AI Data Discovery: How LLMs Find PII and PHI in Your Database

You can't protect what you don't know exists. That single sentence is the entire reason AI data discovery matters. Every enterprise database holds personally identifiable information (PII), protected health information (PHI), financial data, or authentication secrets - and almost no enterprise can produce an accurate, current list of where those data elements actually live.

In this post we'll explain how AI data discovery works, why large language models are dramatically better at this than the regex-based scanners of the last decade, and how a platform like SWOT DAM 3.0 uses an LLM-driven "Data Sentinel" to classify sensitive columns automatically - even when the column names tell you nothing.

The Problem AI Data Discovery Solves

A modern enterprise database can have hundreds of tables and tens of thousands of columns. The schemas are typically the product of decades of organic growth - different teams, different naming conventions, different abbreviation rules, and a lot of legacy. The result is column names like:

• USR_EML_V2 - an email address
• TX_AMT_GROSS - a financial transaction amount
• PT_DOB_ENC - a patient's date of birth, lightly encoded
• SSN_TAIL_4 - the last four digits of a social security number
• AUTH_BLOB - a hashed password column

A regex scanner looking for email, amount, dob, ssn, or password will miss every single one. Compliance frameworks, data loss prevention, and breach disclosure laws don't care about your naming convention - they apply regardless. The gap between "what your scanner found" and "what's actually sensitive" is where data breaches hide.

What AI Data Discovery Actually Does

An AI data discovery module takes a different approach. Instead of pattern matching on column names, it sends the full schema context - table names, column names, data types, sample data shape - to a large language model and asks it to classify each column.

The LLM understands intent. It knows that USR_EML_V2 is almost certainly an email column because of the surrounding context: it sits in a USERS table, near columns called USR_NM and USR_PHN. It knows that TX_AMT_GROSS in a table called TRANSACTIONS is a financial amount. The model is reading the schema the way a human DBA would.

The result is a structured classification per column:

• PII - names, emails, phone numbers, addresses, IDs
• PHI - patient identifiers, medical record numbers, diagnoses, treatments
• Financial - account numbers, transaction amounts, balances
• Authentication - password hashes, tokens, API keys, secrets
• Classified / confidential - anything tagged as restricted by your organization's policy

Each classification comes with a confidence level (high, medium, low) so analysts can prioritize their review.

How It Works End to End

Inside SWOT DAM 3.0, the "Data Sentinel" module runs as a scheduled job - typically every 24 hours. The flow looks like this:

1. Iterate every active database collector in the platform.
2. Connect to the target database with read-only metadata permissions.
3. Extract the schema: table names, column names, data types, sample row count.
4. Construct a natural language prompt describing the schema.
5. Send the prompt to the configured LLM (OpenAI, Google Gemini, or a self-hosted Ollama model).
6. Parse the structured JSON response identifying sensitive columns.
7. Diff against the previous scan - flag new sensitive columns as discoveries.
8. Surface findings in the dashboard with a "new since last scan" indicator.

The 24-hour cadence is configurable. The model provider is configurable - which is important for organizations with data sovereignty requirements.

Why LLM-Based Discovery Beats Regex

Three reasons:

1. Context. A regex looks at one column name at a time. An LLM sees the entire schema and uses the surrounding columns as evidence for what each column actually contains.
2. Obfuscation tolerance. A regex pattern for "email" won't match USR_EML_V2. An LLM does.
3. Maintenance burden. Regex libraries have to be updated as your business evolves. An LLM-driven classifier generalizes.

Data Sovereignty: Why Provider Choice Matters

The biggest objection most security teams raise to AI data discovery is reasonable: "I don't want my database schema to leave my network." This is a legitimate concern in regulated industries - healthcare, finance, government - and the answer is provider flexibility.

SWOT DAM 3.0 supports three different LLM providers:

• OpenAI - strong model quality, cloud API
• Google Gemini - strong model quality, cloud API
• Ollama - local, self-hosted open-source models that run inside your network

For organizations where schema metadata cannot leave the network, the Ollama deployment gives you AI data discovery with no external API call at all. The platform is identical; only the model location changes.

What You Do With Discovery Results

Discovery is the input; the value is in what you do downstream.

• Tighten alert rules. Auto-generate "alert on bulk SELECT" rules for every newly discovered PII column.
• Inform PAM scoping. Make sure any role that touches PHI is flagged as high-sensitivity for SWOT PAM's risk engine.
• Tag for compliance. Map every discovered column to the relevant compliance frameworks - GDPR, HIPAA, PCI DSS, CCPA - and produce a per-framework data inventory.
• Inform DLP and masking. Hand the column list to your data loss prevention and dynamic-masking tools so they protect the right fields.
• Audit readiness. "Where do you store PHI?" stops being a research project and becomes a one-click report.

The Bottom Line

AI data discovery is the inventory layer that the rest of database security depends on. Without it, you're defending against threats to data you can't fully list. With it, every other control - alerting, PAM, DLP, compliance reporting - gets sharper.

If you're running multiple databases and your current data discovery story is "we ran a regex scanner three years ago," you have a gap worth closing. SWOT DAM 3.0 includes AI-powered discovery as standard - and you can run the model locally if your environment requires it. Book a demo with our team to see it on real schemas.

- Rudratic Technologies. Related reading: What is DAM?