AI, explained

How do I keep my data private with AI?

Privacy with AI is a configuration and architecture problem, not a matter of trust, and the controls are more concrete than most vendors explain.

The concern behind this question is usually specific: will our customer records, contracts, or internal documents end up training someone else's model, or surface in another company's output. The answer depends entirely on which product tier you use and how you build, and the difference between the careless and the careful path is large.

Start with the deployment tier

Consumer chat products and business or API tiers are governed by different terms. Business and enterprise tiers from the major providers do not train on your inputs by default, offer configurable retention, and provide data processing agreements. Free consumer tiers frequently do use conversations for improvement unless you opt out. The single highest-impact privacy decision most companies make is simply which door they walk through, and staff pasting client data into personal accounts is the most common real-world leak.

Then reduce what you send at all

• Redact before transmission. Strip names, national identifiers, card numbers, and health details when the task does not require them. Most tasks do not.
• Scope retrieval by permission. If an assistant can search all company documents, it will eventually show someone a document they should not see. Filter retrieval by the requesting user's access rights, not after the fact.
• Keep secrets out of prompts. API keys and credentials belong in a secret store, never in a system prompt or a tool description.
• Constrain tools. An agent should hold the narrowest permissions that let it do its job, and destructive operations should be unavailable to it entirely.

Retention and logging

Logs are where privacy programmes quietly fail. Teams configure zero retention at the provider and then log every full prompt and response to their own observability stack indefinitely. Decide retention deliberately at both layers, redact logs, and set an expiry. You need enough history to debug and evaluate, not a permanent archive of customer conversations.

The GDPR layer

Under GDPR you need a lawful basis for the processing, a record of it, a data processing agreement with the provider, clarity on where processing occurs and what transfer mechanism applies, and honest disclosure in your privacy notice. If the AI materially affects decisions about people, assess that too. None of this is exotic, but it needs to be written down before launch rather than reconstructed afterwards.

When to run models yourself

Self-hosting an open-weights model removes the provider from the equation and adds real cost and operational burden, usually with a capability trade-off. It is justified for genuinely restricted data such as clinical records or classified material. For most businesses, a business-tier API with a signed DPA, redaction, and scoped access is both safer in practice and far cheaper than a self-hosted deployment nobody has time to patch. An AI audit should tell you which side of that line your data sits on.

Frequently asked questions

Related