FIELD GUIDE

Data Classification Before Generative AI: A Four-Tier Workflow

October 25, 2026Verified: 2026-10-25

The most important question is usually not which prompt to write, but whether the underlying material may leave a given system. “Company data” is too broad a category for decisions about AI uploads.

The workable objective is not to remove every human touch. It is to produce a specific, inspectable output with known sources, clear ownership and an escalation path. The method below is intended for security, privacy and operations teams. It can be implemented with approved documents, office software or a restricted AI assistant; no feature or licensing tier is assumed.

Decide whether AI belongs in this process

Define four internal tiers for pilot purposes: public, internal, confidential and restricted. These are proposed organizational categories, not a universal law or product feature.

Before evaluating vendors, write a one-sentence acceptance test: who needs the result, which facts must be correct, what may be omitted, and what must never happen. Compare the result against the current procedure, including the cost of review. If ordinary software can reliably solve the task, adding a language model introduces an unnecessary failure mode.

A practical implementation workflow

1. Establish the real input and the authority to use it

Identify where the original data resides and whether the operator is authorized to use it for the proposed purpose.

Before making a tool request, distinguish facts drawn from the source system from instructions written by staff. Record which materials are approved, who owns the process and what might be missing. If the evidence cannot be trusted, defer the experiment rather than trying to fix uncertainty with prompt wording.

2. Design an inspectable assisted step

Label test inputs and classify mixed documents at the highest relevant tier, including attachments and inferred personal information.

Make the proposed output visible alongside the evidence the reviewer needs. Separate transformation of information from the authority to send, publish, decide or change a record. Where a step involves sensitive information or irreversible consequences, require a more restrictive control path.

3. Define quality criteria and exception routes

For each tier approve specific environments, retention settings, sharing permissions and any required redaction or pseudonymization.

Use difficult examples from the actual task distribution, not only clean demonstrations. Include missing context, conflicting records and requests that must be declined. Record false positives, false negatives and work passed back to the original team so the trial measures total work rather than the model response alone.

4. Review, decide and retain the evidence

Check generated outputs for leaked names, account details or copied confidential passages before redistribution.

The accountable human should be able to correct, reject and trace the result without searching through a raw conversation. Archive the approved artifact, source version and decision using organizational retention rules. Recheck the workflow when the data source, model behavior or business policy changes.

Approval checkpoints, privacy and controls

Use role-based access, least-privilege source folders and an escalation path to privacy/legal specialists. A contract promising no model training does not alone settle access, logging, residency or retention.

Set an explicit release gate for the output: a draft can be accepted internally, but any external message, business-system write, financial consequence or decision about a person needs the authorization appropriate to its risk. Record the role of the approver rather than relying on the assistant’s confidence. Access should be scoped to the least information needed for the particular task.

What can go wrong

Redaction can fail when context re-identifies someone. Generative summaries may combine non-sensitive fragments into sensitive inferences; treat output classification as a separate review.

Governance measures are controls that reduce exposure; they are not guarantees of safety or universal evidence of statutory compliance. Keep technical ownership, business accountability and legal interpretation separate.

Illustrative workflow, not a tested case study

A finance team tests expense-note categorization with synthetic records first, then a limited approved dataset with names masked and reviewer comparison to originals.

To make this example an actual pilot, the organization would select representative historical cases, remove or protect personal information, agree on expected outputs before testing, and record reviewer edits. A fabricated success rate is less useful than a small set of well-documented failures that the team can learn from.

Measure outcomes rather than plausible productivity

  • Classification Coverage: define the denominator and measurement period in advance; compare with a valid non-AI baseline and record corrections.
  • Unauthorized Disclosure Attempts: define the denominator and measurement period in advance; compare with a valid non-AI baseline and record corrections.
  • Access Exceptions: define the denominator and measurement period in advance; compare with a valid non-AI baseline and record corrections.
  • Redaction Misses: define the denominator and measurement period in advance; compare with a valid non-AI baseline and record corrections.
  • Approved-Data Usage: define the denominator and measurement period in advance; compare with a valid non-AI baseline and record corrections.

Keep three ledgers separate: process quality, staff time and full operating cost. Cost includes the service or seats, security review, training, integration, human checks, support and possible rework. Do not treat time released as realized revenue or wage savings unless the organization can demonstrate that conversion.

Limits, alternatives and regional applicability

Choose on-premises processing, manual review or deterministic filtering where external processing is not authorized.

Privacy law can regulate personal data regardless of internal labels. UK GDPR, EU GDPR and local sector rules require separate assessments.

Primary-source context and applicability

  • The ICO describes its AI data protection guidance as best practice and legal interpretation, not a statutory code; the page was under review when checked.
  • The ICO supplies an AI and data protection risk toolkit for organizations; its guidance page notes an ongoing review.
  • NIST published the Generative AI Profile, a companion to the AI RMF, on 26 July 2024.
  • Microsoft documentation states organizational Copilot access follows existing user permissions; exact protections depend on offering and contract.
  • Google Workspace privacy documentation updated 6 October 2026 describes data treatment for Gemini in Workspace and related business offerings.
  • OpenAI Enterprise Privacy, updated 8 January 2026, states business data is not used to train models by default; terms depend on product.

These dated institutional publications and vendor documents establish the limited context stated above. They do not constitute proof that the suggested workflow is effective or compliant in every organization.

Tool and model availability, enterprise agreements, data-handling terms, feature switches and pricing change frequently. This article deliberately does not claim a current price, unlimited entitlement, universal release date or a tested superiority of one vendor. Confirm product, tier and geography from dated first-party documentation before purchase; use the AI 10 price ledger for price snapshots.

This guide is an editorial operating method, not a product benchmark, a controlled experiment, a legal opinion or a promise of savings.

Sources & Further Reading

CONTINUE THE RESEARCH
CONTINUE EXPLORINGExplore related files ↗