Why your LLM’s passport matters under DPDP and GDPR

For a regulated institution, the question is no longer only which model you use. It is where the model runs — and which jurisdiction’s law travels with every token your risk data passes through it.

When a risk analyst asks an AI to assess a control or quantify an exposure, that request can carry deeply sensitive material: customer identifiers, loss-event detail, account data, internal control weaknesses. If the model answering sits in a shared cloud in another region, that data has just crossed a border — often without anyone in the second line intending it to. Under modern data-protection regimes, that crossing is a regulated act in itself.

Residency is now a design decision, not a setting

India’s DPDP Act 2023 draws a hard line around personal financial data and where it may be processed. The RBI IT Framework expects inference and storage to stay within India-approved boundaries. GDPR Article 25 demands data protection by design and by default — which is hard to claim when sensitive payloads leave your environment to reach a model. And under SOX, every instance of AI touching financial data needs to land in an audit trail you control.

The common thread: where your model runs determines which regime applies, what you must prove, and how much of the burden falls on configuration after the fact versus architecture from the start. A bolt-on AI feature that calls a shared external endpoint cannot give you that control. It was not built to.

Match the deployment to the mandate

The practical answer is not a single global model. It is the ability to choose, per institution and per jurisdiction, where inference happens — and to have every mode governed identically so security never depends on which one you picked.

Redact before you reason

Choosing where the model runs handles residency. It does not, on its own, handle what the model sees. Even inside an approved region, a well-designed system should strip personally identifiable information — keys, email, national IDs, phone numbers — before any payload reaches inference, and sanitize the model’s output before it returns to the user. In a fully air-gapped deployment, payloads stay local-only and never traverse an external boundary at all.

That combination — pick the jurisdiction, then minimize what crosses any boundary within it — is what lets a CISO sign off on AI in a regulated environment without a carve-out or a leap of faith. Sovereignty stops being a clause in a vendor contract and becomes a property of the architecture.

The institutions moving fastest on AI in risk are not the ones with the loosest rules. They are the ones who made residency a first-class choice, so that adopting AI never meant renegotiating their obligations.