Real-time PII protection in Legal Tech

Discover how a legal tech solution for immigration law firms is helping legal teams to protect privileged client data in AI-powered workflows and to have control of prompts themselves.

Legal tech tools are increasingly leveraging AI assistants for client intake, summarizing case files, writing correspondence, answering status inquiries, and compiling filing information. In immigration practice, these conversations might include some of the most private information that a firm has: passport numbers, alien registration numbers, dates of birth, addresses of the client's home, employer, and family, medical records submitted as evidence, financial records submitted as evidence and the client's own account of his circumstances.

Challenges with Legal Tech: Safeguarding Privileged Client Data

The more AI capabilities that are integrated throughout a platform, the more data will be moved. It is carried by third-party model providers, prompt logs, error monitoring and product analytics, all things that make the product better, and all things that are not under the firm's control. That is a legal issue which doesn't have an easy fix: attorney-client privilege and confidentiality do not cease when you ask an API for information. It is important for companies to have sensitive identifiers that can be captured when companies leave the platform, and not afterwards.

The second complication is the immigration work. Narratives, affidavits and supporting documents must be supplied in the language of the client and the pattern-matching rules are written for English text, missing identifiers in Arabic script or transliterated names. No redaction layer in one language is a failure, NOT a control.

Attorneys, paralegals and case managers are the most distant from the code, in between that and the people who will be most directly impacted by the way that these AI assistants work. Information on immigration can be updated daily: forms are released, memos are issued, instructions for filing are revised. The legal team will make a request when the assistant requires an update, and the engineer will have to implement, test and deploy the update. Experiments stop and the client facing output is no longer guided by the prompt.

Our solution: Real-time AI Data Protection and Prompt Management.

RezunateLLM is middleware. It sits between the application code and the AI models that are   called by the platform and looks at each request before it is sent to a third party provider, log or analytics tool.

Middleware detects and classifies sensitive data (names, national ID and passport numbers, addresses, dates of birth, contact details, etc.) in English and Arabic in 18 categories of PII and flags, redacts, or blocks according to the platform's own rules. Other custom detection rules can be developed along with the default detection rules for firms with identifiers like alien registration numbers or USCIS receipt numbers. Hosted by RezunateLLM or completely self-hosted, the platform provides the firm with data residency choices.

Prompt management pushes the assistant's directions from the codebase to the central dashboard. Prompts are written and edited by legal teams. Every edit creates a new static version that never replaces the last one; this implies that the platform will always have the complete history of the actions the assistant has been asked to carry out and when, and it will roll back instantly if the output quality is lowered.

Both capabilities are seamlessly integrated into an existing AI stack without the need to rebuild or retrain the AI assistant, as RezunateLLM works at the data layer.

Impact Highlights

• Make sure that privileged client information is being safeguarded in real time before it can be accessed by third party AI service providers, logging systems and analytics vendors.
• Identifiers detected in both English and Arabic, filling the gap left by English-only pattern matching, in immigration case work.
• To ensure clients and staff privacy and data management needs and to keep the product operational.
• Maintained a history of each prompt change and had the ability to roll back to previous versions.
• Gave attorneys and paralegals the freedom to adjust AI instructions as guidance changes, without having to go through a developer deployment.
• Implementing at the data layer, without having to rebuild or retrain the current AI assistant.

Create and build AI with control.

Protect privileged data. Provide expert control. Have prompts that don't slow down the developers. Get real-time AI guardrails and prompt management all in one infrastructure layer with RezunateLLM.