Practical guide
How to operationalize data classification in AI workflows
A data-classification policy names categories and handling expectations. An operating workflow turns those expectations into decisions at the moments when data enters, moves through, and leaves an AI-enabled process. The workflow needs accountable owners, machine- and human-readable labels, routing rules, retained evidence, time-limited exceptions, and a review loop. This method is for.
A data-classification policy names categories and handling expectations. An operating workflow turns those expectations into decisions at the moments when data enters, moves through, and leaves an AI-enabled process. The workflow needs accountable owners, machine- and human-readable labels, routing rules, retained evidence, time-limited exceptions, and a review loop. This method is for implementing an organisation's approved classification scheme. It is not a substitute for jurisdiction-specific legal review, a claim of compliance, or proof that a control is effective in operation.
Define the control record
For each workflow, record the control objective, approved classification scheme, system boundary, data owner, workflow owner, security and privacy reviewers, allowed tools, prohibited transfers, evidence retained, exception authority, and review date.
NIST SP 800-60 provides a method for mapping information types to security categories and impact levels. NIST SP 800-53 Rev. 5 includes controls for security and privacy attributes, including their assignment, retention, and interpretation. NIST's Data Classification Practices project, currently published as an initial public draft, illustrates classification as a lifecycle problem that combines discovery, identification, communication, enforcement, and monitoring. Use the draft as implementation context rather than as a final requirement.
Operate the CLASSIFY route
Use eight states: Capture, Locate, Assign, Stamp, Send, Inspect, File exception, and Yearly or event-driven review. “Yearly” is a memory aid; the actual cadence should match the organisation's risk and policy.
### Capture the use case
The requester describes the business task, intended AI use, source systems, expected outputs, recipients, retention, and decision consequence. Reject vague requests such as “use customer data with AI.” Name the actual fields, documents, transformations, and destinations.
### Locate the data and owner
Inventory direct inputs, retrieved context, prompt content, system logs, intermediate files, model outputs, feedback, and downstream exports. Link each dataset to its authoritative source and accountable data owner. Unknown ownership blocks routing until resolved.
### Assign classification
Apply the approved scheme to each input based on information type, sensitivity, regulatory or contractual constraints, and potential impact. When items with different labels are combined, apply the organisation's combination rule. A conservative default is for the working set and derivative output to inherit the highest applicable restriction until an authorised reviewer documents a lower classification.
### Stamp attributes
Attach the classification, owner, handling rule, source, review date, and relevant caveats in formats that survive the workflow. Use visible labels for people and structured attributes for systems. Preserve labels when content is copied, summarised, embedded, exported, or logged.
### Send through an allowed route
Translate labels into executable rules: approved tool and account, permitted region, access group, encryption, retention, logging, export destination, and deletion route. The workflow should block or quarantine a transfer when attributes are missing, inconsistent, expired, or unsupported by the destination.
### Inspect the transaction
Before and after processing, test that the actual data, tool, permissions, output, and destination match the approved route. Retain enough evidence to reproduce the decision: request identifier, source references, labels, rule version, tool, actor, timestamps, disposition, and output location. Sample operating records for misclassification, lost labels, inappropriate access, and outputs that reveal restricted input.
### File a time-limited exception
An exception record states the business need, affected data, rejected alternatives, compensating measures, authority, start and expiry, monitoring, and revocation trigger. The requester cannot approve their own exception. Expired exceptions block the route rather than silently becoming permanent practice.
### Yearly or event-driven review
Review when policy, data, tool behaviour, ownership, contracts, law, incidents, or destinations change, and at the scheduled control date. Reconcile records against the real workflow. Close obsolete routes and update labels or rules before resuming.
Worked artifact: classify a support-summary workflow
A synthetic customer-operations team wants an assistant to summarise support cases for a weekly service review. The source set contains public product documentation, internal operating notes, customer contact details, free-text tickets, and an account-risk flag. The proposed destination is a shared management workspace.
The operating record produces this route:
| Item | Assigned handling | Workflow action | Evidence | | --- | --- | --- | --- | | Public product documentation | Public | May enter approved assistant | Source URL and retrieval date | | Internal operating notes | Internal | Restricted to approved workspace group | Owner, label, access list | | Customer contact details | Restricted personal data | Remove before processing unless separately approved | Field-level transform log and sample check | | Free-text tickets | Restricted pending review | Detect and redact identifiers; quarantine failures | Rule version, exception queue, reviewer | | Account-risk flag | Restricted decision context | Exclude from the summary task | Block event and owner decision | | Generated summary | Inherits restricted classification | Publish only to named management group | Output label, source trace, destination record |
During the first test, a ticket contains a phone number in an unexpected field. The redaction rule misses it, and the pre-send inspection quarantines the batch. The workflow owner updates the detector, reruns a representative test set, records the false-negative finding, and obtains reviewer sign-off before resuming. The output does not get downgraded merely because obvious identifiers were removed; its sources and remaining content still determine handling.
An urgent request to email the summary to an external adviser is not routed automatically. The owner must establish the adviser's authority, destination controls, minimum necessary content, and applicable agreement. If the exception authority approves a narrower export, the record expires after the stated review and cannot authorise future transfers.
Test whether the workflow operates
Test both expected and adverse cases. Include missing labels, conflicting labels, mixed-classification inputs, copied content, unstructured identifiers, expired exceptions, unauthorised destinations, tool changes, and deletion requests. For each test, record the expected route, actual route, evidence, defect owner, correction, and retest result.
Useful operating evidence includes the proportion of transactions with complete attributes, blocked transfers, label-loss events, exception age, access mismatches, detected restricted fields, completed deletion actions, and unresolved defects. These measures show how the workflow behaved; they do not by themselves establish legal compliance or overall control effectiveness.
Two Prune can assess how a participant navigates evidence, applies constraints, verifies outputs, and communicates a decision in a simulated workflow. It does not certify an organisation's data controls or make automated employment decisions.
Sources
NIST SP 800-60 Volume I 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]
Sources
- 1.NIST SP 800-60 Volume I Revision 1, Guide for Mapping Types of Information and Information Systems to Security Categories · National Institute of Standards and Technology
- 2.NIST SP 800-53 Revision 5, Security and Privacy Controls for Information Systems and Organizations · National Institute of Standards and Technology
- 3.NIST SP 1800-39 Initial Public Draft, Data Classification Practices · National Institute of Standards and Technology