Practical guide

Data classification policy for AI-enabled work

A data classification policy defines how an organisation identifies information types, assigns approved classification values, carries those values through AI-enabled workflows, applies handling rules, and resolves exceptions. It should create decision rights and testable requirements. It should not be a list of labels without owners, procedures, or evidence. This guide is educational and is not.

By Two Prune

A data classification policy defines how an organisation identifies information types, assigns approved classification values, carries those values through AI-enabled workflows, applies handling rules, and resolves exceptions. It should create decision rights and testable requirements. It should not be a list of labels without owners, procedures, or evidence. This guide is educational and is not legal advice or a compliance determination. Classification obligations depend on jurisdiction, contract, sector, information type, and system design. Obtain current specialist review before adoption.

Define scope before labels

State which people, repositories, systems, prompts, outputs, attachments, exports, and third parties are in scope. Include temporary files and transformed content where relevant. Define exclusions and who can approve them.

NIST SP 800-60 provides a method for mapping information types and systems to security categories based on potential impact. Start with the information and business process, not the storage location or file extension. A public press release and a restricted incident record may both be PDFs, but their classification basis differs.

The policy should also define its relationship to access control, retention, privacy, incident response, records management, procurement, and acceptable AI use. A classification value can trigger those processes, but it does not replace them.

Write seven enforceable clauses

### 1. Purpose and control objective

Explain why classification exists and what it should enable. A testable objective might be:

> In-scope information is assigned an authorised classification based on content and context before it is entered into, processed by, or exported from an AI-enabled workflow, and the classification remains traceable through defined transformations.

### 2. Permitted values and interpretation

For each value, define criteria, examples, non-examples, default treatment when uncertain, and the owner of interpretation. Keep the taxonomy small enough to apply consistently. If another policy supplies the enterprise taxonomy, link to its version rather than copying it.

### 3. Roles and authority

Name the policy owner, data owner, system owner, user responsibilities, security and privacy reviewers, exception authority, and audit or assurance role. State who may assign, change, challenge, and remove a classification.

### 4. Assignment and inheritance

Define when classification occurs, whether it can be proposed automatically, how inherited attributes work, and when a person must confirm the result. NIST SP 800-53 control AC-16 addresses the association, permitted values, interpretation, and retention of security and privacy attributes. A tool suggestion should not silently become an authoritative decision unless the policy explicitly permits that route.

### 5. Handling in AI-enabled work

Connect each value to approved systems, identities, sharing routes, prompt and attachment rules, output review, logging, retention, export, and deletion requirements. State how classification applies when AI-generated output combines or summarises multiple sources.

### 6. Exceptions and incidents

Require a reason, scope, owner, compensating action, expiry, approval, and review date. Distinguish a planned exception from a suspected misclassification or disclosure event. Route incidents through the applicable incident process rather than treating them as ordinary exceptions.

### 7. Evidence and review

Define the records retained: taxonomy version, assignment events, overrides, transfers, scan coverage, failed classifications, exception decisions, remediation, and periodic owner review. Set a review trigger for system, law, contract, taxonomy, or workflow changes.

Add an operating procedure

The policy should point to a procedure that a worker or system owner can execute:

For automated classification, document the rule version, coverage, review route, and failure handling. NIST's initial public draft practice guide on data classification describes using technologies to discover, identify, and label sensitive unstructured data. Because it is a draft, use it as implementation context rather than a final requirement.

  • Identify the information type, source owner, purpose, and workflow.
  • Select the permitted class using the current criteria.
  • Record the basis and classifier, whether person or authorised rule.
  • Apply the required attribute and handling controls.
  • Confirm that the attribute persists through allowed transfers and transformations.
  • Escalate uncertainty, conflict, or unsupported formats.
  • Review changed context and remove or update the value only with authority.

Worked example: classify material for an AI-assisted supplier review

Consider a synthetic supplier-review workflow with four permitted values: Public, Internal, Confidential, and Restricted. The policy says Restricted information cannot enter the approved AI assistant; Confidential information requires an approved workspace and named purpose; Internal information is allowed in the approved workspace; Public information follows the normal content process.

A reviewer receives three items:

| Item | Classification decision | Handling | | --- | --- | --- | | Published supplier sustainability report | Public, after source and version confirmation | May be used in the approved workspace | | Internal service-performance table | Confidential, assigned by service owner | Use only in the approved workspace; retain source trace | | Incident record containing security details | Restricted under the incident taxonomy | Do not enter into the assistant; use the restricted review route |

The reviewer asks the assistant to summarise the first two items. The output remains Confidential because it incorporates the service-performance table. A suggested prompt that would include the incident record is blocked. If the service owner needs an exception, the request must name the purpose, fields, system, duration, compensating actions, and approving authority. The exception expires after the review and does not alter the incident record's classification.

This example shows policy operation but does not establish that a tool enforces the policy or that the organisation meets legal requirements.

Test policy quality before approval

Use a tabletop review with representative information and workflows. Check whether reviewers can:

Record disagreements and repair the criteria rather than averaging them away. After adoption, test actual population coverage, assignment events, samples, and exceptions separately. Policy design evidence is not operating evidence, and operating evidence is not proof of a downstream outcome.

Two Prune may assess how a participant identifies information types, follows a stated policy, handles conflicting evidence, verifies AI-assisted output, escalates uncertainty, and communicates a decision in a simulation. It does not certify the policy, determine legal compliance, or replace the accountable data owner.

  • reach the same value from the written criteria;
  • identify the accountable owner and permitted assignment route;
  • apply handling rules to prompts, attachments, outputs, and exports;
  • preserve the classification through a transformation;
  • route uncertainty and time-limited exceptions; and
  • identify the evidence needed for later operating review.

Sources

NIST SP 800-60 Volume 1 Revision 1, Guide for Mapping Types of Information and Information Systems to Security Categories

  • NIST SP 800-53 Revision 5, Security and Privacy Controls for Information Systems and Organizations
  • NIST SP 1800-39 Initial Public Draft, Data Classification Practices

Sources: [1] [2] [3]