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.
| Regulation | Region | Deployment that fits |
|---|---|---|
| DPDP Act 2023 | India | Local LLM — personal data never exits the network |
| RBI IT Framework | India | Local LLM / BYOK within approved boundaries |
| GDPR (Art. 25) | EU / UK | BYOK · EU region, or Local for maximum control |
| MAS TRM | Singapore | BYOK in your own project, or Local |
| PCI-DSS | Global | Local LLM — cardholder data isolated, no external calls |
| SOX (IT controls) | USA / Global | BYOK / Local — full audit trail of AI access |
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.
Run AI where your regulator can see it
From cloud to fully air-gapped — every deployment mode governed identically, with PII redaction before inference and a full audit trail behind every AI action.