Shadow AI

In plain English

Shadow AI is the AI version of shadow IT. Shadow IT meant staff installing unapproved software or using a personal cloud drive for work files. Shadow AI means signing up for a free chatbot, meeting transcriber or image generator and feeding it work data.

The motive is usually innocent: people want to save time. The problem is that the organization cannot protect, monitor or account for a tool it does not know exists. A related blind spot is an AI feature that a vendor quietly switches on inside software the organization already approved.

Example

Dunmore Legal, a fictional 300-person law firm, reviews its web traffic and finds 120 staff using five browser-based AI assistants. Some paste draft contracts and client emails into them. The firm has no contract with any of these providers, so it cannot say whether they keep prompts or train on them.

The firm responds in four steps:

1. It records all five tools in its AI inventory and assesses the two most popular.
2. It approves one enterprise assistant under a contract that bars training on firm data, and blocks the rest.
3. It updates its acceptable use policy: approved tools and prohibited inputs (client confidential and personal data). The policy also requires human review before anything reaches a client, and adds an intake form for new tools.
4. It runs short training that explains why the rules exist.

Unapproved use drops sharply, because staff now have a sanctioned tool that is just as quick.

Why it matters

Shadow AI turns ordinary governance duties into blind spots:

• Privacy. Pasting personal data into an unvetted tool can mean disclosing it to a provider with no processor contract. The GDPR requires controllers to use processors with sufficient guarantees under a binding contract (Art. 28) and to keep data secure (Art. 32).
• Confidentiality and IP. Client secrets, trade secrets and source code can leave the organization and may end up in a provider's training data.
• Accuracy. Unreviewed outputs, including hallucinations, can reach clients and regulators.
• Visibility. The NIST AI RMF expects mechanisms to inventory AI systems (GOVERN 1.6) and policies for third-party software and data risks (GOVERN 6). You cannot assess, monitor or investigate a tool you have never recorded.
• Literacy. Under EU AI Act Art. 4, as rewritten in 2026, providers and deployers must take measures to support their staff's AI literacy. Acceptable use training is one common measure. See AI literacy.

How the AIGP exam tests it

Shadow AI shows up in Domain I questions on policies and third-party risk (lesson I.C.5). A typical scenario: hundreds of employees are pasting confidential material into unvetted chatbots. Which instrument governs this most directly?

The trap is the vendor contract. Contracts bind vendors, and there is no contract with a tool nobody procured. The instrument that binds employees is the acceptable use policy, enforced through HR. A second trap is the outright ban with no alternative, which pushes use underground.

The tell: the best answer pairs a clear acceptable use policy with an approved alternative and a fast route to request new tools. If the question is about visibility ("which customers' data was touched?"), the answer is usually the AI inventory (lesson I.C.4).