Agent Skills: Microsoft Purview Unified Catalog

Microsoft Purview Unified Catalog

The Unified Catalog provides business-friendly data discovery and governance on top of the Data Map: organising assets into governance domains and data products with glossary terms, ownership, and data quality. This skill is the catalog/business layer; the Data Map (purview-data-map) is the technical scanning layer underneath.

When to use

Use this skill when the user wants analysts and data consumers to find, trust, and request data - not when they want to scan sources (that is the Data Map's job).

Do not use this skill for:

• Scanning lakehouses, databases, SaaS sources (use purview-data-map)
• Sensitivity labels for security/compliance (use purview-data-classification)
• Lifecycle/retention rules (use purview-data-lifecycle)
• AI data discovery (use purview-dspm-ai)

Pick the right object for the job

| If you want to... | Object | Owned by | |---|---|---| | Define a business area with accountability (Finance, HR, Supply Chain) | Governance domain | Domain owner (business) | | Bundle related assets into something an analyst can request and consume | Data product | Data product owner (business) | | Give a business term a shared definition tied to data assets | Glossary term | Domain steward | | Measure trustworthiness of a critical data element | Data quality rule + score | Data steward | | Show how a column flows from source to report | Lineage (auto, from Data Map scans) | System-generated | | Grant a consumer access to a data product | Access policy / request workflow | Data product owner |

Rule of thumb: start with 1-2 governance domains and 5-10 data products. A catalog with 200 thinly-curated products is less useful than one with 10 great ones. Curation depth beats catalog breadth, every time.

Approach

A catalog rollout dies when IT "loads everything" without business ownership. Follow the order; each step gates the next.

1. Pre-requisite: Data Map is scanning and producing technical metadata — The Unified Catalog sits on top of the Data Map. Confirm scans are running, lineage is populating, and classifications are firing before designing domains. Verify: Data Map shows scanned assets with classifications (e.g. "Credit Card Number") and at least one lineage link source → sink.
2. Pick the right governance domains — Start with 2-3 high-pain business areas, not the org chart. Good first picks: Finance (regulatory pressure), Customer (consent and CRM sprawl), HR (privacy). Each domain needs a named business owner who has time, not a delegated IT proxy. Verify: each domain has a named accountable owner from the business side with a recurring 30-min weekly slot for curation.
3. Publish 3-5 data products per domain — A data product groups related assets for a specific consumer use case (e.g. "Customer 360 for marketing analysts"). Each product has: description, glossary terms linked, owner, access guidance, and at least one data quality rule on a key column. Do not publish empty shells. Verify: a representative consumer can find the product, understand what is in it, and request access via the in-portal workflow without external help.
4. Build the glossary in parallel, not first — Building a 500-term glossary up front produces a graveyard of terms with no asset links. Define terms as you build products; each term must link to at least one data product or asset at publication. Verify: every published glossary term is linked to ≥1 data asset.
5. Assign stewardship with named individuals, not teams — "Owned by the data team" = owned by nobody. Each governance domain and data product has a single named steward. Stewardship workload should be < 4 hours/week per steward, or it stops happening.
6. Configure data quality on critical data elements only — Do not scan everything. Pick 5-10 critical data elements per data product (e.g. customer_id, order_amount, date_of_birth) and define rules: completeness, uniqueness, validity, freshness. Publish the score on the product page. Verify: each high-priority data product shows a quality score and the rules behind it.
7. Drive adoption with discoverability — Wire the catalog into Power BI (catalog endorsement surfaces in the BI service), Microsoft Search, and Teams. If consumers cannot find products where they already work, they will not use the catalog. Verify: a Power BI dataset surfaces its catalog endorsement; a Teams search returns a linked data product.
8. Operate as a continuous programme — Monthly: review unowned assets, stale products, broken lineage. Quarterly: domain owner check-in on curation completeness. Annually: sunset retired products.

Guardrails

• Align catalog access with Data Map collection permissions. A consumer with no scan-level access cannot see the underlying asset even if the catalog product is published to them - permissions are AND, not OR. Plan this together.
• Curation is a continuous programme, not a one-time load. Without weekly steward effort, the catalog goes stale within a quarter and trust collapses.
• Do not auto-bulk-publish. Importing 10,000 assets via scan ≠ a catalog. Assets without curation are noise that hides the good products.
• Glossary terms without asset links are graveyards. Enforce "must link to publish".
• Data quality rules cost compute. Each rule runs on a schedule and pulls data. Limit to critical data elements; do not put quality on every column.
• Domain count creep kills the programme. 3 domains with depth beats 30 with names only. Add a domain only when an existing one has > 10 well-curated data products.
• Personal data needs governance, not just publishing. If a product contains personal data, coordinate with privacy (Priva) before exposing it to consumer search.

Common anti-patterns

• "Build the glossary first." 500 terms, 0 asset links, programme dies in 6 months. Build terms with the products that need them.
• "Auto-publish every scanned asset." Catalog floods, consumers cannot find anything, trust drops, never recovers.
• "One steward owns 50 products." Not stewardship - a queue. Cap at ~10 products per steward.
• "Catalog as compliance theatre." Published to satisfy an audit, no consumer ever uses it, no curation budget allocated. Sunset it or invest properly.
• "Skip discoverability integration." Catalog exists but consumers search Teams/SharePoint for data instead. Wire it into existing workflows or accept zero adoption.
• "Data quality on everything." Compute cost explodes; meaningful signal drowns in noise. Critical data elements only.

Example prompts

• Set up the Purview unified catalog with governance domains and data products.
• Which governance domains should I start with?
• Build a business glossary and data quality rules.
• How do analysts discover and curate data assets?
• Establish data stewardship in the data catalog.
• How many data products should one steward own?

Microsoft Learn

• Unified Catalog overview: https://learn.microsoft.com/purview/unified-catalog
• Governance domains: https://learn.microsoft.com/purview/concept-governance-domain
• Data products: https://learn.microsoft.com/purview/concept-data-products
• Glossary terms: https://learn.microsoft.com/purview/concept-business-glossary
• Data quality overview: https://learn.microsoft.com/purview/data-quality-overview
• Roles in Unified Catalog: https://learn.microsoft.com/purview/catalog-permissions
• Power BI integration: https://learn.microsoft.com/purview/how-to-search-catalog