Skip to main content
LEARNING CENTER

AI Data Classification: What to Share and What to Protect

Every time your team uses ChatGPT, Claude, or Copilot, data leaves your machine. Understand the four classification levels, classify your data types, and learn what should never reach a third-party AI model.

PublicInternalConfidentialRestricted

Why AI Data Classification Matters

When an employee pastes a customer's account details into ChatGPT to draft a response, that data leaves your network as plaintext. When a developer asks Claude to debug proprietary code, your intellectual property enters a third-party system. These aren't hypotheticals. They happen thousands of times a day in organizations that haven't classified their AI data flows.

AI data classification gives you a framework for deciding what's safe to share, what needs protection, and what must never leave your machine. It's the foundation every AI security program is built on, because you can't protect what you haven't classified.

The four classification levels, Public, Internal, Confidential, and Restricted, provide a common language that security teams, compliance officers, and end users can all understand. When everyone agrees that customer PII is Confidential and credit card data is Restricted, the rules for AI tool usage become clear.

The Four Classification Levels

Select a level to see what data it covers, the AI sharing rule, and how Shield enforces it.

Confidential

PROTECTION REQUIRED

Sensitive business information that would cause harm if disclosed, customer data, financial figures, proprietary algorithms, contracts, and strategic plans. Requires automated protection before any AI use.

Examples

Customer PII and account detailsFinancial projections and revenue dataProprietary algorithms and source codeLegal contracts and NDAsStrategic roadmaps and acquisition plans
AI Sharing Rule
NEVER share without automated redaction. All confidential data must be stripped or tokenized before prompts leave your machine. Manual review is not sufficient, use automated tools.
Shield enforcement: Shield applies all active filter packs, PII, PCI, secrets, credentials, and custom patterns. Confidential data is redacted at the token level before any model provider sees it.

Data Type to Classification Map

See how common data types map to classification levels, and the implications for AI tool usage.

Personal Identifiable Information (PII)
confidential

Names, email addresses, phone numbers, social security numbers, passport numbers, and any data that can identify an individual.

john.doe@company.comSSN: 123-45-6789Phone: +1 (555) 123-4567Passport: A12345678
Protected Health Information (PHI)
restricted

Medical records, diagnoses, treatment histories, insurance IDs, and any health data protected under HIPAA or similar regulations.

Patient ID: 98765, Diagnosis: hypertensionInsurance: BCBS-123456789Lab results and imaging reports
Payment Card Data (PCI)
restricted

Credit card numbers, CVV codes, expiration dates, bank account numbers, and any data within PCI-DSS scope.

Card: 4111-1111-1111-1111CVV: 123, Exp: 12/28Routing: 021000021, Account: 9876543210
Credentials & Secrets
restricted

API keys, database passwords, SSH private keys, OAuth tokens, connection strings, and any authentication secrets.

sk-abc123def456...DATABASE_URL=postgres://...-----BEGIN RSA PRIVATE KEY-----
Proprietary Source Code
confidential

Unreleased code, proprietary algorithms, internal libraries, and code containing business logic that provides competitive advantage.

function calculatePricing() { ... }// Proprietary ranking algorithm v3.2Internal npm package: @company/core
Legal & Contracts
confidential

Signed contracts, NDAs, settlement agreements, M&A documents, litigation materials, and attorney-client privileged communications.

Master Services Agreement dated 2026Exhibit A: Pricing ScheduleConfidential Settlement Terms
Financial Data
confidential

Revenue figures, profit margins, EBITDA, cap tables, valuation models, budget allocations, and any non-public financial metrics.

Q3 revenue: $12.4M (+18% YoY)Cap table: Series B investors2027 budget forecast spreadsheet
Internal Strategy
internal

Product roadmaps, launch plans, competitive analyses, hiring plans, and organizational changes not yet announced.

Q4 product launch: Project TitanCompetitive analysis: Competitor X feature gapOrg restructuring plan, draft

AI Data Classification Self-Assessment

Answer 5 questions to assess your organization's AI data classification posture. This is a self-assessment, not a formal audit.

Question 1 of 5

Do you know what types of data your team sends to AI tools?

Can you enumerate the categories of data, PII, financial, code, legal. That flow through ChatGPT, Claude, Copilot, or other LLMs in your organization?

Classification That Works Without Thinking

The weakness in every data classification program is the same: it relies on humans to remember the rules. Shield removes that dependency by automatically classifying and redacting data at the egress point, before any prompt leaves your machine.

Auto-detect
Four filter packs (PII, PCI, secrets, custom) classify data in real time as prompts are composed.
Auto-redact
Confidential data is tokenized before it reaches the network layer, the model never sees the original.
Auto-block
Restricted data triggers a hard block, the prompt is intercepted with an immutable audit record.
Auto-audit
Hash-chain verification provides tamper-evident logs for every classification decision, audit ready.

Frequently Asked Questions

Classify every prompt. Automatically.

Shield sits on your machine and classifies data before it reaches ChatGPT, Claude, Copilot, or any AI model. No policy to remember. No manual steps. Just protection that works.