Start with one real data flow
Consider an internal document assistant: a colleague asks a question, the application checks their access, retrieves relevant document passages, sends the permitted context to a model and returns an answer with sources. Separately, it may write logs or backups. Map each of these steps before choosing a provider.
For every step, record the purpose, necessary data, recipients, location, retention and access rules. This puts the GDPR principles of purpose limitation, minimisation and storage limitation into practical design decisions. A cookie banner does not resolve these decisions.
Cloud API, EU region or local AI?
A label such as “GDPR-compliant LLM API” is not a conclusion about your application. The service terms, configuration and actual use all matter. Compare these deployment options against the same task:
Managed cloud API
Check the contracting entity, processing terms, subprocessors, model-training use, retention and support access. Personal data is not automatically prohibited in every cloud API; its use needs an appropriate legal basis and safeguards for the specific processing.
A service with an EU region
Clarify what the regional setting covers: model processing, stored inputs, monitoring, backups and support access may have different arrangements. For any transfer outside the EEA, assess the relevant transfer mechanism; an adequacy decision and contractual safeguards are different routes. EDPB guidance on international transfers explains the conditions.
Local or privately hosted AI
Run the model on your infrastructure or on private infrastructure operated by MyForge. This can give more control over the processing environment. Check the whole chain: embeddings, document extraction, telemetry, updates and backups can still call external services. Hardware capacity, access control, patching and model licensing remain part of the design. Explore the two local AI deployment options →
Hybrid processing
Keep selected work local and send only approved data to a cloud service. Removing names alone does not prove anonymity: context can still identify someone. Treat redaction as a control to evaluate, including failure cases; the EDPB’s AI opinion also calls for a case-by-case assessment of model anonymity.
RAG needs real access controls
Retrieval-augmented generation (RAG) supplies relevant source material to a model. It does not by itself prevent prompt injection or disclosure between users. OWASP describes these risks for RAG and fine-tuned systems alike.
For the document-assistant pilot, enforce permissions before retrieval, separate each organisation’s data and caches, restrict connected tools, and test hostile instructions inside uploaded files. Let the user inspect source passages. A model’s instruction to “keep data private” cannot replace application-level authorization.
Decide what is stored and how it is removed
Set retention separately for prompts, documents, search indexes, operational logs and backups. Keep only what the stated purpose requires; there is no universal “90-day chat / one-year AI log” rule for every application. A security log often needs event metadata rather than the complete customer message.
Provide an accessible route for rights requests and implement the applicable deletion process across the actual stores. A delete button alone is not the whole process, and erasure is subject to legal conditions and exceptions. See the EDPB guide to individuals’ rights.
What a useful first pilot delivers
- A scoped task and a data-flow diagram, including external services.
- An access matrix and documented operator responsibilities.
- A small representative test set: answer quality, sources, denied access and malicious document content.
- Measured response time and running cost on the proposed deployment.
- A retention/deletion plan and a list of decisions needed before production.
Work out controller and processor roles from the actual responsibilities, and document processing terms where required. The EDPB’s guide to these roles is a useful starting point. A pilot provides technical evidence for the decision; legal assessment and any formal compliance review need their own agreed scope.
Bring one process, not your entire company archive
Tell us the task, who will use it, the types of documents involved and whether the model must run on your premises. Begin with synthetic or appropriately approved sample data. We can then define the pilot, the deployment options and the acceptance criteria together.
Related services: AI integration for business workflows · Local AI on your infrastructure or operated by MyForge
