Data or metadata: what an AI agent should get from you

Someone on your team builds a dashboard in Claude on a Tuesday afternoon. A few sentences, a few questions to the systems, and there is a working view of revenue per region. Impressive. Until the question arrives that you cannot answer straight away: what did that agent see along the way?

That is the right question. And the answer should not depend on luck.

The reflex: all closed, or all open

The moment an agent can reach your systems, a choice lands on the table that is usually made too coarsely. The connection closes entirely and nobody can do anything. Or the connection opens and the agent reaches the raw data, in the hope that it goes well.

There is a third option, and it is almost always the right one. Give the agent what it genuinely needs and nothing more.

What an agent really needs is meaning

Metadata is the layer of meaning above your data. For almost every agentic task that is enough.

For almost every agentic task, metadata is enough. Metadata is the layer of meaning above your data: which fields exist, how they are defined, who owns them, how fresh the numbers are and where they came from.

An agent building a revenue dashboard does not need to see the individual records. It needs to know that "revenue" exists, how it is defined, which table it lives in, at what level and from when. It reasons on that meaning. The raw rows add nothing except risk.

An agent needs meaning, not bulk.

When data is genuinely needed

Sometimes you need the values themselves. A calculation at record level, an export with permission, a check on one specific transaction. That is legitimate. The point is not that data is never allowed. The point is that data should be a deliberate, recorded choice rather than a by-product of an open connection.

Metadata is the default. Data is the exception you allow on purpose.

One switch, valid everywhere

Recorded once per connection, and it holds for people, systems and agents alike.

This is where the metadata switch comes in. Per connection you decide whether only metadata flows or data as well. What, when, where, by whom. You record it once and it holds everywhere that connection is used, whether the answer goes to a person, a system, or an agent in Claude or ChatGPT.

That makes Tuesday afternoon's question simple to answer. The agent saw exactly what you released. The recorded meaning, not the underlying source.

This is precisely why you connect once, not per agent: governance and lifecycle underneath an agent on an MCP server are not a side issue. The switch is where that governance becomes tangible.

This is data minimisation in practice

The GDPR has asked for data minimisation for years: use no more data than the purpose requires.¹ Metadata as the default makes that concrete in the age of agents. You share meaning, you keep data where it belongs, and you can show exactly who could reach what. The fact that the agent arrives through an open standard changes none of that.²

Datahub is the system of intelligence for your AI, over your data and knowledge. The meaning is recorded, access is governed and it stays yours. An agent is then not a risk to contain, but a guest with exactly the right keys.

How to start

• Make metadata the default for every new connection.
• Treat access to data as an explicit, recorded exception.
• Record it once, so people, systems and agents follow the same governed path.

Governance is not the brake on agents. It is what makes them usable.