Skip to main content
The state of ai impact assessment
Written AI Policy Won't Stop Data Leaks. Pick Your ControlsPrivacy and Security
5 min readFor Enterprise IT Leaders

Written AI Policy Won't Stop Data Leaks. Pick Your Controls

Research from BlackFog found that 47% of security leaders have experienced sensitive data exposure through an AI tool. In 77% of those incidents, the tool was company-approved. Meanwhile, 98% of organizations say they have an AI usage policy. The policy exists. The exposure happened anyway.

This guide helps you decide how to close that gap. It's built as a decision tree because the right control path depends on what you can already see and stop.

The Decision You Are Facing

You have a written AI usage policy. The question is whether you can enforce it, and at what layer.

The same research shows how uneven enforcement is. Just over half of organizations (56%) back their policy with comprehensive technical controls. About a third (35%) have a written policy on what employees must not enter into prompts, with no technical enforcement at all. And 39% of security leaders say AI adoption is moving faster than they can govern it. Of those, only 27% have comprehensive controls.

A policy standard without a mechanism is an aspiration, not a compliance control. Auditors and regulators will treat it accordingly. Your choice is which enforcement path to build first.

Key Factors That Affect Your Choice

Answer these four questions before you pick a path.

1. Can you see what data is leaving? In the BlackFog findings, 62% of respondents could identify which users or departments sent data to AI tools. Fewer than half (48%) could produce a record of the types of data that left. Knowing who is not the same as knowing what. Regulatory obligations around sensitive personal data attach to the data, not the user.

2. Can you block or restrict unsanctioned tools? Only 54% of security leaders reported this ability. About a third (32%) can detect shadow AI but can't consistently stop it. Detection without prevention gives you an incident log, not a control.

3. Are autonomous agents in scope? 56% of respondents said autonomous AI agents acting on behalf of employees fall outside their existing controls. That figure reached 70% in technology and telecoms. If agents can read, write, and transmit data under an employee's credentials, your user-centric controls don't see them.

4. What is your cyber risk appetite for AI-related data exposure? If your cybersecurity risk register has no entry for AI data leakage, or the entry has no linked key control indicator, start there. The paths below assume the risk is documented and prioritized.

Path A: Visibility First

Choose this path if: you can't answer "what data types left the organization through AI tools last quarter?"

This is your situation if you sit in the 52% that can't produce a record of data types, or if you can identify users but not content.

What to do:

  • Deploy data-level monitoring at the egress point, covering prompts, file uploads, and API calls to AI services. Network-layer and endpoint-layer tooling both apply; the requirement is content awareness, not just destination awareness.
  • Map detected data types to your policy classification scheme so findings translate directly into your existing handling rules.
  • Feed results into your risk dashboard as a key control indicator, such as the volume of classified data observed in AI-bound traffic.

Requirements driving this path: Any regulatory obligation that requires you to know where sensitive personal data goes (data mapping, records of processing, breach assessment) presupposes this visibility. If you can't scope an information security incident, you can't meet notification timelines. For public companies, scoping also feeds the materiality assessment behind any Form 10-K or Form 10-Q cybersecurity disclosure.

Limit: Visibility alone doesn't reduce exposure. Treat it as a first phase with a defined exit date.

Path B: Enforcement First

Choose this path if: you already know where the data goes, but your policy relies on employee compliance.

This fits the 35% with written prompt rules and no technical enforcement, and the 32% who can detect shadow AI but not stop it.

What to do:

  • Convert each policy standard into a technical rule. "Do not enter customer data into AI prompts" becomes a data loss prevention rule keyed to your classification labels.
  • Maintain an allowlist of sanctioned AI tools and block the rest at the proxy or secure web gateway.
  • Because 77% of exposure incidents involved approved tools, don't stop at tool approval. Apply content controls inside sanctioned tools too: tenant-level configuration, retention settings, and prompt filtering.
  • Document any approved exception with an owner, a compensating control, and an expiry date.

Requirements driving this path: Control design must show the control addresses the stated risk, and automated control testing should confirm it operates. An auditor performing audit fieldwork will ask for evidence the rule fires, not a copy of the policy.

Limit: Content filtering produces false positives. Plan a tuning period and a documented review process so the control doesn't get quietly disabled.

Path C: Agent Governance

Choose this path if: autonomous AI agents are already running, or are on your 12-month roadmap, and your controls assume a human is at the keyboard.

Choose this path in parallel with A or B if you're in the 56% whose agents fall outside existing controls.

What to do:

  • Build an entity-level controls requirement: no agent deploys without a registered owner, defined scope, and documented data access.
  • Issue agents distinct, least-privilege identities rather than letting them inherit a user's full access.
  • Log agent actions at the data layer so Path A monitoring covers them.
  • Add a human approval gate for agent actions that move sensitive personal data outside your environment.
  • Add agent-specific scenarios to your incident response structure, including how you'd assign severity level when an agent, not a person, caused the exposure.

Requirements driving this path: Accountability expectations in your enterprise risk oversight don't disappear because the actor is software. Someone must own the agent's risk.

Limit: Standards for agent governance are still maturing. Design controls around principles (identity, scope, logging, approval) so you can adapt as guidance firms up.

Summary Matrix

Your situation Start with Primary evidence an auditor will want Main risk if skipped
Can't produce data-type records Path A Content-level logs mapped to policy classification Can't scope an information security incident
Policy exists, no technical enforcement Path B Working rules, test results, approved exception records Policy standard is unenforceable on its face
Detect shadow AI, can't block it Path B Block/allowlist configuration and testing Detection with no prevention
Agents operate outside controls Path C (plus A or B) Agent registry, scoped identities, action logs Ungoverned data movement at machine speed
AI adoption outpacing governance A and B in sequence, C as needed Risk register entry with linked key control indicators Risk grows faster than control coverage

If you can only fund one path this quarter, fund visibility, then enforcement. You can't write a defensible control design for a data flow you can't see. But don't let Path A become permanent. The research points to the same lesson: a policy tells employees what you expect, and only a technical control tells you whether it happened.

Promotional banner for the Pentest Readiness checklist download

You Might Also Like