Investment firms hold information whose value depends on controlled access: pipeline, investment memoranda, valuation assumptions, investor records, portfolio performance, contracts, management discussions, and diligence evidence. Leakage can happen through a public chat interface, an over-permissioned agent, copied spreadsheets, logs, generated output, or a connected service—not only through a cyberattack.

Start with the data, not the model

Build a register of AI use cases and the data each one requires. Classify data by sensitivity and business consequence, then define whether it may be processed, where, by whom, and under which retention and review rules.

Purpose → Data class → Approved environment → Minimum access → Human authority → Logging → Retention

If one link is undefined, the use case is not ready for production.

A practical classification

Classification should drive controls. A public-content assistant and a diligence agent should not share the same permissions, logs, or deployment pattern.

Control the complete data path

Common failure patterns

“Enterprise AI” is not itself a control. Teams still fail when a connector inherits broad access, a shared account removes accountability, source permissions are ignored during retrieval, prompts and outputs are retained unnecessarily, or an agent can send, publish, or update records without approval.

A minimum governance pack

  1. AI use-case and system inventory.
  2. Data classification and approved-use matrix.
  3. Vendor and model assessment.
  4. Role and tool-permission map.
  5. Human-approval and exception rules.
  6. Logging, retention, deletion, and incident process.
  7. Periodic access and output testing.

Primary references

This operating model aligns with the governance and lifecycle approach of the NIST AI Risk Management Framework, the management-system requirements described by ISO/IEC 42001, and the sensitive-information disclosure risks documented by the OWASP GenAI Security Project.

Govern the data path before scaling the agent.

Structishen maps the use case, data boundary, permissions, approvals, and operating controls before connecting AI to live work.

Discuss a governed implementation →