“Microsoft Purview” names three solution areas, and only one is a catalog. Data governance holds two products: the Purview Data Map, and Purview Unified Catalog, the curation layer over it that is authoritative for the governance state of an asset (Purview overview, ms.date 2026-05-15). Atlan builds a different layer, the enterprise context layer; the useful fact is that Microsoft already split those two jobs inside its own stack.
Agent grounding went to Fabric IQ, part of Microsoft IQ, “a set of capabilities that form the enterprise intelligence layer of the Microsoft stack,” whose two core items are an ontology (preview) and a Power BI semantic model (Fabric IQ overview, ms.date 2026-07-08, updated 2026-08-31). Purview appears nowhere in that document. The live question on an Azure estate is which layer answers when an agent asks what a metric means, and what happens to governed context on the way out, the other half of what a context layer is.
| Dimension | Purview Unified Catalog | Enterprise context layer |
|---|---|---|
| What it is | SaaS curation on the Purview Data Map | A neutral layer holding agreed definitions |
| Authoritative for | Whether an asset is governed | What was agreed, and who is accountable |
| Enforcement reach | Azure sources only | Owners and approvers across the estate |
| Agent interface | REST API in public preview, GA features only. No MCP server documented as of 2026-09-16 | MCP, SQL and APIs over one connection |
| Cost model | Governed assets per day, plus DGPUs per run | Platform subscription |
What is Microsoft Purview Unified Catalog, and what is it authoritative for?
Microsoft’s definition: “Unified Catalog is a searchable catalog of your scanned data where you curate, grant access to, and improve the health of your data. The catalog is a software as a service (SaaS) experience based on a single-tenant model” (data governance overview, ms.date 2025-02-19, updated 2025-11-18). It reached general availability on September 1, 2024, announced by Rohan Kumar, Corporate Vice President, Microsoft Security Platform, Microsoft Purview and Microsoft Priva (Microsoft Security blog, 2024).
Its authority is precisely bounded, because every object inside it is a governance object and none of them is the data. Governance domains hold data products, terms, OKRs and critical data elements (governance domains, 2024-04-08 / 2025-11-18). Data products group assets under owners, a use case, a terms-of-use attestation and one access policy (data products, 2024-04-08 / 2025-11-18); that attestation is a weaker object than a data contract for AI. Glossary terms do more than define. Microsoft calls them “active values that provide context” that “also apply policies,” propagating to whatever the term lands on (glossary terms, ms.date 2026-04-24). All of it is curated above the Data Map, roughly the seam between the types of metadata for AI agents.
Then critical data elements, which reconcile columns into one concept, mapping CustID and CID into a single “Customer ID” container (critical data elements, 2026-04-20 / 2026-09-05). That page has read Preview for over two years. The object carrying the most semantic weight in the product is the object a programmatic interface cannot read yet.
Which capabilities carry a status label
Microsoft’s 2026 data quality work is labelled generally available: standalone, incremental and configurable-threshold scans, custom SQL-expression rules, and advanced resource sets (what’s new, 2026-08-26 / 2026-09-15). Preview covers critical data elements, governed-assets semantic search and term-to-column linking (data assets search, 2026-09-08), natural language product search (data products search, ms.date 2026-05-21), and live view (live view, 2024-05-22 / 2025-11-18). The REST API is Public Preview (API overview, 2026-04-30). Governance domains, data products, glossary terms, OKRs, access policies and health controls carry no stage label at all (unified catalog, 2024-08-28 / 2026-03-16): no Microsoft sentence asserts GA or Preview for them, checked 2026-09-16, so read them as shipped, not GA-claimed.
What Purview can scan, and what it can apply a policy to
The Data Map scans Snowflake, Amazon S3, Amazon Redshift, Google BigQuery, Oracle, Teradata, SAP S/4HANA, Salesforce, Tableau and Looker, among others (data map data sources, ms.date 2026-07-01). That is a multicloud observation surface on Microsoft’s own list, a starting point for any multi-cloud context layer argument.
The policies column down all four tables in that doc says Yes for Azure multiple sources, Blob Storage (preview), ADLS Gen2 (preview), Azure SQL Database, SQL Managed Instance and SQL Server on Azure Arc. Zero non-Microsoft sources. Unified Catalog’s access policies are a request-and-approval workflow on a data product, a different mechanism from enforcement at the engine, the gap that separates observing an estate from governing AI agents across multiple clouds. Its table-level read of Unity Catalog is covered in Purview versus Databricks Unity Catalog.
Where did Microsoft put the layer an agent reads?
Fabric IQ “provides structured grounding for copilots and agents, so answers reflect your enterprise language as defined in your ontology (preview),” and Unified Catalog is not on that path. That puts two artifacts most teams govern least at the center of agent reasoning: an ontology in preview, where ontology design for AI starts deciding answers, and a modelling surface built for reporting, the substance of Power BI semantic model versus a dedicated semantic layer.
Purview’s documented contribution is governance over the agent: “risk discovery in prompts and responses, audit coverage for AI interactions, and retention and eDiscovery applicability to AI-generated content” (Microsoft Purview and Fabric, ms.date 2026-02-09, updated 2026-05-11). Real job. Different job from supplying business meaning, whether built in Azure AI Foundry or via Microsoft Copilot Studio agents. No Microsoft document states either way whether a Fabric data agent can read Unified Catalog data products or terms as grounding context.
So whatever Unified Catalog knows has to leave the portal first. The REST API covers eight object types and “only covers the Unified Catalog features that are available in General Availability (GA)” (API overview, 2026-04-30), which leaves out critical data elements, search by meaning and term-to-column linking. No first-party MCP Server is documented for Unified Catalog as of September 16, 2026; the 2026 Data Governance release notes list no MCP entry at all (what’s new, searched 2026-09-16), and that absence is not a roadmap statement. The choice today is a preview REST API against a protocol agents already speak, the trade in why MCP matters for AI agents and when to use MCP versus an API.
The bulk way out is the finding. Microsoft’s own note on the self-serve analytics export:
“The Catalog implements role-based access control (RBAC), ensuring that not all users can view all domains or data products. However, for self-serve analytics, Purview publishes all data, allowing anyone with access to this data to view the entire catalog.”
Only part of the estate travels, because “Purview only publishes governed assets data” (self-serve analytics, ms.date 2026-06-23), and known issue 10 logs the same RBAC behavior (known issues, ms.date 2026-04-15). Governed context either stays inside the portal, or it reaches the agent surface with the permission model that made it trustworthy removed. That is why context layer role-based access control belongs at the front of a retrieval-path conversation. One more bound: “metadata changes are reflected in Unified Catalog within 24 hours” (critical data elements, 2026-09-05). Whether the agent reading that data product knows it holds a day-old record is context drift nobody is instrumenting.
What does Unified Catalog cost, and what do its instruments measure?
Scanning is effectively free once you are on pay-as-you-go, and two meters carry the rest. Unified Catalog charges $0.0165 per governed asset per day on Standard, about $0.50 a month. Enterprise Data Management runs on Data Governance Processing Units, one DGPU equal to 60 minutes of managed compute, at $15 Basic, $60 Standard and $240 Advanced; the Data Map’s Advanced Resource Set bills $0.21 per vCore hour (Microsoft Purview pricing, Data Governance tab, East US, USD, read 2026-09-16). Managed assets are priced uniformly across regions and prorated by days governed.
The first meter runs on an unusual definition: “A governed asset is a data asset that you attach or associate to governance concepts, such as data products, critical data elements, glossary terms, and data quality. If you don’t associate a technical asset to a governance concept, it isn’t a governed asset” (data governance billing, ms.date 2026-04-24). Microsoft’s worked example: 500 assets in the Data Map with only 200 governed costs $100 a month over 30 days, because “the 300 assets that aren’t linked to governance concepts aren’t considered governed assets and therefore not counted.” The processing meter works the same way on activity rather than inventory.
So the coverage instrument is the billing dimension itself, not a number a marketing team picked. The health scoreboard works the same way from the other side: twelve out-of-the-box controls ship, every one a percentage-of-data-products measure, and custom ones are unavailable (controls, 2025-07-27 / 2025-11-18). Traditional catalogs were measured by coverage. A catalog built for agents gets measured by answers.
| What Purview’s instruments count | What an agent needs answered |
|---|---|
| Governed assets per day, the billing meter | Did the agent find what it needed |
| Percentage of data products with an owner | Who is accountable when the number is wrong |
| Percentage with “a description greater than 100 characters” | Was the description enough to answer the question |
| Percentage aligned to an OKR | When retrieval failed, why it failed |
One hundred characters satisfies that control whether or not an agent could answer anything from the description, and neither instrument has a place to record an answer. That is a statement about what a metric can see, not about Microsoft’s engineering. Reading outcomes instead of percentages is where business context for AI starts to matter, and whether the bundled trade is right is the comparison run in context layer TCO.
Purview supplies context automatically and serves it collaboratively, with a steward curating after the scan and nothing in the docs routing usage back into context creation. A catalog built for agents needs Context Agents creating context from evidence, a Context Lakehouse storing and vectorizing it, consumption over MCP, SQL and APIs, and a learning loop that finds the gaps. Vectorization is the difference between an MCP-connected data catalog and a portal with a search box, between semantic search and keyword search.
When is Purview Unified Catalog the right answer on its own?
There is a clear condition under which Unified Catalog plus Fabric IQ covers what a team needs. If the estate is Azure, Fabric and Power BI, the agent surface is Copilot and Fabric data agents, no definition has to hold outside what Microsoft observes, and the governed slice is the slice that matters, then Microsoft’s two products are a coherent answer. The per-asset price is low, scanning is free, and the health controls are sensible. The data quality engine is real and maturing fast, and the access-request workflow solves a problem most teams have. Microsoft is candid about its own preview AI too: the application card publishes the scope limits, the language constraints and a requirement that “Human oversight is required before operational use” (application card, ms.date 2026-06-01). For the rest of the portfolio, there is Microsoft data governance tools, and for evaluation criteria in that market, Purview alternatives.
What changes the answer is a condition, not a verdict. A second agent vendor in the building. A source Purview can register but cannot enforce on. One definition that has to resolve identically in two places. Any of those turns the question into the single-stack lock-in versus neutral context layer decision: governing an agent and supplying it context are separate jobs, a split also written up for Google Knowledge Catalog.
Real customers, real stories: Modern data catalog in action
53 % less engineering workload and 20 % higher data-user satisfaction
"Kiwi.com has transformed its data governance by consolidating thousands of data assets into 58 discoverable data products using Atlan. 'Atlan reduced our central engineering workload by 53 % and improved data user satisfaction by 20 %,' Kiwi.com shared. Atlan's intuitive interface streamlines access to essential information like ownership, contracts, and data quality issues, driving efficient governance across teams."
Data Team
Kiwi.com
🎧 Listen to podcast: How Kiwi.com Unified Its Stack with Atlan
One trusted home for every KPI and dashboard
"Contentsquare relies on Atlan to power its data governance and support Business Intelligence efforts. Otavio Leite Bastos, Global Data Governance Lead, explained, 'Atlan is the home for every KPI and dashboard, making data simple and trustworthy.' With Atlan's integration with Monte Carlo, Contentsquare has improved data quality communication across stakeholders, ensuring effective governance across their entire data estate."
Otavio Leite Bastos, Global Data Governance Lead
Contentsquare
🎧 Listen to podcast: Contentsquare's Data Renaissance with Atlan
How do Unified Catalog and an enterprise context layer divide authority?
The line worth drawing runs per class of context. The Data Map is authoritative for schema, scan results and classification; invert that and a stale copy of the schema drives a query plan. Unified Catalog is authoritative for the governance state of a governed asset and for enforcement on Azure sources; invert that and approval and quality state diverge from the workflow that produced them. An enterprise context layer is authoritative for the reconciled definition of a metric and the accountable owner, serving all of it to any agent vendor with the access model intact.
Microsoft’s docs already say what goes wrong when nobody sets the rule: a curated record can lag the Data Map by up to 24 hours, the bulk export drops RBAC, and nothing alerts when two layers hold the same term with different meanings. So decide the precedence rule per object, set a reconciliation cadence, and name an owner. Authority, not connector counts, is where context layer evaluation criteria should start, and why a model-agnostic context layer is a design decision rather than a slogan. The operating version lives in the context layer for data governance teams.
Atlan is the Context Layer for AI. It holds the class of context no cloud platform observes: the definition two teams argued their way to, the approval that happened in a review meeting, the named human accountable when the number turns out wrong. Atlan AI Labs found that agents grounded in governance metadata achieve 38% higher SQL accuracy than agents working from raw schema alone.
Microsoft’s authority is real, cheap to run, precisely bounded, and published with dates on it. It just put agent grounding in a different product and doc set with no bridging document, so the handoff between its two layers is where the decision lives. Sequencing is the next problem: how to implement an enterprise context layer for AI. When a Copilot agent and a Claude agent both ask what a metric means, which layer answered, and did anyone choose that.
FAQs about Microsoft Purview Unified Catalog
1. What is Microsoft Purview Unified Catalog?
Microsoft Purview Unified Catalog is the SaaS curation and discovery experience on top of the Purview Data Map, based on a single-tenant model. It is authoritative for the governance state of an asset: its governance domain, data product, glossary terms and quality score.
2. What is the difference between Purview Data Map and Unified Catalog?
The Data Map scans sources and stores technical metadata: schema, classifications and lineage. Unified Catalog curates governance concepts on top of it: domains, data products, terms and quality scores. Changes reflect within 24 hours.
3. What counts as a governed asset in Microsoft Purview?
A governed asset is one attached to a governance concept: a data product, a critical data element, a glossary term, or data quality. Assets collected but never linked to a concept are not governed, and are not billed. Microsoft’s example bills 200 out of 500.
4. Which parts of Unified Catalog are generally available and which are in Preview?
Microsoft labels the 2026 data quality scan improvements, custom SQL-expression rules and advanced resource sets as generally available. Critical data elements, natural language search, semantic search, term-to-column linking and live view carry its Preview label. Governance domains, data products and terms ship with no stage label.
5. Can Microsoft Purview catalog non-Microsoft sources like Snowflake and BigQuery?
Yes. The Data Map scans Snowflake, Amazon S3, Redshift, BigQuery, Oracle, Teradata, SAP, Salesforce, Tableau and Looker, so Purview is not Microsoft-only. Enforcement is separate: in Microsoft’s own source grid, no non-Microsoft source supports a Purview policy.
6. Does Microsoft Purview Unified Catalog have an MCP server or an API for agents?
It has an API. The REST API is in public preview and covers GA features only, so Preview objects like critical data elements fall outside it. No first-party MCP server is documented for Unified Catalog as of September 16, 2026, and that absence is not a roadmap statement.
Sources
- Microsoft Purview overview
- Data governance overview
- Unified Catalog
- Data Map supported data sources
- Unified Catalog governance domains
- Unified Catalog data products
- Unified Catalog data assets search
- Unified Catalog data products search
- Unified Catalog glossary terms
- Critical data elements
- Unified Catalog data quality
- Unified Catalog health controls
- Unified Catalog self-serve analytics
- Unified Catalog application card
- Live view
- Data governance billing
- Data governance known issues
- What’s new in Microsoft Purview
- Unified Catalog API overview
- Fabric IQ overview
- Microsoft Purview and Microsoft Fabric
- Microsoft Purview pricing
- Microsoft Purview Data Governance general availability announcement, 2024
- Data catalog vs context layer, Atlan
