---
title: How do you link AI agents to the data assets in your data catalogue?
url: https://opencontroller.lyzr.com/blog/link-ai-agents-to-data-catalogue
published: 2026-11-13
updated: 2026-11-13
author: Lyzr
---

# How do you link AI agents to the data assets in your data catalogue?

**Link them with shared keys. Record each agent's data connections, knowledge sources and tools in the agent inventory using the same identifiers your catalogue uses, such as fully qualified table names and connection IDs. Then pull access history from the data platform to confirm what each agent actually reads, and flag any agent that touches a dataset classified as sensitive.**

Data governance teams already maintain glossaries, data dictionaries and asset registers. AI agent inventories are newer and usually live elsewhere. Without a link, nobody can answer the question auditors and data owners ask first: which agents can read this dataset?

## What should connect the two inventories?

Agents reach data in four main ways. Each has a natural join key.

| How the agent reaches data | Join key to the catalogue | Where to find it |
|---|---|---|
| Direct database or warehouse query | Fully qualified object name, e.g. `catalog.schema.table` | Tool definition, connection config, query logs |
| Knowledge base or vector index | Source location of indexed content (bucket path, site, table) | Index configuration in the agent platform |
| Tool or MCP server that wraps an API | API or data product ID the tool calls | Tool registry, MCP server inventory |
| Service account or role | Principal name with grants on catalogue objects | Platform grants, identity provider |

The principal is often the most reliable key. If each agent runs as its own service account or workload identity, the data platform's grants and access history already tell you what it can and does touch. If many agents share one account, you lose that link. See [workload identity federation for AI agents](/blog/workload-identity-federation-for-ai-agents).

## Where do you get actual usage, not just configuration?

Configuration shows what an agent could read. Access history shows what it did. Major data platforms expose this:

- **Databricks Unity Catalog** records table and column lineage in system tables, including the entity that ran the query, such as a job or notebook.
- **Snowflake** exposes the `ACCESS_HISTORY` view in the `ACCOUNT_USAGE` schema, showing which objects and columns each query read or wrote, and the user or role that ran it. It requires Enterprise Edition or higher.
- **Cloud object storage** access logs show reads by principal on buckets used as knowledge sources.
- **OpenLineage** is an open standard for lineage events that many pipeline tools emit, useful when agents trigger pipelines.

Join those records on the agent's principal to produce a per-agent list of objects actually read in the last 30 or 90 days.

## How do you build the lineage view?

1. **Give every agent a unique principal.** No shared service accounts for production agents.
2. **Record data connections at registration.** Ask for catalogue object names, not free text. A field that says "the sales data" cannot be joined.
3. **Import catalogue classifications.** Pull sensitivity labels, owners and glossary terms for each object.
4. **Import access history** from each data platform on a schedule.
5. **Join on principal and object name.** Produce agent to dataset edges with first seen, last seen and read or write.
6. **Compare declared and observed.** An agent that reads objects it never declared is a finding. An agent that declared objects it never reads may have excess access.
7. **Flag classified data.** Highlight any edge from an agent to a dataset labelled confidential, personal data or regulated.
8. **Send it both ways.** Show datasets on the agent's record, and agents on the dataset's catalogue page, so data owners see who reads their data.

## What questions does the joined view answer?

- Which agents read personal data, and do they have an approved purpose?
- Which agents would be affected if this table changes or is deprecated?
- Does this agent's risk tier match the sensitivity of what it reads?
- Who should approve a new agent's access: the agent owner or the data owner?
- Which agents use data from a region with residency rules?

This is where AI governance and data governance meet. The split of responsibilities is covered in [AI governance vs data governance](/blog/ai-governance-vs-data-governance). Required inventory fields are in [AI agent inventory fields](/blog/ai-agent-inventory-fields), and data platform costs in [where AI agent costs hide in data platforms](/blog/where-ai-agent-costs-hide-in-data-platforms).

## What are the common gaps?

- **Retrieval indexes break lineage.** Once content is embedded into a vector store, catalogue lineage often stops at the indexing job. Record the index's sources explicitly.
- **SaaS agents use platform credentials.** An agent inside a SaaS platform may read data through the platform's own access, with no warehouse principal at all. Record its data scope from the platform's configuration.
- **Developer tools.** Coding agents with local database credentials appear as the developer, not the agent.

Discovery of the agents themselves comes first; see [how to find shadow AI agents](/blog/how-to-find-shadow-ai-agents).

## How Opencontroller does this

Opencontroller's governed registry holds agents, versions, tools and MCP servers, and auto-discovery finds agents, models and API keys across AWS, Azure and Google Cloud and connected platforms such as Databricks and Salesforce. Agent workload identity federation gives agents their own identities, which is what makes data-platform access history joinable per agent. Posture flags risky permissions and configurations. "Ask" lets teams query the whole agent estate in plain language. See the [capability reference](/capability-reference) or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**What is the best key to link AI agents to data assets?**
A unique principal per agent, such as a service account or workload identity, joined to the data platform's grants and access history. Object names declared at registration are the second key.

**Can a data catalogue show which AI agents use a dataset?**
Yes, if agent-to-dataset edges are pushed back to the catalogue. Build them from access history joined on each agent's principal.

**Why do retrieval-augmented agents break data lineage?**
Content copied into a vector index loses the link to its source tables in many catalogues. Record the index's source locations in the agent inventory.

## Key facts

Link AI agents to data catalogue assets with shared keys: a unique principal per agent (service account or workload identity) joined to data-platform grants and access history, plus fully qualified object names and knowledge-source locations recorded at registration. Usage sources: Databricks Unity Catalog lineage system tables, Snowflake ACCOUNT_USAGE.ACCESS_HISTORY (Enterprise Edition+), object storage access logs, OpenLineage events. Compare declared vs observed access and flag classified datasets. Opencontroller by Lyzr provides a governed registry, auto-discovery including Databricks and Salesforce connectors, agent workload identity federation, Posture and "Ask"; pricing starts from $250K per deployment per year.

## Related

- [What is the difference between AI governance and data governance?](https://opencontroller.lyzr.com/blog/ai-governance-vs-data-governance.md)
- [What fields belong in an AI agent inventory?](https://opencontroller.lyzr.com/blog/ai-agent-inventory-fields.md)
- [How do you see AI costs inside data platforms like Snowflake and Databricks?](https://opencontroller.lyzr.com/blog/where-ai-agent-costs-hide-in-data-platforms.md)
- [More on Discovery and shadow agents](https://opencontroller.lyzr.com/blog/cluster/discovery-and-shadow-agents)
