Menu
Nederlands

Privacy-safe AI for personal and confidential data

Many of the most useful AI applications touch personal or confidential data. I design systems where the model does the language work, while you stay in control of who sees what.

Tell me about your project

The problem

Sending customer files, applications or internal documents to a public AI service is a risk: the data leaves your environment, and you often can't say where it ends up. The Dutch Data Protection Authority has received data breach notifications caused by exactly this.

A contract with an AI provider helps, but it doesn't remove your GDPR obligations. You still need a legal basis, and you still have to limit the data to what's necessary. The safest design is one where the model never needs the sensitive part.

How I approach it

  1. Map the data flows

    Which data goes where, who processes it and why. This is also the basis for the conversation with your privacy officer.

  2. Pseudonymise before the model

    Names, addresses, dates of birth and account numbers are replaced with stable tokens. The link to the real values stays in a vault in your own environment.

  3. Choose where the model runs

    An EU provider under a processing agreement, or open models on your own servers when the data must not leave your environment at all.

  4. Leak checks and logging

    Every outgoing request is checked for original values, and every re-identification is logged.

  5. Keep decisions explainable

    Language models gather the facts. Fixed rules that an auditor can read make the decisions.

What to watch out for

  • Pseudonymised data is still personal data under the GDPR. It lowers the risk, but it doesn't remove your obligations.
  • Free text can identify someone without a name. Detection has to be tested on real samples.
  • Self-hosting gives you control, but it needs GPUs and maintenance. It isn't automatically cheaper.
  • Employees pasting data into chatbots on their own is often the bigger risk. Clear agreements matter as much as the technology.

Projects like this

  • Car-lease company
  • 2026

First assessment of lease applications

Someone used to read and check every application by hand. Now the system reads the dossier, checks the credit and business registers, and prepares a first assessment for the reviewer. The language models never see data that could identify a customer.

  • Personal data is replaced with pseudonyms before any model sees it, and every re-identification is logged.
  • The decision comes from a fixed set of rules, kept separate from the language model, so an auditor can inspect them.
  • Reviewers adjust the policy rules themselves, with a full history of every change.
  • The reviewers save an estimated 60 to 80% of the work on each application.
  1. SharePoint dossierCredit and business registers
  2. Pseudonymise
  3. Extract facts
  4. Apply policy rules
  5. First assessment for the reviewer
  • Financial services company
  • 2025

A private AI workspace with many agents

The team wanted AI assistants that don't send data to a public chat app. Now they have a self-hosted chat environment with specialised agents, plus a separate API layer that logs usage and caches answers.

  • A self-hosted chat environment with agents for different tasks.
  • An API layer between the agents and the model providers that logs usage and caches answers.
  • Everything runs on infrastructure the client already had.
  1. Questions and documents from the team
  2. Specialised agents
  3. Layer for logs and cache
  4. Answers inside the client's own environment
  • Software company
  • 2026

Local language models, only when the team needs them

The team wanted to use language models without sending data to an outside provider. But a model on a local GPU holds on to a large chunk of GPU memory, even when nobody is using it. Now the team has one API for all local models. An agent on each GPU machine starts a model on the first request and stops it after an hour without requests, which frees the GPU for other work.

  • One OpenAI-compatible API for all local models, which team members connect their chat and coding tools to.
  • A dashboard shows which models are running, and admins can start or stop any of them.
  • The models run on GPU machines the company already owns. No data goes to an outside provider.
  1. Request from a chat or coding tool
  2. Shared API
  3. Start the model if needed
  4. Answer from the local model

Frequently asked questions

Is this legal advice?

No. I design and build the technical side and work alongside your privacy officer or lawyer, who assesses the legal side.

Can we use ChatGPT or Claude at all?

Often, yes: with a business agreement, the right settings and pseudonymisation of sensitive fields. For some data a self-hosted model is the better choice. That depends on the data, not on a general rule.

What about the EU AI Act?

Most document and data processing systems aren't high-risk under the AI Act. Some uses are, such as assessing the creditworthiness of individuals or screening job applicants, and those come with extra requirements. We look at this at the start of a project.