Metadata management best practices, as most published lists give them, produce a catalog with good coverage. That was the right target when the reader of metadata was a person. The reader is now as likely to be an agent, and coverage tells an agent nothing about whether a definition is right. This post is for the architect or data leader who owns a metadata program and needs it to hold up when an AI system acts on the answer. The ten practices below turn catalogued metadata into authoritative context: a record a machine can read, a named expert has approved, and every tool shares. Each one has a test you can run today.
The usual list runs: assign a team, write a strategy, adopt standards, deploy a tool, automate collection, scale out. All six are sound. None of them says who approved a definition, whether a rule can check it, or whether the record can leave the tool it lives in. The ten below start where that list stops.
| # | Practice | What it produces | The test |
|---|---|---|---|
| 1 | Scope to consequence | A short list of terms where a wrong answer costs something | You can name the twenty terms agents resolve most, and who owns each |
| 2 | One authoritative source per domain | One record wins; every other copy is marked as a copy | For any term, the catalog says which record is authoritative |
| 3 | Approve at the concept, map to the columns | One definition, mapped as data to every column that carries it | "Active customer" resolves to all forty of its columns from one record |
| 4 | AI proposes, your experts confirm | Drafts and mappings arrive scored; a steward accepts, edits, or rejects | No definition reaches an agent without an approver's name on it |
| 5 | A name, a date, and a version on every record | Approver, approved-on, review-due, version, successor | One query returns every record whose review date has passed |
| 6 | Attach a validation rule | A SHACL shape per definition, run on every load | The rule reports pass or fail; a definition with no rule is flagged |
| 7 | Hold the record on open standards | RDF, SKOS, OWL, and SHACL, exportable in full | You can export everything, with history, tonight |
| 8 | Serve the record at query time | One interface for people and agents, review trail attached | An agent can cite the approver and refuse an expired record |
| 9 | Keep the catalog, sync it from the record | Catalog descriptions reflect the governed definition | Changing a definition once changes it everywhere it is shown |
| 10 | Measure the governed slice | Share of consequential terms holding a current approval | The number falls when a review date lapses, and someone notices |
Start with the questions an agent will answer where a wrong answer has a cost. A credit limit, a regulatory figure, a clinical lookup. List the terms those questions depend on. That list is the authoritative slice, and it is small. In our post Context Management: The Complete Guide we state that the authoritative slice is consistently under 5% of all data in TopQuadrant customer engagements. The rest can stay inferred, because a person reads it and corrects it as they go. Our post Ground Truth: SME-Approved vs Inferred Metadata makes the full case for that split. 85% accurate isn't good enough for the slice. It is fine for everything else.
Then decide, for each domain in the slice, which source is authoritative. One record for "active customer", owned by Finance. One for "counterparty", owned by Risk. Everything else becomes a copy of it. This is a governance decision, and it is the one that stalls longest, because it means telling four teams that their definition lost. Our post What Is Data Governance for AI? covers how that call gets made and by whom. Our post Metadata Management vs Data Governance: What Each One Decides draws the line between the decision and the record that enforces it.
A model can ingest a definition typed into a description field. It cannot follow the sentence to a column, test a value against it, or tell whether it expired. Practices three to seven fix that. They are easiest to see applied to one term.
Our post Metadata Management for AI: Why Machine-Readability Beats Catalog Structure sets out the four tests a record passes once this is done: resolvable, checkable, dated, and attributable.
Regulated industries asked these questions before AI did. Each text below was written for human-run reporting, and each already asks for something a governed metadata record holds. The table covers the record. Our post Enterprise Metadata Management: Ownership, Lineage, and Business Value at Scale covers the organizational side for financial services and life sciences.
| Who it applies to | Text | What it asks for | What the record must hold | Practices |
|---|---|---|---|---|
| Banks under Basel supervision | BCBS 239, Principle 3, Basel Committee, January 2013 | Integrated data taxonomies that include metadata, and one authoritative source for risk data per type of risk, with other sources reconciled to it | One winning definition per risk term, every copy marked as a copy, reconciliation recorded as data | 2, 3, 9 |
| Providers of high-risk AI systems in the EU | EU AI Act, Article 10 | Data governance practices covering "data collection processes and the origin of data" and "the formulation of assumptions, in particular with respect to the information that the data are supposed to measure" | Provenance for each field, and an approved definition of what the field measures | 4, 5 |
| Any controller of EU personal data | GDPR, Article 30 | A record of processing activities listing "the purposes of the processing", the "categories of personal data", "the categories of recipients", and "the envisaged time limits for erasure" | Classification, purpose, recipients, and retention as queryable fields on each dataset, with an owner | 5, 7, 8 |
| Life sciences firms keeping FDA-regulated electronic records | 21 CFR 11.10 | "Secure, computer-generated, time-stamped audit trails" of actions that "create, modify, or delete electronic records", and "accurate and complete copies of records in both human readable and electronic form" | Who changed a definition and when, held as data, and a full export on demand | 5, 7 |
Three of the ten do most of the work: a name and a date on every record, a full export, and one authoritative source. None of these is new to a compliance team. What changes with AI is who reads the record. An agent reconciling a risk figure needs the same trail an examiner does, at query time.
Your catalog stays. It does discovery well: a person searches for an asset and finds it. The governed record does meaning: one definition the forty columns share, with its steward and its history. The relationship between the two is sync. The approved definition flows down into the catalog descriptions it governs, so the catalog shows approved meaning and stops inventing it. Our post Metadata Management vs Data Catalog: Keep the Catalog, Govern the Meaning makes the case in full, and our Knowledge Graph vs Data Catalog: What Each One Solves comparison article covers the storage side.
| Flow | From | To | What moves |
|---|---|---|---|
| Definitions | Governed record | Catalog | Approved text, steward, review date, status |
| Mappings | Governed record | Catalog, pipelines, BI | Term-to-column assignments, scored and approved |
| Inventory | Catalog and scanners | Governed record | New tables, columns, and lineage edges, as proposals for a steward |
| Serving | Governed record | Agents and applications | Definition, scope, rule, status, and provenance, at query time |
The serving row is the one most programs skip. Our post Active vs Passive Metadata: Why a Well-Organized Catalog Still Goes Unused covers why an approved record that never reaches the tools is a policy document. Can an agent call one interface, get the definition with its approver and date, and decline it when the date has passed? If yes, the metadata is managed.
Five show up again and again.
Our post Metadata Management Framework: Four Stages and a Self-Assessment turns the ten practices into stages, so you can see which ones your program has already done. Tool selection is in our post Metadata Management Tools for Governance and AI Readiness.
TopQuadrant does not sell the catalog your analysts search. Your catalog stays, and it syncs. Our metadata management use case page shows how TQ Data Foundation does this in practice.
TQ Data Foundation is the governed record the ten practices produce. Every term gets an agreed definition, a named steward, and a date the record was last confirmed true. AI proposes, your experts confirm: suggested mappings arrive scored, and a steward approves before anything becomes authoritative. Each definition is validated with SHACL. Everything is held in RDF, SKOS, OWL, and SHACL, standards TopQuadrant helped author, and exportable in full at any time. We maintain 300 reusable standards-based models, so the first approval starts from governed structure.
Ten practices, one test each. Run the tests against your program this week. The ones you fail are the ones TopQuadrant was built for.