Protecting business data in an AI system means knowing what enters the system, who can read it, where it is processed and when it is deleted. A private or local model is an architectural choice. It does not replace checks on the application, permissions, logs and providers.

Local, private and cloud are different concepts

“Local” describes execution location. “Private” may describe a dedicated environment or access restricted to authorized users. “Cloud” refers to external infrastructure, which may be shared or dedicated. Compare proposals using a data flow description rather than a label.

An Internet-facing service does not become secure because its model runs in an office. Cloud use does not automatically mean every prompt is used for training either: check the specific product, settings and applicable agreements.

Map a document’s complete journey

  1. Input: who uploads the file and for what task?
  2. Preparation: do OCR, conversion and extraction run internally or through another service?
  3. Inference: which model receives the content, on which infrastructure?
  4. Output: who can view the answer and source documents?
  5. Storage: where do temporary files, history, logs and backups remain?
  6. Deletion: what is removed, when, and how is removal checked?

Include embeddings, search indexes and monitoring services in the map. A local archive combined with remote OCR is not an entirely local workflow.

Controls to test before release

  • Unauthorized users must not receive another department’s documents, including through generated answers.
  • Service credentials must stay on the server, outside code delivered to the browser.
  • Logs should support diagnosis without unnecessary full copies of documents or prompts.
  • Administrator access, backups and restoration need accountable owners and verifiable procedures.
  • Failure behaviour must be explicit: an external fallback may change where data goes.

Document retrieval and training are not interchangeable

A system can retrieve relevant passages from an up-to-date document collection and supply them as context to a model. Often called RAG, this workflow needs permissions consistent with the original archive. Fine-tuning instead changes model behaviour through training. It is not an automatic replacement for retrieval or access control.

First determine whether the task is finding information, producing a consistent format or changing model behaviour. These needs call for different evaluations.

A verifiable example on my website

“Ask about my CV” uses a curated server-side profile as context for RunAI. API credentials are not embedded in the page. The component bounds the question and previous exchanges, and instructs the model not to invent facts missing from the profile.

This is a focused integration using public biographical information, not a certification for handling confidential documents. The RunAI case study describes its architecture. For infrastructure choices, start with local AI benefits, limitations and costs.

Compliance depends on the actual processing

An on-premise installation does not establish GDPR compliance by itself. Purpose, legal basis, parties involved, retention and security measures need assessment for the actual processing activity. EDPB Opinion 28/2024 also explains that an AI model’s anonymity requires a concrete assessment; it cannot simply be assumed.

For a project involving personal data, take the flow map and technical evidence to the person responsible for data protection in your organization. They provide a more useful starting point than a generic “100% private” promise.

Reference

EDPB Opinion 28/2024 on AI models and data protection.