PII Detection & Redaction in Enterprise Documents: A Complete Guide

Why PII Detection Matters More Than Ever

When you feed documents into an AI system — whether for search, summarization, or analysis — you're potentially exposing sensitive data to models, logs, and retrieval pipelines. A single SSN in a financial report, a patient name in a medical record, or an AWS access key in a configuration file can trigger compliance violations, data breach notifications, and regulatory fines.

The solution isn't to avoid AI. It's to scan every document before it enters the AI pipeline and automatically redact or flag sensitive content. This guide covers the three detection approaches, their strengths and weaknesses, and how to implement them at scale.

The Three Layers of PII Detection

Modern PII detection uses a layered approach. No single technique catches everything, but together they provide comprehensive coverage.

Layer 1: Regex Pattern Matching

The first line of defense. Regex patterns match structured identifiers with predictable formats:

• SSN (US): \d{3}-\d{2}-\d{4}
• Credit card numbers: \d{4}[\s-]\d{4}[\s-]\d{4}[\s-]\d{4} with Luhn algorithm validation
• Email addresses: [a-zA-Z0-9._%+-]+@[a-zA-Z0-9.-]+\.[a-zA-Z]{2,}
• Phone numbers: Multiple formats across countries
• Passport numbers, driver's licenses: Country-specific patterns

Strengths: Fast, deterministic, no ML required. Weaknesses: High false-positive rate (e.g., "123-45-6789" could be a part number, not an SSN). Misses unstructured PII (names, addresses in prose).

Layer 2: ML-Based Named Entity Recognition (NER)

Machine learning models trained to identify entities in natural language text:

• Person names — "John Smith submitted the report"
• Organizations — "Acme Corp's Q4 filing"
• Locations — "shipped to 123 Main St, Springfield"
• Dates of birth, medical terms — contextual identifiers

Modern NER models (like DeBERTa-v3 fine-tuned on PII datasets) can detect PII in unstructured prose where regex fails completely. They understand context — distinguishing "Apple released a new iPhone" (organization) from "Apple is my favorite fruit" (not PII).

Strengths: Catches unstructured PII that regex misses. Context-aware. Weaknesses: Requires GPU for real-time processing. Model quality varies by language and domain.

Layer 3: Entropy-Based Secret Detection

API keys, access tokens, connection strings, and private keys have high entropy (randomness). Secret detection uses Shannon entropy and pattern heuristics to identify:

• AWS access keys — AKIA... pattern + high entropy
• GitHub tokens — ghp_... pattern
• JWT tokens — eyJ... base64-encoded structure
• Private keys — -----BEGIN RSA PRIVATE KEY-----
• Connection strings — Server=...;Database=...;User Id=...;Password=...

Strengths: Catches secrets that regex and NER miss. Weaknesses: Can flag high-entropy non-secret strings (base64-encoded images, hashes).

Detection Pipeline Architecture

In production, these three layers run as a pipeline:

1. Document ingestion — file is uploaded or synced from a connector
2. Text extraction — OCR for scanned PDFs/images; native text extraction for DOCX, PPTX, etc.
3. Regex scan (Layer 1) — fast pass to catch structured PII
4. NER scan (Layer 2) — ML model identifies unstructured PII in prose
5. Secret scan (Layer 3) — entropy analysis for API keys and tokens
6. Policy evaluation — each finding is scored (confidence 0–1) and matched against configurable rules
7. Action — auto-redact, flag for review, or mark safe based on policy

Redaction Strategies: Auto vs. Manual

Once PII is detected, you have three options:

• Auto-redact — high-confidence findings (≥0.90) are automatically replaced with [REDACTED] or category labels like [SSN], [EMAIL]. Best for high-volume, low-risk scenarios.
• Flag for review — medium-confidence findings (0.70–0.90) are surfaced in a review queue for human verification. Best for legal, healthcare, and financial documents where false positives are costly.
• Mark safe — low-confidence findings or known false positives are marked safe and ignored in future scans. This builds a learning feedback loop.

Common Pitfalls to Avoid

1. Regex-only approaches — they miss 40–60% of PII in unstructured documents. You need NER.
2. Scanning too late — if you scan after documents enter the vector index, PII is already exposed. Scan before indexing.
3. Ignoring OCR text quality — scanned PDFs with poor OCR produce garbled text that regex can't match. Invest in good OCR.
4. No re-scan on update — when a document is updated, re-scan it. New versions can introduce new PII.
5. Storing raw findings in logs — if your detection log includes the actual SSN that was found, you've just created a new PII leak. Log finding types and positions, not values.

Compliance Frameworks and PII Detection

Different regulations require different detection coverage:

How CordonData Handles PII Detection

CordonData's compliance scanning pipeline runs all three detection layers before any document enters the vector index or LLM context window:

• Regex engine — 50+ built-in patterns for SSN, email, phone, credit card, passport, driver's license, and more across multiple countries
• ML-based NER — DeBERTa-v3 fine-tuned on PII datasets, running on CPU for deployment flexibility
• Entropy-based secret detection — identifies API keys, tokens, connection strings, and private keys
• Configurable policies — per-knowledge-base rules for auto-redact vs. flag-for-review thresholds
• Re-scan on update — new document versions are automatically re-scanned
• Audit trail — every finding is logged with confidence score, detection method, and reviewer action

Getting Started with Automated PII Detection

1. Inventory your documents — what types of PII exist in your document corpus?
2. Define your policy — which PII types require redaction vs. flagging? What confidence thresholds?
3. Run a baseline scan — scan a representative sample to understand your PII surface area
4. Tune your rules — adjust patterns and thresholds based on false positive/negative rates
5. Integrate into ingestion — make scanning a mandatory step before any document enters your AI pipeline
6. Monitor and iterate — review flagged findings, mark false positives, and continuously improve

CordonData includes automated PII/NHI/secret detection as a built-in feature. It runs on-premise or in your cloud — your documents never leave your infrastructure during scanning. Join the waitlist for early access.