GDPR, the AI Act, sector rules, a customer contract, each with its own question. Strip the data and the model learns nothing. Copy it and a breach would be of everything.
GDPR, the AI Act, sector regulation, a works council, a customer contract. Each asks a different question, and the project has to answer all of them before it starts.
Strip the data until it is safe and the model learns nothing. Keep it rich and legal says no. Without a method, the team picks one and loses.
Data copied into a training environment with wide access is a second copy of the crown jewels. If it leaks, so does the whole history.
We assess the risk with your data protection officer, choose the technical safeguards that keep the data useful, and write the governance so each use is allowed, logged and reviewable. The project passes the review because it was designed for it.
Three parts: know the risk, put the safeguards in, and keep them in force.
What data, what use, what could go wrong.
Built into the pipeline and the model.
The rules, the owners, the reviews.
Four phases, with your DPO at each gate.
Data, rules, threats, risk ratings.
The safeguards, designed.
Built into the pipelines and systems.
Proof, then routine.
The project passes review, the data stays useful, and a leak would not be a catastrophe.
The DPIA, the safeguards and the logs come out of the design. The DPO reviews a working system.
Safeguards chosen for the task keep the signal in the data. Accuracy is measured under the safeguard and reported next to the guarantee.
Minimised data, pseudonyms, least access and logs. What an attacker could take is less, and what they took would be traceable.
| Side by side | Privacy at the end | Privacy by design |
|---|---|---|
| Data handling | Full copies in a training environment | Minimised, pseudonymised, in place where possible |
| Controls | Access lists, hopefully | Least access, encryption, logs, enforced |
| Regulatory risk | Discovered at review | Rated in assessment, mapped to each rule |
| If there is a breach | The whole history | Minimised data, traceable |
| What you can build | Whatever legal did not stop | The use cases the safeguards allow, on record |
Some cost accuracy, some cost nothing. Minimisation and pseudonymisation rarely change what a model learns. Differential privacy trades accuracy for a guarantee, and the amount is measured: the model under the safeguard against an unconstrained baseline, reported with the privacy budget so the DPO and the business owner decide together.
GDPR is the baseline and the AI Act adds obligations for some uses. Sector rules, health, finance, public sector, add their own. The assessment maps each use case to the rules that apply, and each safeguard to the article or clause it satisfies, so the DPIA is written from the design.
Any personal or confidential data an AI project touches: customer records, patient data, employee data, transactions, documents with names in them, logs. The technique differs by type: structured records suit minimisation and differential privacy, documents need redaction and access control, partner data suits federated or secure computation.
The safeguards are enforced by the pipelines and systems themselves. Access and model runs are logged, retention and deletion run automatically, and a review cadence with a named owner checks the logs, the budget and the rules against new use cases. An incident path says who does what if something goes wrong.
We are a small team of senior specialists. We pick the right model and the right layer, and we build the least machinery that does the job. You get a call with an engineer, not a sales deck.
