog, AI agents can navigate autonomously through your data ecosystem without manual guidance

Making agents more autonomous by giving them access to a metadata catalog is one of the best investments a company can make to scale its agents. This provides them with a roadmap to solve issues independently. Without it, you have to manually prompt each task for every agent, resulting in a huge time investment and limited agent capabilities.

Example: Metadata catalog in action

At the Port of Antwerp-Bruges, we have leveraged their metadata catalog to build a SQL agent that interacts with nautical data. 

Without metadata, the agent might find a table ship_logs with columns like id, stat, cat, time_at_port and spend significant time guessing what each column represents, whether stat is ship status or priority, what cat codes mean, or whether time_at_port is in hours or days. Often, it would make incorrect assumptions or give up entirely.

With the metadata catalog, the agent instantly sees:

• Table: Ship Logs
  stat = ship status (1 = Docked, 2 = In Transit, 3 = Delayed)

• cat = ship category (CARGO = Cargo, PASS = Passenger, TANK = Tanker)

• time_at_port = duration in hours

Thanks to this structured information, the agent can quickly identify which tables to query, how to combine them, and retrieve the necessary information to answer complex questions. Tasks that would normally take a business analyst an entire day can now be resolved in minutes. Metadata gives the agent the roadmap it needs to navigate the data efficiently and accurately.

Key challenges in building a data catalog

The biggest challenge organizations face is not technical, but rather deciding where and how to begin. Much of a company’s data often lives deep within large systems, making it difficult to know which data exists, which is most critical, and where to start documenting. With big amounts of uncataloged information scattered across teams, identifying high-priority datasets can feel overwhelming. On top of that, data engineers, analysts, and business stakeholders approach data from different perspectives and with different vocabularies. Without structured collaboration and a clear process for uncovering gaps in understanding, organizations risk creating incomplete and inconsistent metadata that fails to support decision-making.

How to build a metadata catalog?

We are not talking about tooling. A tool is a means that is only a part of the solution. Which tool you use is almost irrelevant. The success of any metadata catalog initiative lies in the acceptance of this tool by the end users. Without their acceptance, your tool will end up somewhere on a shelf, gathering dust. 

So where do you start building? By involving the business. While most companies start with documenting the systems where the data resides, Port of Antwerp-Bruges starts with the business perspective of the data, by modelling the communication over the data instead of the physical representation of the data, using a knowledge graph-based approach to identify and relate the concepts. This gives context to the data, relates the data at a conceptual/semantic level, without being biased on how it is stored in operational systems. Because, let’s be frank, how data is stored, in which system it can be found; that information changes frequently, while the meaning of the data remains the same.

Starting with the business has another advantage: it supports identifying the high-priority datasets. “What the heart thinks, the mouth speaks”. Or in this case: “what the essence is of an organisation, the people talk about”. And more often than not, strategic projects are identified that involve these datasets. Getting involved in these strategic projects, showing the advantage of having a clear understanding of what is meant with the data, and how that data can be used for further evolvement of the organisation, is where you can show the real power of having a metadata catalog.

But it’s not only about the definitions and relations between definitions. A good metadata catalog also links those business concepts with the underlying data, forming the bridge between the IT implementations of data and the semantics of that data as used by the business. 

This requires structured collaboration between the business on the one side, the data engineers on the other side, and the data analysts that form the bridge between business and IT; since: 

• Data engineers understand the physical data

• Data analysts know how to translate business questions into cleansed data, using a process that aids in uncovering gaps in understanding such that the risk of creating incomplete and inconsistent metadata that fails to support decision-making is mitigated.

• Business stakeholders provide the data needs, they are the driving force in identifying where to start building the catalog. Their questions highlight the priorities, which can then be used to create a clear roadmap for documentation.

Reaching (data) agreements among all of them should result in a solid metadata catalog. 

Superlinear is here to guide the way. We use a feedback-driven methodology that turns agent mistakes into metadata improvements. By giving agents access to your initial metadata catalog and systematically classifying their errors, we can pinpoint exactly where the metadata “map” needs improvements. This creates a continuous improvement cycle where each agent deployment makes the framework more accurate and complete.

In addition, Superlinear applies GraphRAG for definition mining. GraphRAG identifies key concepts and relationships within organizational documentation. By running it over materials that describe your organization, we can surface critical concepts that explain how your business works. These mined concepts are then suggested to the data catalog team, who can decide whether to formally document them in the catalog. This ensures that the metadata catalog evolves not only from human collaboration and agent feedback, but also from systematically mined organizational knowledge. Read the following article if you want to unlock real intelligence with Knowledge Graphs and GraphRAG.

Conclusion: why you should master metadata

When you master metadata, you unlock the full potential of your data. AI agents become more accurate and far more autonomous, ready to scale across every part of your organization. If you’re serious about scaling AI, it’s time to get serious about metadata.