Insights
Private AI agents and GDPR — engineering checklist
An engineering checklist for private and self-hosted AI agents under GDPR-minded Swedish and EU constraints — without fake compliance seals.
HumanBee AB
AI engineering & consulting · About HumanBee
Private AI and GDPR are related but not identical. Running an agent on infrastructure you control can reduce some cloud risks. It does not automatically make processing lawful, minimized, or well-documented. This checklist is for engineers and security partners — not legal advice.
1. Map data classes
List what enters the agent loop: user content, customer PII, secrets, embeddings, logs, tool payloads. Classify retention and residency needs before choosing a model host.
2. Document every egress
For each step, write where data goes: model API, vector store, ticket system, analytics, support chat. “We run locally” fails if logs or embeddings still leave the boundary.
3. Separate model choice from orchestration
Self-hosting the orchestrator while calling a public model API is a common hybrid. Be explicit. See Private & Self-Hosted AI and the regulated industry page.
4. Minimize tools’ blast radius
Authorize tools by side effect. Irreversible writes need human-in-the-loop. Default deny beats “the model seemed sure.”
5. Logging and access control
Who can read prompts and tool outputs? How long are they kept? Are admin roles separated from tenant data? Multi-tenant products (e.g. Zynli) make isolation a first-class concern.
6. Vendor and DPA reality
If you use a hosted model or vector service, involve procurement and counsel on DPAs. Engineering cannot invent a compliance seal.
7. Honest customer language
Do not claim “GDPR certified agent” unless you have a real basis. Prefer: topology described, risks documented, controls implemented.
HumanBee AB is a Swedish company. We design private paths with honest tradeoffs — consult if you want a topology review for your estate. Swedish overview: /sv/tjanster/privat-ai.