AI systems are hungry for data, and much useful data is personal. That puts privacy at the centre of almost every serious AI project. Handled well, it is a design constraint that produces better systems. Handled late, it becomes the reason a project stalls, or worse, the reason it causes harm.
This is general guidance, not legal advice. Your obligations depend on where you operate and whose data you hold, so involve your own legal and compliance teams early.
Know what data you have, and why
Start with an inventory: what personal data the system will touch, where it comes from, where it goes and why each piece is needed. For each item, you should be able to say what it is used for. Most privacy problems begin with data collected “in case it is useful”, with no clear purpose.
The inventory should cover the whole lifecycle: collection, processing, storage, sharing with third parties, and deletion.
Minimise by design
Use the least personal data that does the job. Remove or mask identifiers the model does not need. Keep data only as long as it serves its purpose, then delete it. Minimisation reduces both the likelihood of a problem and the harm if one occurs, because data that was never collected cannot be exposed.
A useful test: for each field, ask what would break if it were removed. If the answer is nothing, remove it.
Consider synthetic data
For many tasks, realistic synthetic data can stand in for real personal data. It can be generated to match the statistical patterns of real data without corresponding to any real individual.
Our Multimodal Synthetic Surveys produce research insight without relying on a live respondent panel, and they track closely with live benchmarks. Our Astra showcase uses synthesized product and catalogue data, so it can be demonstrated without exposing any client’s information.
Be careful with what goes into prompts
Language-model systems introduce a specific risk: personal data included in a prompt is sent to the model, and may be logged along the way. Treat prompts as data flows. Strip identifiers that the model does not need to complete its task, and be deliberate about what context is retrieved and passed in.
Control where data travels
Understand which third-party services, including model providers, will see your data, what they do with it, how long they keep it and where it is stored. Choose deployment options and contractual terms that match your requirements. For sensitive data, that may mean models deployed in your own environment rather than shared services.
Restrict access
Apply role-based access so that people and services can reach only the data they need. Our Dashboard & Billing System builds role-based access into the platform for exactly this reason. Access should be granted by need, reviewed regularly, and removed promptly when roles change.
Keep an audit trail
Log who accessed data, when, and what the system did with it. When someone asks how a decision was made or where their data went, you need to be able to answer precisely. Audit trails also make problems visible early, before they grow.
Respect the rights of the people in the data
Depending on where you operate, people may have rights to see, correct or delete data about them. Design so those requests can actually be honoured, including in anything derived from the data. That is much harder to retrofit than to build in, particularly once data has been used to train a model.
Security is part of privacy
Encryption in transit and at rest, careful management of credentials and secrets, and regular review are not optional. Privacy commitments mean little if the data can be taken. For AI systems, security also includes protecting against inputs designed to make a model reveal information it should not.
Bring privacy in on day one
Every item above is cheaper at the start of a project than at the end. The best time to involve privacy and compliance is before the first line of code, when design decisions are still inexpensive to change.
A short checklist
- Do you know every piece of personal data the system touches, and why?
- Have you removed everything the system does not need?
- Could synthetic data replace any of it?
- Do you know where data travels, including to model providers?
- Is access restricted by role, and logged?
- Can you honour a request to see or delete someone’s data?