LLM Application Development, security and the DPDP Act: a compliance checklist
Is LLM application development compliant with the DPDP Act?
LLM application development is not compliant by default. India's DPDP Act governs the personal data inside the application: lawful consent, a stated purpose, bounded retention, access controls that survive retrieval, and evidence you can produce on demand. Here is the clause-by-clause checklist.
Not by default. The DPDP Act 2023 governs how you handle the personal data inside an LLM application, not the fact that you built one. An application is compliant when it has lawful consent, a stated purpose, bounded retention, access controls that survive retrieval, and evidence you can put in front of a regulator or an enterprise buyer.
What follows is the checklist we work through on every Indian LLM build, obligation by obligation, with the architectural control that satisfies each one and the artefact that proves it. Security and compliance are treated together here because in practice the same controls answer both the statute and the security questionnaire.
What the DPDP Act actually asks of an LLM application
The Digital Personal Data Protection Act, 2023 applies to digital personal data processed in India, and to processing outside India that relates to offering goods or services to people in India. It casts you as the Data Fiduciary and your user as the Data Principal, and the duties are notice, consent for a stated purpose, accuracy, reasonable security safeguards, erasure when the purpose is served, breach notification to the Data Protection Board, and a grievance mechanism. The Ministry of Electronics and Information Technology, which administers the Act, publishes the text and the subsequent draft rules on its own site, and that text rather than a vendor summary is what your legal team should read. A plain-language version sits in our DPDP Act 2023 glossary entry.
Three features of LLM applications make the duties harder than in ordinary software. Personal data spreads: the same customer detail can sit in a source document, an embedding, a cached response, a prompt log and a trace. Purpose blurs: a support transcript collected to resolve a ticket is not automatically available to fine-tune a model. And boundaries leak: retrieval will happily surface a document the asking user was never entitled to see unless the index enforces permissions.
Cross-border transfer is the one place the DPDP Act is more permissive than the GDPR. The Act allows transfer to any country except those the central government restricts by notification, which is a negative-list approach rather than an adequacy regime. That makes using a model API hosted outside India lawful in most cases, but it does not make it acceptable to every enterprise buyer, and contractual residency commitments frequently bind you more tightly than the statute does.
Two further provisions change the design if they apply to you. The Act allows the government to designate an organisation a Significant Data Fiduciary on the basis of volume and sensitivity of data, which brings additional duties including a data protection officer based in India, independent audits and data protection impact assessments. And processing the personal data of anyone under eighteen requires verifiable parental consent, with tracking and behavioural advertising directed at children prohibited. An education product with an LLM tutor sits squarely inside that second rule, and retro-fitting age assurance into a live product is unpleasant work.
The clause-by-clause checklist
| Obligation | What it means in an LLM application | Evidence to hold |
|---|---|---|
| Notice and consent | Users are told, in clear language, that their content is processed by an AI system and for what purpose | Consent record with timestamp, version of notice shown, purpose string |
| Purpose limitation | Ticket data used for answering is not silently reused for training or evaluation without a fresh basis | Data-flow map naming each store, its purpose and its lawful basis |
| Storage limitation | Prompts, traces, caches and embeddings inherit a retention clock, not just the source database | Retention policy per store plus evidence of automated deletion jobs |
| Security safeguards | Encryption in transit and at rest, scoped credentials, PII redaction before logging, tenant isolation | Architecture diagram, access-control matrix, penetration test report |
| Access and correction rights | A subject request reaches source records, the vector index and the log store | Runbook with target turnaround and a re-index trigger |
| Erasure | Deleting a record removes its chunks and embeddings, not only the row | Deletion test showing the record is no longer retrievable |
| Breach notification | A prompt-log exposure is a personal data breach if it contains personal data | Incident plan naming the Data Protection Board notification owner |
| Grievance redressal | A named contact and a stated response window inside the product | Published contact, ticket trail, resolution times |
The four controls that carry most of the weight
Most of the checklist collapses into a small number of architectural decisions taken early. These are the ones we refuse to leave until the second phase.
- Permission-aware retrieval. Every chunk carries the access-control metadata of its source, and the query filters on the asking user's entitlements before ranking. The approach is set out in permission-aware retrieval.
- Redaction at the boundary. Personal identifiers are stripped or tokenised before anything reaches a log, a trace or a third-party evaluation tool. The model may see them; your observability stack usually does not need to.
- One retention clock per store. Source documents, the vector index, prompt logs, traces and semantic caches each get an explicit lifetime, and deletion is a scheduled job rather than an intention.
- Tenant and environment isolation. Separate indexes and separate credentials per tenant, with production data kept out of development. Shared embeddings across tenants are the quiet cause of most cross-customer leakage.
- A model-routing policy. Which model may see which data class is a written rule, enforced in code, not a preference. Sensitive classes route to a self-hosted or region-pinned endpoint.
- Prompt-injection defences. Treat retrieved content as untrusted input. OWASP maintains a public list of the top risks for LLM applications, and prompt injection and sensitive information disclosure head it.
One practical test tells you whether these controls are real. Ask an engineer to take a single customer identifier and trace it through every store the application writes to, then delete it and prove it can no longer be retrieved by any query, retrieval call or cached response. Teams that can do this in an afternoon are compliant in substance. Teams that need a week of investigation to answer have a data-flow map that no longer matches the system, which is the state most LLM applications drift into within two quarters of launch.
Where data residency actually bites
The statute rarely forces Indian data residency. Contracts, sector regulators and procurement teams often do, and RBI-regulated entities in particular arrive with outsourcing conditions that go well beyond the DPDP Act. If your buyers are banks, NBFCs, insurers or hospitals, design as though residency is required, because renegotiating the architecture during a security review costs more than building it correctly.
That usually means open-weight models served inside your own cloud account or a private VPC, with no data leaving your boundary. We build these as self-hosted agentic systems, starting at $31,500 or ₹20,80,000 plus infrastructure. It is a real premium over an API-based build, and it is worth paying only when a regulator or a signed contract requires it.
What compliance adds to cost and timeline
On an ordinary build, the compliance work is two to three weeks of the programme rather than a separate project: the data-flow map, the retention design, the redaction layer, the deletion test and the security questionnaire pack. LLM application development starts at $21,000 or ₹13,60,000 and runs to $84,000 or ₹56,00,000, with these controls inside scope. Starting figures for every service are on the pricing page, and Indian clients are invoiced in INR with GST.
Self-hosting, a formal penetration test or an audit trail built to a regulator's specification pushes the figure and the calendar up. Budget an extra three to five weeks for the first regulated deployment, most of it spent on evidence rather than code.
When this checklist is the wrong tool
If your LLM application never touches personal data, most of this is ceremony. A code assistant over your own open-source repositories, a pricing explainer over public documents, or a drafting tool that operates on text the user pastes and never stores does not need a consent ledger. Applying the full checklist there slows delivery and teaches your team that compliance is theatre.
The checklist is also the wrong starting point when the real question is whether the use case should exist. Inferring sensitive attributes about customers, or scoring people in ways they cannot contest, can be lawful and still be a bad idea. We have told clients no on exactly that basis. Compliance sets the floor, not the standard.
What this looks like in a regulated build
For an NBFC processing KYC and loan onboarding documents, the constraint was absolute: customer documents could not leave the client's infrastructure. The system runs on open-weight models inside their own environment, with extraction, validation and human review inside one audited boundary, described in the KYC document intelligence case study. The compliance artefacts, a data-flow map, a retention schedule and an access matrix, were produced during the build rather than reconstructed for the first audit.
Our own posture and sub-processor position are published on the trust page, and we expect clients to interrogate a supplier the same way, using something like a security questionnaire for AI vendors.
Related reading
DPDP Act 2023 and AI: what Indian companies must do covers the statute in business terms, self-hosted LLMs for BFSI goes deeper on regulated infrastructure, and what makes an LLM application production-ready sets the wider launch bar.
Compliance in an LLM application is an architecture decision taken in week one, not a document written in week twelve.
Frequently asked questions
Does the DPDP Act allow Indian data to be sent to an overseas LLM API?
▾
Generally yes. The DPDP Act 2023 permits transfer to any country other than those the central government restricts by notification, which is a negative-list approach. Sector rules and customer contracts are often stricter, so regulated buyers in banking, insurance and healthcare frequently require Indian residency regardless of the statute.
What personal data does an LLM application usually hold that teams forget?
▾
Prompt logs, model traces, semantic caches, evaluation datasets and the vector index itself. Each can contain personal data copied from source records, and each needs its own retention clock and deletion path. Deleting the source row while leaving its embeddings retrievable is the most common failure we find.
Do we need a self-hosted model to be DPDP compliant?
▾
No. Compliance depends on consent, purpose, retention, security safeguards and evidence, not on where the model runs. Self-hosting inside your own cloud account becomes necessary when a regulator, an outsourcing guideline or a signed enterprise contract requires the data never to leave your boundary.