Security, Privacy, and Information Protection

Insurance documents contain personal, medical, financial, legal, contractual, and commercially sensitive information. Document Intelligence expands the number of systems and processing stages that may touch this content. This chapter examines the security and privacy controls required across the full lifecycle.

A useful AI capability can create unacceptable risk if information is exposed, retained improperly, used beyond its approved purpose, or transmitted to an unauthorized service. Security and privacy must be built into architecture and operating procedures from the beginning.

Data protection is not a feature added after the model works. It is a condition of whether the model should be allowed to work at all.

Documents should be classified by sensitivity, confidentiality, regulatory treatment, privilege, and business impact. Classification should drive storage, access, processing, sharing, and retention controls.

Users, applications, services, vendors, and models should receive only the access required for their approved purpose. Access should be role-based or attribute-based and reviewed regularly.

Sensitive information should be encrypted in transit and at rest. Key ownership, rotation, separation, backup, and incident procedures should be defined independently from application convenience.

Production processing should occur in approved environments with controlled network paths, hardened configurations, logging, vulnerability management, and separation from development and testing.

Privacy controls should address purpose limitation, data minimization, notice, consent where applicable, access, correction, deletion, and secondary use. The organization should know which information is necessary for each use case and avoid collecting or retaining more than required.

Contracts and technical controls should define whether document content, prompts, outputs, embeddings, corrections, and metadata may be retained or used for model training. Vendor terms should not be assumed to match the insurer’s obligations.

Documents may contain malicious instructions, hidden text, links, or content designed to influence a model. Systems should treat documents as untrusted input, isolate tool authorization from model output, and validate actions independently.

Embeddings and extracted features may retain sensitive meaning even when they do not resemble the original document. They require access, retention, deletion, and incident controls comparable to the underlying content.

Logs are necessary for security and audit but can become secondary repositories of sensitive information. Logging should capture what is needed for accountability while masking or minimizing content.

Original documents, processing copies, extracted data, prompts, outputs, and caches may have different retention requirements. Destruction should propagate across copies and derived stores where required.

Incident plans should address data exposure, unauthorized model access, compromised credentials, malicious documents, service outages, and corrupted outputs. Business continuity should include safe manual fallback.

Vendors should be evaluated for security architecture, privacy practices, subprocessors, hosting, incident notification, model use, vulnerability management, audit evidence, and exit support.

Create a data-flow map for every use case showing what information enters, where it is processed, what is derived, who can access it, how long it is retained, and how it is deleted.

An insurer planned to use a general-purpose AI service to summarize claim documents. A privacy review found that the proposed configuration retained prompts and outputs for service improvement and lacked the required regional and contractual controls. The insurer selected a controlled enterprise deployment, disabled secondary use, restricted data categories, added content filtering, and documented retention and deletion. The use case proceeded only after the information flow matched the organization’s obligations.

Security and privacy decisions belong to executive risk management, not only technical configuration. The organization must understand the complete information lifecycle, including derived data and vendor use, before approving production deployment.

Document Intelligence security requires classification, least privilege, encryption, secure environments, privacy-by-design, controlled vendor use, protection against content-based attacks, retention discipline, incident response, and third-party oversight.