AI & Data Privacy: Which Risks Really Count and Which Are Overrated

AI data privacy tends to be either dramatized or ignored. Both are expensive. This article assesses the risks soberly, by likelihood and by the concrete case in which AI is used. Because whether a risk is real rarely depends on the model and almost always on the setting it runs in.

88%

of companies use AI in at least one business function

McKinsey Global AI Survey, 2025

21%

of German companies have an AI strategy

Bitkom, 2026

5 years

possible retention of consumer chats when training is enabled

Anthropic, 2025

Contents

• Why AI data privacy is a question of setting
• The real risks
• How likely is each risk?
• What is overstated
• What companies must watch out for
• Frequently asked questions
• Sources

Why AI data privacy is a question of setting

88% of companies use AI in at least one function (Source: McKinsey Global AI Survey, 2025), but only 21% have an AI strategy (Source: Bitkom, 2026). That gap is the actual privacy risk: usage is high, governance is low.

The setting is decisive. The same prompt is harmless in a business account with a data processing agreement and risky in a free personal account that uses inputs for training. The risk lies in how AI is used, not in using it.

The real risks

Four risks are real and relevant:

• Inputs used for training: In consumer accounts, inputs feed model training by default unless you opt out. Confidential content can be retained long-term.
• Data transfer to third countries: With US providers, data can be processed outside the EU. The US Cloud Act can, under certain conditions, grant US authorities access, even to data stored in the EU.
• Loss of confidentiality: Trade secrets, HR data, or customer data in an unsecured tool is both a GDPR and a competitive risk.
• Shadow AI: Employees secretly use private tools because there is no sanctioned alternative. Then nobody controls which data goes where.

The biggest AI data privacy risk is not the model. It is the ungoverned usage next to it.

How likely is each risk?

Risk is likelihood times impact. By that measure the field sorts clearly:

• Highly likely: Shadow AI and training use in free accounts. Both happen without active intent, daily, in almost every organization.
• Medium: Data transfer to third countries. Real, but manageable through provider choice, EU hosting, and contracts.
• Occasionally high: Loss of confidentiality for especially sensitive data (health, HR, trade secrets). Rare in frequency, but severe in impact.

The lesson: spend energy first on the highly likely everyday risks, not on rare extreme scenarios.

What is overstated

Two worries are usually smaller than assumed. First, that a business or enterprise account “secretly” uses inputs for training. The major providers do not train on business, enterprise, or API data by default. The difference from a consumer account is substantial. Which provider regulates what exactly is shown in the provider comparison.

Second, that AI is inherently “not GDPR-compliant”. Compliance depends on the use, not the tool: on legal basis, contract, data minimization, and settings. With the right setting, compliant operation is possible.

What companies must watch out for

Three things belong on the radar before AI goes wide:

1. A clear setting: Sanctioned tools, the right contract tier (business/enterprise with a data processing agreement), a defined data region.
2. Clear input rules: What may go in and what may not, understandable for everyone, not just IT.
3. Governance over bans: A sanctioned, good solution displaces shadow AI more reliably than any ban.

At netzstrategen this is part of AI Operations: the Strategy Layer keeps context EU-hosted and anonymizes it before the AI, the Admin Layer filters what reaches the model at all. What the setup looks like in practice is shown in Configure AI for data privacy. The legal guardrails come from the EU AI Act.

Frequently asked questions

Is it dangerous to use AI in a company?

Not inherently. What is dangerous is ungoverned use: free accounts with training enabled and shadow AI. In the right setting (business/enterprise account, clear rules, EU hosting) the real risks are well manageable.

Are my inputs used to train the AI?

In a consumer account, often yes, unless you opt out. In the business, enterprise, and API tiers of the major providers, not by default. That difference is the most important lever in AI data privacy.

What is the most underestimated risk?

Shadow AI. Where there is no sanctioned, good solution, employees fall back on private tools, and nobody controls which data goes where. Where the biggest risk sits is shown fastest in a free introductory call.

Sources

• McKinsey: The State of AI, Global AI Survey, 2025
• Bitkom: Artificial Intelligence in Germany, 2026
• Anthropic: Updates to Consumer Terms and Privacy Policy, 2025
• European Commission: EU AI Act