Confidential data and AI: classify before processing

Short answer

Whether data may be processed with AI depends on content, purpose, legal basis, contract, provider, storage, access and technical settings. A blanket rule such as “no personal data” is insufficient.

What the concept actually means

Data classification translates protection need into action: permitted, approved environment only, anonymise first or prohibit entirely.

Why it matters in the enterprise

Data minimisation remains effective: provide only passages and attributes needed for the task. Test data and abstracted examples reduce risk during development and training.

A controlled method

The CTPM practice framework for controllable AI applications uses seven stages: understand the task, clarify context and data, apply AI deliberately, review professionally, handle deviations, approve accountably and document transfer. It is a transparent working framework, not a certification.

Define task and impact
Clarify data, context and permissions
Review against domain criteria
Control deviations, approval and evidence

CTPM practice example

CTPM practice example: For an HR analysis, names and direct identifiers are removed, sensitive free text is excluded and only aggregated criteria are processed in a contractually approved environment.

Quality and test criteria

The following criteria make quality observable for this use case:

Data class and purpose are recorded before use.
Provider, storage and training settings are reviewed.
Access follows need to know.
Deletion and incident handling are defined.

Risks and common misconceptions

Anonymisation can fail through context. Pseudonymous data remains personal. Copy-and-paste bypasses technical controls, requiring clear work rules and training.

Example transfer artefact

Transfer artefact: a data-approval matrix with class, permitted environment, minimisation, retention, owner and review evidence.