Shadow AI: why bans fail and what a one-page policy should say

"Shadow AI" is the use of AI tools at work that the organization has not approved or does not know about. It is the successor to shadow IT, and like its predecessor it grows fastest where the approved option is slow, missing or worse than what employees can get for free.

How common it is

A survey of 2,000 employees at UK and US companies with more than 500 staff, run by security vendor BlackFog and reported in January 2026, found that 49% use AI tools their employer has not approved, and 51% have connected AI tools to other work systems without IT approval ^[1]. Among those using unapproved tools, 58% rely on free versions ^[1]. Microsoft's 2024 Work Trend Index, covering about 31,000 workers in 31 markets, found 78% of AI users bringing their own tools to work, rising to 80% in small and medium-sized companies ^[3].

Vendor-run surveys differ in sample and definition, so treat the percentages as indicators rather than precise measures. The direction is consistent.

Why it matters

The risk is mostly what employees put into the tools. In the same survey, about a third said they had shared research or datasets through unsanctioned tools and more than a quarter had shared employee data such as payroll or performance information ^[1]. Senior staff were not more careful: 69% of C-level respondents agreed that speed outweighs privacy or security ^[1].

IBM's 2025 breach research gives the cost side. One in five organizations reported a breach due to shadow AI, and those with high levels of shadow AI had $670,000 higher average breach costs than those with low or none ^[2]. Only 37% had policies to manage AI or detect shadow AI, and among organizations with AI governance policies, only 34% performed regular audits for unsanctioned AI ^[2].

Why bans do not work

In the BlackFog survey 63% of employees said it is acceptable to use AI without IT oversight if no approved option exists, and 60% said unsanctioned tools are worth the risk if they help meet deadlines ^[1]. A ban without an alternative asks people to be slower than their peers. It drives use out of sight, which is the worst outcome for visibility.

A one-page policy

A useful policy fits on a page and answers seven questions.

1. Which tools are approved? Name them, and say which tier (business accounts, not personal logins).

2. What data may go in? Use three classes. Green: public or non-sensitive content, allowed in approved tools. Amber: internal business information, allowed only in approved business-tier tools. Red: customer personal data, health, financial account, credentials and other restricted records, not allowed in any tool unless specifically cleared.

3. Which accounts? Work identities only; no personal accounts for work content.

4. What about connectors and plug-ins? Require approval before an AI tool is connected to email, files or business systems. Agent frameworks deserve extra scrutiny: Microsoft's security team advises, for the OpenClaw agent runtime, treating it as untrusted code execution with persistent credentials, and recommends dedicated identities, least privilege and isolated deployment ^[4].

5. Who checks the output? A person is accountable for anything sent to a customer or used in a decision.

6. How do I ask for a new tool? Publish an owner and a target turnaround, such as a week. A slow exception path recreates the problem.

7. What happens if something goes wrong? Say where to report an accidental paste, and make clear that prompt reporting is not punished.

Making it stick

Policy needs three supports. Provide a good approved option, since a tool people prefer is the strongest control. Add visibility: review what sanctioned tools are in use and what network or identity logs show, and audit for unsanctioned use on a schedule, which only 34% of organizations with AI governance policies did in IBM's data ^[2]. Train briefly and concretely, using examples from your own business rather than general warnings.

For most businesses, the practical implication is to publish the one-page policy and an approved business-tier tool in the same week, then review actual usage after 30 days and adjust.