If more than about 80% of your models, features and agents live in Databricks, use Unity Catalog and do not buy a second catalog yet. It is the only product here that blocks a user or an agent from reading a model, a feature table or a vector index, and it carries no separate licence on the Premium and Enterprise tiers. Buy Collibra when you must push one access policy into Databricks and Snowflake at once. Buy Atlan when your problem is discovery and business context across many engines and you want the cheapest credible contract. Alation is the one to skip: its price sits close to Collibra's, and its new AI governance product, as described, documents and routes approvals without pushing access policy into the engines.
| Unity Catalog (Databricks) | Atlan | Collibra | Alation | |
|---|---|---|---|---|
| Licence fee | None; included in Premium/Enterprise, you pay compute | Contact sales; Vendr median $50,420/yr | Contact sales; Vendr median $197,142/yr | Contact sales; Vendr median $141,394/yr |
| AI assets it catalogs | Models, functions, vector (AI Search) indexes, volumes, agents/MCPs per vendor | UC models and model versions, Genie spaces, plus its own AI Governance Studio | AI models via Edge (Databricks, Snowflake Cortex AI), AI use-case registry | Models, agents, tools; connectors to Databricks MLflow, Snowflake Cortex, Bedrock, SageMaker, Foundry, Copilot Studio |
| Blocks access? | Yes, natively (grants, ABAC row filters and column masks) | Syncs tags back to Snowflake and Databricks; you build the masking policy | Yes, by pushing masks and row filters into Snowflake and Unity Catalog (Protect) | Workflow and audit trail; no push-down described |
| Coverage outside Databricks | Iceberg REST reads and writes on UC-managed Iceberg; foreign tables read-only | Broad, connector-based | Broad, connector-based | Broad, connector-based |
| Who should not buy it | Teams with most AI on Snowflake, Azure or AWS-native stacks | Teams that need a policy engine | Teams under ~$200k governance budget or with one engine | Most AI governance buyers in 2026 |
Vendr figures checked on each vendor's Vendr page on October 26, 2026. None of the three standalone vendors publishes a list price; their pricing pages route you to sales.
What Each Catalog Actually Governs for AI
Unity Catalog governs the AI asset itself, while the standalone catalogs mostly govern a record that describes it. That difference decides which purchase you need.
A data catalog is a system that inventories data assets, records who owns them and how they connect, and, in some products, controls who can use them. For AI, the assets are the model, the features and tables it trained on, the vector index it retrieves from, the agent that calls it, and the lineage that links all of them.
In Databricks, registered models are a subtype of the FUNCTION securable, so access is granted with GRANT ON FUNCTION, and lineage to the training table is saved automatically when the run logs its input. MLflow 3 points at the Unity Catalog registry by default. Vector indexes, which the docs now call AI Search indexes, "appear in and are governed by Unity Catalog". Databricks' product page adds agents and MCP servers to that list; treat that as the vendor's own claim until you see it in your workspace.
Atlan reads those objects in. Its Databricks connector crawls AI models and AI model versions, but only those registered in the Unity Catalog model registry, and carries lineage for model versions and Genie agents (the latter in private preview). Vector search indexes and functions are not on its crawl list. Atlan launched its AI Governance Studio at Activate 2025 and pitches it as visibility into every AI app and model.
Collibra registers models through Edge-based integrations that include Databricks and Snowflake Cortex AI, and links each model to a registered AI use case. That use-case record is what a risk or compliance team needs for the EU AI Act. Collibra bought trail ML in October 2026 to deepen it, which we covered in our report on the acquisition.
Alation launched AI Governance on May 11, 2026 as "a single system of record for every AI model, agent, and tool," with model cards, a regulation registry mapped to the EU AI Act, NIST AI RMF and ISO 42001, and an append-only audit trail. In September it added generally available connectors that pull models from AWS, Databricks, Microsoft and Snowflake into one registry.
Which Ones Block Access and Which Only Describe It?
Only Unity Catalog and Collibra enforce. Atlan and Alation document, and Atlan can hand tags to the engine so the engine enforces.
Unity Catalog's attribute-based access control evaluates governed tags in policies attached at the metastore, catalog, schema or table level, and applies row filter and column mask policies automatically to every matching table. Some pieces are still Beta, including metastore-level policies, views and DENY policies. Check those three before you design around them.
Collibra Protect is the enforcing option outside Databricks. On Snowflake it turns a classification into a tag and applies tag-based masking, and creates row access policies when you specify row filters. On Databricks, Protect has been GA since the Q4 2024 release and applies its policies to Unity Catalog to mask columns and rows. So a Collibra policy ends up executing inside Unity Catalog anyway. That is the strongest argument for Collibra: one policy, written once, landing as native masks in two engines.
Atlan's enforcement path is tag reverse sync. It is set per tag, and on Snowflake it only changes tags that already exist; it neither creates nor deletes them. Someone still writes the masking policy in Snowflake or Databricks that reads the tag. That works for a team with a strong platform engineer. In practice Atlan supplies the context and a tag pipe, and the engine does the enforcing.
Alation's AI Governance routes approvals by regulation applicability, opens remediation tasks for missing evidence and logs actions. Nothing in the launch release says it blocks a deployment. William McKnight of McKnight Consulting told TechTarget that enterprises must pair it with runtime guardrails, evaluation suites and orchestration tools. The same is true of Atlan and Collibra for agents. None of the three standalone catalogs documents stopping an agent mid-call today (Collibra says trail ML adds runtime blocking, but gives no availability date); for that you need the controls in our agent audit logging guide and an MCP gateway.
Unity Catalog's Free Licence and Where It Stops
Unity Catalog costs nothing extra inside Databricks and gets thin quickly outside it.
Databricks says Unity Catalog requires no separate licence: you pay for the compute and storage you use. AI Search, for example, needs serverless compute, so the governance spend shows up on your compute bill instead of a licence line.
Outside Databricks, the managed product exposes an Iceberg REST catalog. External clients such as Spark, Flink and Trino can read and write managed Iceberg tables, but foreign Iceberg tables are read-only and need a manual REFRESH FOREIGN TABLE, managed Delta tables are read-only to them, and external clients cannot run snapshot expiry or orphan-file cleanup. Tables with row filters or column masks need a separate cross-engine ABAC setup. Oz Katz of lakeFS, which sells a competing data-versioning product, argues that Databricks' governance and AI features work fully only on tables Unity Catalog manages, and that tables in Glue, Hive or another Iceberg catalog are read but not written. Weigh it as a critic's reading, but it matches the Databricks doc's own table of what is writable.
The open-source Unity Catalog is a different product. The GitHub project is Apache 2.0, covers tables, volumes, functions and AI models, speaks the Iceberg REST and Hive metastore APIs, and is a sandbox project at the LF AI and Data Foundation. Its README describes none of the ABAC, automated lineage or quality monitoring that the Databricks product page sells. If someone on your team proposes OSS Unity Catalog as the free cross-engine catalog, ask who will build the lineage and policy layers, because the project leaves those to you.
We covered the platform side of this decision in Snowflake Cortex vs Databricks Mosaic AI and Databricks alternatives; the catalog follows whichever engine holds most of your AI.
What Atlan, Collibra and Alation Cost
None of the three publishes a price, and the third-party figures do not reconcile. Treat them as a negotiating anchor and get a written quote.
Atlan's pricing page and Alation's both route to a sales conversation with no tiers or units shown, and Collibra lists no public price. The best public numbers are Vendr's buyer data, checked October 26, 2026:
- Atlan: median $50,420 a year, range $30,078 to $177,847, based on 44 purchases, page updated February 2026.
- Alation: median $141,394 a year, range $77,526 to $302,223; Vendr does not state the sample size.
- Collibra: median $197,142 a year, range $174,740 to $225,360, on a page still titled for 2025, with no sample size shown.
Other aggregators do not agree. CostBench, which cites Vendr data, lists per-user Collibra prices that its own 25-seat example multiplies out to about $4.3M in year one, far above any reported deal. The ranges are wide because the price moves with users, connectors and governance modules. Use the medians to sanity-check a first quote, and nothing more.
On the defined workload for this page (Databricks for most ML, Snowflake for BI, a few hundred registered models and agents, a governance team of three to five), the licence gap is roughly $0 for Unity Catalog, about $50k for Atlan and $140k to $200k for Alation or Collibra, before implementation.
Deployment Time and the Stewardship You Will Actually Staff
We found no independent, verifiable go-live data for these three products. The timelines that circulate are published by one vendor about its rivals, so we are not quoting them; ask reference customers instead.
What is documented is why catalogs fail after go-live. A May 2026 DATAVERSITY piece by Samuel Bocetta says most organizations treated their catalog like "a glorified spreadsheet with better branding", loaded it once, and watched it go stale "within weeks." Enterprise Knowledge's adoption paper reaches the same place: adoption needs a user-driven program design and integration into day-to-day tasks, and both of those cost staff hours.
Here the platform catalog has a structural edge. Unity Catalog's metadata is written by the platform as jobs run, so lineage and model registration stay current without a steward. A standalone catalog adds a second place where descriptions, owners and classifications must be kept true, and a team of three to five cannot curate thousands of tables by hand. Before you sign, ask each vendor for two reference customers your size and ask them one question: how many hours a week do your stewards spend in the tool, a year after go-live?
Who Should Not Buy a Standalone Catalog at All
If Databricks already holds your models, features and indexes, a standalone catalog duplicates metadata the platform keeps current for free. Skip it until one of three things is true: a regulator wants an AI use-case inventory that spans vendors, a second engine holds sensitive data your AI touches, or business users cannot find what they need.
Do not rely on Unity Catalog as your cross-engine answer if most of your AI runs on Snowflake, Azure or AWS-native services. Its reach outside Databricks is reads, plus writes to its own managed Iceberg tables.
Do not buy Atlan if you need a product that writes and enforces policy. You will still build masking in each engine.
Do not buy Collibra if you run one engine or your governance budget is under about $200k a year. You would pay for push-down you do not use.
Do not buy Alation for AI governance this year unless you already own it. At a median near Collibra's, it gives you a well-specified registry and workflow, and Atlan gives you the registry for roughly a third of the price.
How to Decide Without Regret
Most catalog regret comes from buying a product that cannot enforce policy in the engine where your sensitive AI data lives. Start from that map.
This Week:
- Count where your registered models, feature tables and vector indexes live. If Databricks holds more than 80%, turn on ABAC on governed tags and stop the catalog evaluation.
- List every engine that holds data your AI reads and mark which ones need masking.
This Month:
- If two or more engines need the same mask, run a Collibra Protect pilot on one classification across Snowflake and Unity Catalog and confirm the policies appear natively in both.
- If the gap is discovery and context, price Atlan against the $50,420 Vendr median and ask for the per-connector terms in writing.
Before Your Next Renewal:
- Ask two reference customers for steward hours per week a year after go-live.
- Check the Beta status of Unity Catalog's metastore-level ABAC and DENY policies again; both change what a free catalog can do.
The answer changes if your AI moves off Databricks, if Unity Catalog's cross-engine ABAC lands for external engines, or if Atlan or Alation starts writing policies into engines instead of tags. Recheck the prices when you renew; the Vendr medians above are dated October 26, 2026.
Continue Reading
- Collibra Acquires Munich Startup trail ML for AI Governance
- Snowflake Cortex vs Databricks Mosaic AI: Pick on Exit Cost
- Databricks Alternatives: How to Cut Costs Without Migrating
- Snowflake Alternatives: Fix the AI Bill Before You Move a Table
- EU AI Act Governance Tools: Buy Inventory, Not Policy Packs
- Croissant and OpenLineage Help Document Training Data Lineage