A marketing specialist writes campaign copy in ChatGPT. A finance analyst uploads a quarter-end spreadsheet to a free AI tool for commentary. A developer pastes company code into a coding assistant. None of them means any harm; they're all trying to work faster. But none of them asked whether these tools were approved by the organisation.
This is called shadow AI, and today it exists in almost every organisation.
Why does shadow AI happen?
- The tools are easy to reach: a browser and an email address are all it takes.
- The benefit is real and immediate: employees save hours.
- Organisations are slow: policies, approval processes and enterprise licences often lag behind actual use.
- The rules are unclear: employees don't know what's forbidden, often because nobody has told them.
So shadow AI is usually not a discipline problem; it's a sign of an unmet need.
The risks
| Risk | Example |
|---|---|
| Personal data breaches | Customer data entered into a service abroad with unknown processing terms |
| Trade-secret leaks | Strategy documents, prices or source code uploaded to external services |
| Use in model training | Some individual plans may use inputs to improve their models |
| Wrong decisions | Unverified output making its way into reports and customer correspondence |
| Compliance and audit | The organisation can't show which data is processed where |
| Agent risks | Unapproved AI agents given access to email or company systems |
In Türkiye, the data protection authority's November 2025 guide on generative AI and personal data also stresses that personal data processing with these tools falls under Law No. 6698.
Why banning isn't enough
The first reflex is often to ban. But bans without a safe alternative don't eliminate use; they make it invisible. Employees move to personal phones and accounts; the risk doesn't shrink, it just leaves the organisation's field of view.
The effective approach has three legs: approved tools, clear rules and training.
10 essential controls
- A written usage policy: a short, clear document that also says what is allowed.
- Approved, accessible tools: enterprise tools that meet employees' needs.
- Contractual safeguards: a contractual guarantee that inputs aren't used for model training.
- Data classification: a simple table showing which data can be used with which tool.
- Data protection and transfer review: use cases involving personal data reviewed with legal.
- Training: short, hands-on safe-use training for all employees.
- Human review: AI output going outside the organisation checked by a person.
- Visibility: knowing which tools are used, through surveys or network data.
- Incident reporting: defining whom to notify, and how fast, when data is shared by mistake.
- Agent and tool permissions: limiting the system access granted to AI tools, with human approval for critical actions.
Where to start
- Measure: a short anonymous survey shows which tools are used for what.
- Meet the need: offer an approved tool for the most common use cases.
- Write the rules: publish a 2–3 page usage policy.
- Train: explain the policy in a short session that focuses on how to use AI safely rather than on bans.
- Repeat: tools change fast, so review the policy and controls regularly.
See your risk level in 2 minutes
MLTeam's free GenAI Usage Risk Assessment reviews the 10 essential controls in this article and shows your risk level and missing controls, with what needs to be done.
Sources