Policy enforcement in data governance is the act of applying a written data rule at the point where data is read, changed or used. A single data policy often has to hold in three places: the operational database, the analytical warehouse, and the tool call of an AI agent. Each place needs different attributes, so a rule that works in one can fail silently in another. Atlan’s AI Context Platform holds the business policy and the classification tags. Atlan also applies access policies of its own (persona, purpose and deny policies) to queries and previews in Atlan and to requests an agent makes through the Atlan MCP server (MCP is the Model Context Protocol, a standard way for agents to call tools), while warehouse masking policies can key on tags that Atlan syncs.
Take one rule, mask national IDs for everyone outside HR, and follow it through each setting: where it fires, what it has to carry and how it breaks. The general definition and steps live in our guide to policy enforcement.
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 | |
|---|---|
| What it is | A written policy applied at the point where data is read, changed or used |
| Three settings | Operational applications, analytical use, AI agents and the data copies behind them |
| What each setting needs | A named rule, an attribute to key on, and a place that can act |
| Where Atlan fits | Business policy and tags; persona, purpose and deny policies for Atlan queries and MCP requests |
| Related | Data governance vs AI governance |
What does policy enforcement mean when one policy runs in three places?
Policy enforcement is the act of applying a written rule at the point where data is read, changed or used.
Decision point and enforcement point
According to NIST SP 800-162, the work splits in two. A policy decision point computes the access decision. A policy enforcement point carries it out.
Either can be central or spread across systems. Attributes of the subject, the object and, in some cases, the environment drive the decision.
Example: the national ID rule
The running example is one sentence of business policy: mask national IDs for everyone outside HR. In an application it becomes a column-level rule, since row security alone cannot mask a field. In a warehouse it becomes a masking policy. For an agent it becomes a limit on which tools it can call, plus the warehouse mask on any query it runs through them.
A policy can be well written and still fail at the handoff. The rule itself is sound. The system that has to apply it never got the data it needs to decide. The warehouse cannot mask by department unless the policy is bound to the national ID column and the warehouse can tell who is in HR. An agent cannot honor the rule if it queries as an account no one can tell apart from HR.
How the three settings differ
Each row of this table is tied to its source in the sections below:
| Operational applications | Analytical use | AI agents and their data copies | |
|---|---|---|---|
| Where enforcement happens | Inside the source system: database row security, application authorization or the API layer | In the warehouse or query engine at query time: masking policies, row access policies, row filters, column masks | Outside the model: tool-call authorization in the downstream system, with data-layer rules still applying to the queries the agent sends under its identity |
| What the policy must carry | Who the subject is, which rows or fields, and whether it covers reads, writes or both | A classification tag the rule keys off, a role or group mapping, and a mask or filter expression | The user’s identity, the tools the task may use, the sensitivity of the data, and an approval step for high-impact actions |
| Documented failure points | Privileged roles and table owners skip the rule; direct connections skip the app layer | Gaps in classification, feature limits and copies outside the governed table | A broad service account, a rule that lives only in the prompt, or sensitive data copied into a retrieval index |
Read left to right, the table changes who is acting. In an application it is a person. In a warehouse it is a query. For an agent it is software working on someone’s behalf.
Identity is resolved differently in each setting, so the attributes differ. That is why data governance and AI governance overlap on enforcement without being the same job.
A context layer for AI agents is one way to give enforcement points a shared policy record and the same tags.
How is policy enforced in operational applications?
In operational systems the policy is enforced inside the source: the database, the application or the API layer.
What the policy carries
It must name the subject, the rows or fields in scope, and the command it covers. Reading and writing are different questions.
What the database enforces
According to the PostgreSQL row security documentation, once row security is enabled on a table with no policy defined, access defaults to deny.
A policy then carries a read check (USING) that decides which existing rows a user can see or modify, and a write check (WITH CHECK) that decides which new or changed rows are allowed to be stored. When only USING is set, PostgreSQL uses it for both, per the CREATE POLICY documentation, so a separate write check is for rules where writing should differ from reading. Row security decides which rows a user reaches. The national ID mask is a column rule, and PostgreSQL covers it with column-level privileges, where a role is granted SELECT on specific columns only, or with masking in the application or API layer.
Where operational enforcement fails
Enforcement fails for the roles and connection paths the policy never reaches. The same documentation notes that superusers and roles with the bypass attribute skip row security, that table owners skip it unless it is forced on the table, and that referential integrity checks skip it too. Anyone who connects straight to the database also skips whatever the application layer enforces.
Checks for the national ID rule
Four checks matter: who connects around the application, which roles hold a bypass, whether the column grants exclude the national ID for non-HR roles, and whether write rules exist alongside read rules. The same holds for agents, since one that reaches the database through a shared credential inherits every bypass that credential holds, which is the starting point for AI agent access control. Data access governance tools and embedded data governance are two ways to check these paths.
How is policy enforced for analytical use?
In analytics the policy fires in the warehouse or query engine at query time, through masking policies, row access policies, row filters and column masks. The rule has to carry a role or group mapping and a mask or filter expression, plus a classification tag when the rule is tag-based.
How warehouses enforce it
Warehouse documentation describes the same idea in different forms. A row access policy is evaluated when a query runs and filters the rows it reads (row access policy documentation). A masking policy changes values at query time while the stored data stays unchanged, and tag-based masking lets one policy follow a tag to every column that carries it (masking documentation). The documented example policy returns the real value only when the current role is the allowed one (using masking policies).
Another engine expresses row-level security as a table, a list of grantees and a filter expression that works like a WHERE clause, and principals not on the list see no rows (row-level security documentation). A third attaches row filters and column masks as SQL functions, and separate policies can apply them by governed tag (row filters and column masks documentation, listed in Sources).
Where analytical enforcement fails
A tag-based rule only reaches columns that carry the tag, so an unclassified column goes unmasked without any error. Each engine also documents features that do not work with the rule. Test whether a path such as preview, sampling or a clone is blocked or sidesteps it. A copy outside the governed table may not carry the source table’s rule.
Classification as a dependency
Of these dependencies, classification is the one a team controls most directly. A data governance taxonomy gives tags a consistent meaning, and classified data is also a starting point for AI-ready data. Atlan can import tags from a connected warehouse when it scans the warehouse’s metadata (a crawl). You can also turn on syncing so changes made in Atlan flow back to tags that already exist in the warehouse (Atlan documentation). Teams weighing this for BI users can read about self-service analytics governance and the semantic layer.
How is policy enforced for AI models and agents?
For AI, the policy is enforced outside the model, at the tool call and in the downstream system, while data-layer rules still fire when the agent queries. It has to carry the user’s identity, the tools the task may use, the sensitivity of the data, and an approval step for high-impact actions.
What OWASP says to enforce
According to OWASP LLM06:2025 on excessive agency, the causes are excessive functionality, excessive permissions and excessive autonomy. Its mitigations include least privilege, running actions under the user’s identity with minimal scope, and complete mediation, meaning authorization is checked in the downstream system rather than decided by the model.
Research on runtime checks
Research points the same way. A preprint by Palumbo and colleagues, Formal Policy Enforcement for Agentic Systems (arXiv preprint, 2026, not peer reviewed), describes a runtime monitor where policy is written in Datalog and each tool call is checked against the history of events before it runs. The decision sits outside the model, the same split between decision and enforcement points that NIST SP 800-162 describes.
Where agent enforcement fails
Three failure patterns follow from OWASP’s mitigations. When an agent runs on a broad service account with no verified end-user context, the data layer checks the account’s privileges instead of the person’s. A policy that lives only in the prompt leaves the decision to the model. And logging records what happened without preventing it: OWASP notes that logging and rate limiting limit damage but do not stop an action.
The national ID rule for an agent
Ask whose identity the agent carries, which tools the task allows, and what the warehouse rule sees when the query comes from an agent. A masking policy that checks the current session’s role evaluates whichever role the agent’s connection holds. If the agent connects as a shared service account, the policy sees that account’s role, so the national ID is either unmasked for everyone or masked for HR too. If the agent connects as the user, the session carries the person’s role and the same mask applies.
Copies of the data
Copies are the other gap. If national IDs are pulled into a retrieval index, an embedding store or a training set, no warehouse mask follows them. OWASP’s LLM08:2025 on vector and embedding weaknesses recommends fine-grained access controls and permission-aware vector and embedding stores, so classify the column before it is copied into the index, and leave it out unless retrieval enforces the permissions of the person asking. The enterprise AI agent guardrails checklist and data privacy for AI agents cover the surrounding controls, and how to secure multi-agent systems covers several agents at once.
How do you carry one policy across all three settings?
You carry a policy across settings by writing it once in business terms and making sure each enforcement point receives the attributes it needs.
Five steps
- Write the policy once in business language, with a named owner who can say what it means.
- Classify the data it covers, so that tags exist for rules to key off.
- For each setting, name the enforcement point and the attribute it needs.
- Pass the user’s identity through to that point instead of rebuilding it in every system. In the example, that means the agent connects as the person, so the session role the masking policy checks is the person’s.
- Test each point with the same case, here the national ID rule, and record which setting lets it through.
The fifth step shows where a policy and its enforcement disagree. Run it for HR and non-HR users, and for both per-user OAuth and a service account.
Central or native enforcement
The wider method is in how to implement an enterprise context layer, and context architecture for AI agents shows the shape of the result.
Step four involves a choice. NIST describes decision and enforcement points that can be central or distributed: a central point gives one place to change a rule, while native enforcement keeps the rule beside its data.
A shared policy record lets each point act on the same definition while each system still implements it. Either route works better when both people and software can read the rules, which is the case made for computational governance and governance rules that AI agents can read. Metadata management for enterprise AI underlies both.
Gartner and this concept
Gartner uses “policy enforcement” as a use case in its governance research, and this page applies the term to operational, analytical and AI data.
Where the term appears
The Magic Quadrant for Data and Analytics Governance Platforms (12 November 2026) lists policy enforcement as a use case, and you can get the report free from Atlan. The companion Critical Capabilities research has its own page. The same report describes this use case as:
“Operationalize, serve and automate the work of D&A governance stakeholders involved in enforcing policies (business data stewards, analytics stewards)”
Gartner names business data stewards and analytics stewards as the people who enforce policies. The named owner in step 1 is usually one of them.
How does Atlan approach policy enforcement across the three settings?
Atlan holds the business policy and the classification tags. The warehouse runs its own masking and row policies at query time, and those can key on the tags Atlan syncs.
Data policies in Atlan
Persona, purpose and deny policies act on queries and previews in Atlan and on requests made through the Atlan MCP server, which passes the authenticated user identity to Atlan.
Data policies in Atlan control who can query and preview data on Atlan assets, with column masking types such as showing the first four or last four characters, hashing, nullifying or redacting. Deny rules override grants and the Admin role, and policies apply through personas or purposes (Atlan documentation on data policies). In the running example, a masking data policy applied through a persona gives people outside HR masked values when they query or preview the assets it covers in Atlan. Direct warehouse queries stay under the warehouse’s own policies.
Agents and the Atlan MCP server
The Atlan MCP server uses either OAuth per user or a service account with exactly one persona, and Atlan then applies the same persona, role and domain policies as in the Atlan UI. A service account carries one persona’s policies, so per-user policies call for per-user OAuth. Tool allowlisting is fail-closed, so a tool not on the list is denied (Atlan documentation on MCP security). The Atlan MCP overview and the MCP server implementation guide cover setup.
Fit with the context layer
This follows the context layer idea: policies and classifications are held once, in a place each system can read. The core components of a context layer and the unified context layer explain the parts. For the wider argument, see why platform-native context layers fail and why AI agents need an enterprise context layer.
Get the free 2026 Gartner Magic Quadrant for D&A Governance Platforms from Atlan: the full market definition, vendor profiles and evaluation criteria
Get the Gartner reportFAQs about policy enforcement for operational, analytical and AI data
1. What is policy enforcement in data governance?
Policy enforcement is applying a written data rule at the point where data is read, changed or used. A policy has an owner, a scope and a rule, and an enforcement point such as a database, a warehouse or an agent’s tool call carries it out.
2. Is AI part of data governance?
Yes, for the data an AI system reads and the actions it takes on that data. Classification, ownership and access rules carry over, and the agent’s tool call becomes an additional enforcement point.
3. How is enforcing a policy for an AI agent different from enforcing it for a report?
An agent acts on someone’s behalf, so the policy needs that person’s identity and the tools the task may use. OWASP recommends authorizing each action in the downstream system rather than letting the model decide its own permissions.
4. What should a policy carry so it can be enforced?
A named owner, the data it covers, and the attributes the enforcement point keys on. NIST describes these as attributes of the subject, the object and, in some cases, the environment, which an access decision is computed from.
5. What should be included in an AI governance policy?
Scope, an owner, the sensitivity of the data involved, the tools an agent may use, an approval step for high-impact actions, and logging. Each item should map to an enforcement point that can act on it.
6. Can one tool enforce a policy in all three settings?
In practice each system enforces natively: the database, the warehouse and the agent’s tool call each apply the rule themselves. What can be shared is one policy record and one set of tags, so each point acts on the same definition. Atlan holds that record and those tags.
7. What is a data governance policy?
A data governance policy is a written rule about how data may be used, with an owner, a scope and the rule itself. It becomes real when an enforcement point such as a database, a warehouse or an agent’s tool call applies it.
8. Who should be responsible for writing a data governance policy?
Give each policy a named owner who can say what it means, usually the business owner of the data it covers. The teams that run the enforcement points then confirm they can act on it.
9. What pitfalls should you be aware of when writing a data governance policy?
The usual failure is at the handoff: the place that must act on the rule never receives the attribute it keys on. Unclassified columns go unmasked without an error, a broad service account hides who is asking, and copies of the data outside the governed table may not carry the rule.
10. What are the core components of AI data governance?
Classification, ownership and access rules carry over from traditional governance. For agents, add the tools a task may use, an approval step for high-impact actions and logging, each mapped to an enforcement point that can act on it.
11. Why is design-time governance not enough?
A rule written at design time only protects data if the system acting on it receives the right attributes when the request arrives. For agents, OWASP recommends checking authorization in the downstream system at the time of each action, and notes that logging limits damage but does not stop an action.
12. How is AI data governance different from traditional data governance?
The rules and ownership are the same, but an agent acts on someone’s behalf. That adds the tool call as an enforcement point and makes the user’s identity something the policy has to carry through, rather than a service account’s.
13. What is the biggest mistake enterprises make when enforcing data policies for AI?
Giving the agent more access than the task needs, then trusting the agent or its prompt to stay inside the policy. OWASP lists excessive permissions as a root cause of agent failures, for example a tool that only needs to read data but connects with an identity that can also update and delete. Enforce the policy in the downstream system, with the least access the task requires.
14. What are the benefits of a data governance policy?
A written policy gives every system the same rule to enforce, so the national ID rule means the same thing in an application, a warehouse and an agent’s tool call. It also names an owner and a scope, which makes decisions about access reviewable later. NIST notes that policies and procedures contribute to security and privacy assurance.
Sources
- NIST SP 800-162, Guide to Attribute Based Access Control (ABAC) Definition and Considerations, NIST
- Row Security Policies, PostgreSQL documentation
- CREATE POLICY, PostgreSQL documentation
- Understanding row access policies, Snowflake documentation
- Understanding column-level security and dynamic data masking, Snowflake documentation
- Introduction to BigQuery row-level security, Google Cloud documentation
- Row filters and column masks, Databricks Unity Catalog documentation (docs.databricks.com)
- LLM06:2025 Excessive Agency, OWASP GenAI Security Project
- LLM08:2025 Vector and Embedding Weaknesses, OWASP GenAI Security Project
- Column privileges, PostgreSQL documentation
- Using dynamic data masking, Snowflake documentation
- Formal Policy Enforcement for Agentic Systems, arXiv:2602.16708
- Data policies, Atlan documentation
- MCP security, Atlan documentation
- Manage Snowflake tags, Atlan documentation
- Gartner, Magic Quadrant for Data and Analytics Governance Platforms, 12 November 2026, Anurag Raj Et Al.
- Gartner, Critical Capabilities for Data and Analytics Governance Platforms, 12 November 2026, Sarah Turkaly Et Al.
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.