AI agent security and GDPR: what a Finnish business must know
The three questions to ask any AI vendor
Before you let an agent touch a single customer email, ask:
- Where does my data go? (Country, provider, region)
- Is it used to train future models? (Yes = red flag)
- Who can access it inside your organisation? (Answer should be "no one")
If a vendor can't answer these in plain language, walk away.
GDPR-compatible model providers
Not all language models are equal for European data. The ones we use on GDPR-sensitive work:
- Azure OpenAI, EU region: GPT-4o and o-series hosted in Sweden or Netherlands, no training on your data, standard EU DPA
- Mistral: French provider, EU-hosted, strong data agreement
- Anthropic Claude via AWS Bedrock EU: US model, EU inference region, DPA available
- Self-hosted open models (Llama, Mistral): full control, higher cost, sometimes weaker quality
We avoid consumer-grade endpoints (regular ChatGPT accounts, free Claude) for anything with personal data. They may train on prompts.
The guardrails we put on every agent
Every production agent we ship gets these by default:
Least privilege
The agent only sees what it strictly needs. An email agent reading the "sales" inbox has no access to HR or finance folders. A customer service agent can read orders but can't issue refunds without human approval.
Human approval for destructive actions
Any action that can't be undone easily - sending money, deleting records, contacting a customer for the first time - requires a human tap.
Full action log
Every read, decision and write is logged with a timestamp. When something goes wrong, we can reconstruct exactly what happened. Logs are retained per your data policy (usually 90 to 365 days).
Secrets in a vault
API keys, tokens and passwords never live in prompts or code. They sit in encrypted secret storage and are injected only at request time.
Data minimisation
Before content goes to the language model, we strip fields the agent doesn't need. Full names get initialised where possible. Sensitive identifiers (personnummer, bank account) never leave your systems.
The DPA question
You will need a Data Processing Agreement (tietojenkäsittelysopimus) with three parties: us (Bluefour), the model provider (Azure/Mistral/etc), and any integration provider (email, CRM, accounting). We handle the coordination and provide standard EU-compliant templates.
::: callout The one-line audit
An agent is GDPR-safe when a data-protection officer can look at the log, see what personal data touched which system for what purpose, and understand it in one reading.
:::
Sector-specific notes
- Healthcare data: needs a separate risk conversation. We usually don't take these on with off-the-shelf model providers.
- Financial services under EBA rules: agent may need to run in a controlled tenant with additional attestation. Add 4 to 6 weeks to the timeline.
- Public sector: procurement often requires additional certification (like ISO 27001 from providers). Doable, but plan for it.
What to ask us
- Show me the data flow diagram before we sign
- Give me the DPA templates for review
- Explain what happens on failure - who gets notified, what stops
- Show me the log format so my DPO can audit it
We answer all of these in the first proposal, not after signing. If any AI vendor won't, take that as the answer.
For related reading, see [AI agent integrations](/blogi/ai-agentti-integraatiot-pk-yritykselle) and [what is an AI agent](/blogi/mika-on-ai-agentti).
Published · Kaspar Eevald
Bluefour – Helsinki digital agency | hello@bluefour.fi | +358 45 850 4236