To build insights or machine learning models, you should not require unlimited access to fine user data, the statistical results suffice. Align privileges to the minimum amount of information that is necessary.
Do not let data practitioner copy versions of sensitive data. All data remains in the secure infrastructure, and the processing is performed within it.
Every data processing job is logged and can be used for auditing. Unlike when analysts work on their own devices, everything leaves a trace for maximum security.
Differential Privacy guarantees that outputs will not lead to re-identification irrespective of auxiliary information or computing power a third party may have access to.