2.2.4. Classifying AI Tools to Contain Shadow AI
💡 First Principle: People adopt tools that make their work easier, whether or not those tools are approved. If the sanctioned path is slow or absent, usage moves to personal accounts where the organization cannot see what data is shared. The goal is not zero AI use; it is visible, governed AI use, achieved by making the approved path clear and the unapproved path hard.
Shadow AI is the use of AI tools without the knowledge or approval of the organization's IT, security, or governance functions: an employee pasting a customer contract into a free chatbot, a team subscribing to an AI transcription service on a credit card, a developer using an unvetted code assistant. The risks are concrete: confidential data leaving the organization, intellectual property exposure, regulatory violations (such as personal data processed without a legal basis), unreliable outputs used in decisions, and no audit trail.
The exam guide names a specific control: transparent classification of AI tools.
| Classification | Meaning | What employees should do |
|---|---|---|
| Approved | Vetted for security, privacy, legal terms, and fit; may be used within stated limits (for example, "no customer personal data") | Use freely within the published guidance |
| Under evaluation | Requested and being reviewed; not yet cleared | Wait, or use only with non-sensitive test data if the policy allows |
| Blocked | Rejected because of unacceptable risk (data retention, IP terms, security) | Do not use; use the named approved alternative |
Classification works only as part of a system:
- A published, searchable registry so employees can check a tool's status in seconds.
- A fast, clear intake process for requesting new tools, with a committed review time. Slow reviews recreate shadow AI.
- Sanctioned alternatives that meet the underlying need, such as an enterprise generative AI assistant with data protections, so people do not need personal tools.
- Technical enforcement of blocks where possible (network or endpoint controls), paired with monitoring to discover unapproved use.
- Training and clear policy explaining what data may be used with which tools and why.
A total ban without sanctioned alternatives tends to fail. Employees who see real productivity gains will find a way, often on personal devices, where the organization loses all visibility. Combining preventive controls (blocking high-risk tools) with an easy formal path (request, review, approve) is the pattern the exam favors.
⚠️ Exam Trap: In multiple-response questions about shadow AI, the strongest pair usually combines a preventive control (block or restrict unapproved tools) with a governed path (an organization-wide request and approval process or approved alternatives). Options that rely only on employee goodwill, only on after-the-fact monitoring, or on department-level approvals that bypass central review are typically distractors.
Reflection Question: A consulting firm finds that 40% of staff use a free generative AI tool to draft client deliverables. Why might banning the tool alone make the risk worse, and what three elements would you put in place instead?
