- Enterprise h2oGPTe detects PII at three points: when a document is ingested, in the input sent to the model, and in the model's output.
- For each point you choose an action: allow, redact, or fail and stop the step, set per collection of documents.
- Detection combines regular expressions, a fine-tuned language model, and custom labels you define for a large language model to find.
- An agent can build a redaction tool for CSV and text files that keeps the layout and writes a log of what it removed and where.
Every external share starts with a manual scrub
Your audit, legal, and operations teams share documents with auditors, regulators, vendors, and counterparties every week. Before anything goes out, someone reads it line by line to remove customer names, addresses, tax numbers, account numbers, and card numbers.
The work is slow and the misses are expensive. A customer extract, an internal audit report, and an email thread all hold personal data in different shapes, and a person checking the tenth file is less careful than on the first.
The same risk now sits inside your AI tools. If a model can read customer documents, the personal data has to be handled before it reaches the model and before the answer reaches the user.
How PII detection and redaction works
Enterprise h2oGPTe includes configurable guardrails and PII redaction settings that combine several detection methods and let you choose the action at each point.
Step 1: choose what counts as PII for you
Pick the entity types to flag, such as names, addresses, phone numbers, emails, bank account numbers, and national identifiers.
Step 2: combine detection methods
Regular expressions catch fixed formats such as card numbers. A fine-tuned language model uses named entity recognition to find PII in free text, and a large language model can find custom labels you define, such as a date of birth.
Step 3: set the action at each point
For each collection, set what happens at ingestion, on model input, and on model output: allow, redact, or fail. Redact replaces the value with its type, such as a phone number label, or places censor bars over it in the stored document.
Step 4: redact files for external sharing with an agent
For files you share outside the organization, an agent can write and run a redaction tool for CSV and text files. It keeps the format readable, lets you choose which PII types to remove, and produces a log your reviewer checks before release.
Step 5: test on your own data, then scale
Run the settings on a sample of documents your team has already redacted by hand and compare the results. Detection depends on model predictions, so validate it on your data and keep a person signing off each release.
Two limits to plan for
PII detection is off by default in most environments, and it applies to documents ingested for retrieval, not to files uploaded for agent-only use. For those files, use the agent-built redaction tool described above.