Why GenAI demands a new DLP architecture
GenAI DLP must operate on unstructured text in real time, at the speed of human typing. It must understand context, not just patterns. It must cover every AI tool your employees use, not just the ones IT manages. And it must produce audit evidence that satisfies regulators who are specifically asking about AI data handling. Legacy DLP was not built for any of these requirements.
The inspection architecture
Effective AI DLP inspects prompts at three points: at the browser or endpoint (before the prompt leaves the device), at the API gateway (for programmatic AI access), and in the network proxy (as a catch-all). Each layer has different coverage characteristics and latency implications. Most enterprises deploy at least two of the three.
Classification for unstructured text
Enterprise AI DLP classifiers must handle the full range of sensitive data types without relying on regex: PII that appears in natural language, confidential business information that has no consistent format, source code that may be embedded in a paragraph of text, and legal privilege markers that depend on context. Training these classifiers on your organization's own data dramatically improves accuracy.
The redaction vs. block decision
Not every sensitive data detection should result in a blocked prompt. A well-designed policy engine distinguishes between data that should be redacted (PII that doesn't need to be in the prompt), data that requires a confirmation step (confidential documents), and data that should block outright (privileged legal content, regulated financial data). Blanket blocking drives users to workarounds.
Audit evidence that satisfies regulators
The logs your AI DLP generates must capture: the timestamp and user identity, the AI tool used, the data classification that triggered the event, the action taken (redact, block, allow with logging), and the sanitized prompt for audit review. This structured format is what compliance teams need to respond to regulatory inquiries about AI data handling practices.