Data classification means giving data labels that stay attached to it, so people know how it may be used and protected. Tags are how those labels attach to tables, columns, dashboards, and glossary terms. In Atlan’s AI Context Platform, a person or an AI agent that queries data through Atlan sees the label at the point of the query.
A label that stays attached is recorded on the asset itself in the catalog, not in a spreadsheet beside it. An “information type” is the category the data belongs to, such as medical, financial, or proprietary. If you lead data governance or own the data platform, check your tags before AI agents read company data. Unclassified or inconsistently tagged data can reach the wrong person through an agent, and your team then repeats governance work by hand. The sections below follow one label, from the decision that creates it to the agent that reads it.
Gartner published its 2026 Magic Quadrant for Data and Analytics Governance Platforms on 12 November 2026. To see what the report covers and how to get a copy, read our Magic Quadrant explainer.
| Quick facts | |
|---|---|
| Classification | The sensitivity or category an asset belongs to, set by an owner, a steward, a rule, or an import |
| Tag | The label that records a classification on an asset, read by masking rules, purposes, and search filters |
| Propagation | Copying a tag to related assets; off by default in Atlan and chosen per tag |
| Scheme | The agreed set of tiers and categories, each with a handling rule |
| Reader | Policies, people, and AI agents, at the moment they access the data |
What are data classification and tags, and why do policies need them?
Classification says what kind of data something is and how it should be handled, and tags record that where tools can read it. A policy such as “mask personal data” has nothing to act on until a label marks which columns are personal data.
Information types and handling tiers
NIST gives the long-standing vocabulary. In its guide to mapping information types to security categories, an information type is “a specific category of information (e.g., privacy, medical, proprietary, financial, investigative, contractor sensitive, security management) defined by an organization or in some instances, by a specific law, Executive Order, directive, policy, or regulation” (NIST SP 800-60 Vol. 1 Rev. 1, 2008). The guide builds on FIPS 199, which sets three impact levels (low, moderate, and high) for confidentiality, integrity, and availability. Rev. 1 is the final version. NIST released a Rev. 2 initial working draft in January 2024 (NIST SP 800-60 Rev. 2 IWD).
Two ideas often get mixed up here. An information type describes what the data is, such as a customer email or a payment record. A handling tier says how it is treated, such as mask or deny. A tag can carry either, and a good scheme gives every information type a handling rule.
What tags do in Atlan
In Atlan, tags are labels you attach to tables, columns, dashboards, queries, and glossary terms. They sort assets by sensitivity, domain, or compliance scope (Atlan docs, What are tags).
Tags do three jobs at once: they make sensitive data easy to find, group assets for browsing, and drive access control through purposes (named sets of tags with policies attached, explained below).
Related reading: AI-ready data and trusted data for AI. To compare detection software, see data classification tools. The wider practice sits under data governance.
How Gartner covers data classification and tag management
Gartner’s Magic Quadrant for Data and Analytics Governance Platforms (12 November 2026) lists data classification and tag management as a capability. You can get the complimentary report from Atlan. Gartner defines the capability as follows: “Organize data by relevant categories to use and protect more efficiently.” Our Magic Quadrant explainer covers the rest.
How does data get classified: by people, rules, imports, and scans?
Tags reach an asset from four sources: people, rules, imports, and scan results.
Tagging by hand and by rule
People tag by hand. A steward opens an asset, picks one or more tags, and saves them; a tag can also carry an allowed value, such as a region or a level (Atlan docs, Attach a tag).
Rules scale the work. Playbooks run rule-based automations that update certificates, descriptions, owners, terms, and tags in bulk (Atlan docs, Set up playbooks). Tags that arrive through Atlan’s scanning-tool connector can act as the filter. The documentation’s example marks columns tagged Credit Card Number (a data class from the scanning tool) as needing review (Atlan docs, What Atlan crawls from the scanning tool).
Imports and scan results
Imports pull in labels that already exist. Atlan syncs tags from supported source systems, which the documentation lists as AWS Glue, Confluent Kafka, Databricks, dbt, Google BigQuery, and Snowflake (Atlan docs, What are tags), so labels applied at the source show up in the catalog. For Snowflake, you can import tags, attach them to assets with allowed values, and turn on reverse sync, which updates existing Snowflake tags and never creates or deletes any (Atlan docs, Manage Snowflake tags). Databricks tags have reverse sync too (Atlan docs, Attach a tag).
Scan results are the fourth source. Data security tools scan stores and assign classifications. Atlan’s connector for the scanning tool maps them to Atlan Tags on the matching tables and columns. Both system-assigned and user-assigned classifications arrive as tags, and they can be propagated like any other.
Atlan catalogs only what the scanning tool has already scanned, so a data store waiting for its first scan has no classifications yet (Atlan docs, What Atlan crawls from the scanning tool).
Bringing these sources into one catalog puts every label in one place to review. Separate labels in each tool can drift apart, and then it is unclear which one is current. For the detection side, see automated PII classification, PII data discovery software, and data classification tools in finance.
How do tags travel through lineage?
A tag on a source column labels only that column unless it follows the data. Lineage is the map of where data comes from and where it goes, so a raw column connects to the dashboards and derived tables built from it.
Data lineage tracking covers the basics, with best practices and solutions for keeping it accurate.
How tag propagation works
Tag propagation is the setting that carries a tag along that map. Atlan’s documentation is direct about the starting point: tag propagation is disabled by default, and you turn it on for each tag when you attach it (Atlan docs, Tag propagation). When it is on, tagging a column as PII tags the downstream columns as PII too, and a purpose linked to that tag then covers those downstream columns too. The options differ by where the tag goes.
| Setting | Where the tag goes | Use when |
|---|---|---|
| Hierarchy and lineage | Child assets and downstream assets | A sensitive column feeds many derived tables |
| Hierarchy only | Assets below the tagged one in its own tree | You tag a schema to cover its tables |
| Lineage only | Downstream assets across lineage | You tag a column and want its copies covered |
| No propagation | Nowhere | The label is local to one asset |
Propagation also follows glossary terms: with propagation enabled, tagging a term tags every asset linked to it (Atlan docs, Tag propagation).
The Metadata Propagator app
Tag propagation is a built-in setting on each tag. The Metadata Propagator app is a separate tool that copies owners, certificates, tags, and custom metadata (the descriptive information kept about each asset) from source assets to the assets downstream in lineage.
You get a preview run before anything changes, a replace mode, and a depth limit (Atlan docs, Propagate metadata through lineage). It attaches tags directly to each target, so removing a tag at the source does not remove it downstream (Atlan docs, Metadata Propagator).
The app’s “check direct dependency” setting is on by default, so owners, certificates, and other metadata reach only assets that have a single upstream source, which avoids conflicting values. Tags are handled differently: they propagate even when the target has several upstream sources.
A derived table that combines a Restricted source with an Internal one can carry both tags, and a purpose that denies on either tag covers it. Lineage shows where data flows, not whether it is still sensitive. Check the downstream tags on aggregated or de-identified outputs. For AI use, see AI-ready data lineage and data lineage for AI.
What changes when AI agents and unstructured content read the data?
When an agent queries through Atlan, the policies tied to a tag are applied at that moment. So classification works as live context for the agent, not as a report a person reviews later.
How a tag reaches an agent
A person might notice an untagged column that looks like personal data, but an agent retrieves whatever the policies allow, so a stale or missing tag leaves a gap. Tag updates made in Atlan can also be pushed back to existing Snowflake and Databricks tags through reverse sync (Atlan docs, Attach a tag). Access that goes straight to a source system does not pass through Atlan’s purposes.
How a purpose turns a tag into a rule
The link between a tag and what an agent may touch is the purpose. A purpose picks one or more tags, and Atlan treats every asset carrying any of those tags as part of it. Policies on the purpose can deny a query or mask values, new assets inherit the policy as soon as someone tags them, and “denies from purposes always win over grants from personas” (Atlan docs, What are purposes).
Example: tagging PII on three columns
In Atlan’s own example, tagging customers.email, customers.phone_number, and transactions.cardholder_name as PII, under a purpose with a deny-query policy, makes every PII-tagged column unqueryable through Atlan wherever it appears in the catalog, including columns tagged later (Atlan docs, What are tags).
AI assets and unstructured content
AI policies in Atlan control who can add or remove tags and terms on AI models and AI applications, so those assets can carry labels too. Those policies cover who can manage AI assets, not data access, which purposes and data policies handle (Atlan docs, AI policy). More on the access side is in AI agent access control, data privacy for AI agents, and zero trust data governance.
Unstructured content raises a separate question. NIST’s draft practice guide on data classification, published in February 2026, shows how organizations can “discover, identify and label unstructured data” with commercial classification tools (NIST SP 1800-39 ipd). The guide is still a draft.
Atlan holds standard operating procedures, policies, and compliance rules as assets in knowledge folders. People can find them, governance applies to them, and Atlan’s context agents (its built-in AI agents) can use them (Atlan docs, Knowledge folders). Classifications from a data security scanning tool arrive on the tables and columns it has scanned. A knowledge file can be linked to the assets it relates to, which keeps the policy document beside the data it covers.
For more, see unstructured data for AI, document parsing for AI agents, and unstructured data lineage. The metadata layer for AI explains why machines need to read labels.
How do you set up a first classification scheme?
A first scheme should be small enough that people actually use it. Four steps get a team started.
- Pick a few tiers. Atlan’s tag documentation lists information security schemes such as Public, Internal, Confidential, and Restricted.
- Add the categories your regulations require. The documentation lists personal identification (PII and sensitive PII) and compliance scopes such as GDPR, CCPA, HIPAA PHI, and SOX as common tag uses. Name each after the obligation it serves, so the tag tells a reader why the data is restricted.
- Decide the handling rule and a named owner for each tier before anyone tags. Each tier should map to a handling rule, such as mask, deny query, or require review, even if some labels also serve audits and retention.
- Tag the highest-risk assets first, turn on propagation for those tags, and review on a schedule. Propagate lineage-wide only for tags whose handling rule should follow every copy; Atlan’s playbook guidance suggests tagging a chosen set of assets directly and enabling propagation for their children and downstream assets (Atlan docs, Set up playbooks).
Mistakes to avoid
Three mistakes recur. Teams create tiers with no handling rule, so the labels describe data but enforce nothing. They leave tags without a named owner, so nobody fixes a wrong one. And they set propagation badly: left off, a tag sits on a source column while its copies go unlabeled; turned on for everything, a Restricted label lands on outputs that analysts need.
The legal drivers have their own pages: data privacy governance framework, GDPR compliance for AI agents, HIPAA compliance for AI agents, and EU AI Act data governance requirements. For audit evidence, see how to document data access controls for audits and role-based access control for data governance.
How does Atlan approach data classification and tag management?
In Atlan’s AI Context Platform, a tag is governed context that people and agents can read when they work through Atlan. A tag is created once, attached by people, rules, imports, and scan results, and then read by purposes that can deny a query or mask a value.
If a purpose already covers a tag, an asset tagged tomorrow falls under it with no policy rewrite (Atlan docs, What are purposes).
You choose propagation tag by tag. The Metadata Propagator app is a separate route for owners, certificates, and custom metadata, and it only reaches assets with one upstream source by default. Knowledge folders keep policy documents in the same catalog as the data and its tags, and your governance group defines the scheme and decides the exceptions.
Where this fits
This sits inside the larger context layer idea, the shared record of what data means and how it may be used, which agents read. The metadata layer for AI, the data catalog for AI, and the agentic data catalog cover how tags reach agents. Embedded data governance covers putting controls where the work happens. Enforcement has its own pages: policy enforcement across operational, analytical, and AI use, rule management in data governance, policy enforcement, and AI agent governance.
Get the free 2026 Gartner Magic Quadrant for D&A Governance Platforms from Atlan
Get the Gartner reportFAQs about data classification and tag management
1. What is the purpose of a classification label?
A classification label tells people and systems how a piece of data may be handled. It turns a decision such as “this column is confidential” into something a masking rule, an access policy, or an AI agent can read and act on.
2. What is the difference between data classification and tagging?
Classification is the decision about how data must be handled. A tag is the label that records that decision on an asset. The decision lives in your scheme, and the tag carries it to the tools that read it.
3. What is data labeling, and how does it differ from tagging?
In governance, data labeling is attaching a sensitivity or category label to data, which is what a tag does. The term is also used in machine learning for annotating training examples, which is a separate job.
4. Is tag propagation on by default in Atlan?
No. Atlan’s documentation states that tag propagation is disabled by default. You turn it on for each tag when you attach it, and choose hierarchy and lineage, hierarchy only, or lineage only.
5. Can AI agents use classification tags to limit what they read?
Yes. In Atlan, a purpose picks one or more tags and attaches policies that can deny a query or mask values. Every asset carrying those tags is covered, and Atlan documents that a deny from a purpose wins over a grant from a persona. Atlan applies the policies when the agent queries through Atlan.
6. How do you classify documents and unstructured files?
Scan the files with a data security or content classification tool, label them, and carry the labels into your catalog; NIST’s draft practice guide, SP 1800-39, describes discovering, identifying, and labeling unstructured data this way. In Atlan, classifications from a data security scanning connector arrive as tags on the tables and columns it has scanned, and knowledge folders hold policies and procedures as governed assets.
Sources
- Gartner, Magic Quadrant for Data and Analytics Governance Platforms, 12 November 2026, Anurag Raj Et Al.
- Guide for Mapping Types of Information and Information Systems to Security Categories, NIST SP 800-60 Vol. 1 Rev. 1
- Data Classification Practices, NIST SP 1800-39 (initial public draft)
- What are tags, Atlan Documentation
- Attach a tag, Atlan Documentation
- Tag propagation, Atlan Documentation
- Propagate metadata through lineage, Atlan Documentation
- What are purposes, Atlan Documentation
- AI policy, Atlan Documentation
- Knowledge folders, Atlan Documentation
- Manage Snowflake tags, Atlan Documentation
- What Atlan crawls from the scanning tool, Atlan Documentation
- Set up playbooks, Atlan Documentation
- Metadata Propagator, Atlan Documentation
- SP 800-60 Rev. 2 initial working draft, NIST
Gartner does not endorse any company, vendor, product or service depicted in its publications, and does not advise technology users to select only those vendors with the highest ratings or other designation. Gartner publications consist of the opinions of Gartner’s business and technology insights organization and should not be construed as statements of fact. Gartner disclaims all warranties, expressed or implied, with respect to this publication, including any warranties of merchantability or fitness for a particular purpose.
Gartner and Magic Quadrant are a trademark of Gartner, Inc., and/or its affiliates.