AI & personal dataUpdated 22 November 202610 min read

DPDP Act for AI Systems: Implementation Guide for Indian Companies

How the DPDP Act and Rules apply to LLM apps, RAG, agents and ML pipelines in India, with the architecture controls to build before 13 May 2027.

By Suhas Bhairav · Reviewed against the DPDP Rules, 2025 (G.S.R. 846(E), 13 Nov 2025)

AI-specific exemption
None
Lawful basis
Consent (s.6) or legitimate use (s.7)
Security baseline
Rule 6(1)(a)–(g)
Minimum log retention
1 year (Rule 8(3))
Breach reporting
Board detail within 72 hours (Rule 7); CERT-In 6 hours
Max penalty
Up to ₹250 crore (security safeguards)
Obligations from
13 May 2027

Quick answer

The DPDP Act applies to AI systems whenever they process digital personal data: prompts, training sets, embeddings, call transcripts and model outputs about identifiable people. There is no AI-specific exemption. From 13 May 2027, AI teams need a lawful basis per purpose, Rule 6 safeguards such as redaction and access logs, Rule 8(3) one-year log retention, erasure propagation, and processor contracts with model vendors.

  • There is no AI exemption: prompts, embeddings, transcripts and outputs about identifiable people are personal data processing.
  • Map every AI data flow — collection, prompt, retrieval, model call, output, log — before choosing controls.
  • Redact or tokenise personal data before it reaches a third-party model; Rule 6(1)(a) names masking and virtual tokens.
  • Rule 8(3) requires at least one year's retention of processing logs, so design retention tiers deliberately.
  • Model and vector-database vendors are your Data Processors; section 8(1) keeps liability with you.
  • Erasure must propagate to indexes, caches, fine-tuning sets and evaluation data.

Who this is for: CTOs and AI platform teams, ML and data engineers, Product managers shipping GenAI features, CISOs and Data Protection Officers, SaaS and AI startup founders

Does the DPDP Act apply to AI systems?

Yes. The DPDP Act regulates the processing of digital personal data, and its definition of processing covers wholly or partly automated operations — collection, storage, retrieval, use, combination, indexing, sharing and erasure. A large language model call that includes a customer's name and complaint is processing. So is embedding a support ticket into a vector database, transcribing a sales call, or scoring a job applicant.

India has chosen not to pass a standalone AI law for now. MeitY's India AI Governance Guidelines (November 2025) rely on existing laws — the IT Act, the DPDP Act, consumer protection and sectoral rules — and favour voluntary frameworks and techno-legal approaches. In practice, the DPDP Act is the main binding law for AI systems that touch personal data.

What the Act does not do:

  • It has no separate AI or automated decision-making article like GDPR Article 22.
  • It has no "sensitive data" class, so biometrics, health and financial data are protected as ordinary personal data (with sensitivity mattering for SDF designation and penalties).
  • It does not apply to personal data the Data Principal herself made public (s.3(c)(ii)) — relevant to web-scale training, but narrower than it looks.

From 13 May 2027, the fiduciary obligations and Rules 3 and 5–16 apply to every AI system processing personal data in India, or outside India in connection with offering goods or services to people in India.

Step one: map the AI data flow

Controls only work if you know where personal data moves. For each AI feature, draw the flow and record each hop in your data inventory:

StageTypical personal dataDPDP questions
CollectionChat input, uploaded files, call audio, CRM recordsLawful basis? Rule 3 notice? Child user?
Pre-processingRedaction, tokenisation, chunkingWhat is masked before leaving your boundary?
RetrievalEmbeddings, documents fetched by RAGDoes the user have access to every retrieved chunk?
Model callPrompt + context to a model APIWhich processor, which region, training use?
Tool / agent actionsAPI calls, emails, database writesIs the action within purpose? Logged?
OutputGenerated text, scores, summariesAccurate enough for a decision (s.8(3))?
Logging and evaluationPrompts, outputs, traces, feedbackRetention tier? Who can read raw traces?
Training / fine-tuningCurated conversations, labelled dataSeparate consent or basis? Erasure path?

Capture for each hop: system, data elements, purpose ID, lawful basis, processor, region, retention, and access roles. This table becomes the backbone of your notice, DPIA, processor register and erasure design.

Every AI purpose needs its own lawful ground. The most common mistake is assuming that data collected for one purpose can feed a model for another.

  • Serving the user (answering her support query with an LLM): usually covered by the consent or section 7(a) basis for the service itself, if the notice describes it.
  • Improving your models with her data (fine-tuning, evaluation sets): a different purpose. Under section 6, consent must be specific and limited to necessary data; it cannot be a condition of using the core service if the data is not necessary for that service. Ask separately.
  • Employee-facing copilots: often section 7(i) for employment purposes, but proportionate scope matters. See copilots and employee data.
  • Children: verifiable parental consent under Rule 10, and no behavioural monitoring or targeted advertising (s.9(3)).

Implementing consent scope in code

  • Store consents as records with purpose_id, notice version, timestamp and status.
  • Attach purpose_id to every request entering the AI gateway; reject calls whose purpose lacks a live basis.
  • Make training-data export jobs filter on a training_consent = true flag at row level.
  • On withdrawal, stop new processing within a reasonable time and notify processors (s.6(6)), and queue removal from future training sets.

Rule 6 safeguards mapped to AI architecture

Rule 6(1) lists minimum safeguards. Here is how each maps to an AI stack:

  • (a) Encryption, obfuscation, masking, virtual tokens. Run a PII redaction step before prompts leave your boundary: detect names, phone numbers, emails, Aadhaar-like and PAN-like patterns, account numbers and addresses, replace them with reversible tokens held in your vault, and re-insert values only in the final response if needed. Encrypt vector stores and trace stores at rest. See pseudonymisation in AI pipelines.
  • (b) Access control. Scope model API keys per service, enforce document-level permissions at retrieval, and give agents least-privilege, short-lived credentials.
  • (c) Logs, monitoring and review. Log who called which model with which data class, which documents were retrieved, and which tools an agent invoked. Alert on anomalies such as bulk retrieval.
  • (d) Continued processing. Back up configuration, prompts, vector indexes and consent records; plan fallback if a model vendor is unavailable.
  • (e) One-year retention of logs and personal data for detecting and investigating unauthorised access.
  • (f) Processor contracts with model, embedding, transcription and observability vendors.
  • (g) Technical and organisational measures — red-teaming for prompt injection and data exfiltration, an AI usage policy, and staff training.

A prompt-injection attack that leaks another customer's data is a personal data breach: notify the Board and affected people under Rule 7, and report to CERT-In within 6 hours if it is a specified incident.

Prompt and output retention vs the one-year minimum

AI teams often want short retention for privacy, while Rule 8(3) requires every Data Fiduciary to retain personal data, associated traffic data and logs of processing for a minimum of one year from processing, for the Seventh Schedule purposes, then erase them unless another law requires longer. CERT-In separately requires ICT system logs for a rolling 180 days within India.

A workable design uses tiers:

TierContentsRetentionAccess
Hot tracesFull prompts and outputs for debuggingDays to weeksSmall on-call group, redacted by default
Processing logRequest ID, user/pseudonym ID, purpose ID, model, data classes, documents retrieved, tool calls, timestampsAt least 1 yearSecurity and compliance
Restricted archivePersonal data processed, where needed to meet Rule 8(3)At least 1 year, then eraseBreak-glass only, encrypted, access-logged
AnalyticsAggregated, de-identified metricsAs neededProduct teams

Redacting before the model call reduces how much personal data ever enters the archive. The interaction between Rule 8(3) and erasure requests is not fully settled; the e-book illustration says data and logs must be kept for a year even if the user deletes her account, so remove data from live systems promptly while holding the restricted copy only for the minimum period. See LLM logging and observability.

Erasure propagation into embeddings and fine-tunes

Section 8(7) requires erasure when consent is withdrawn or the purpose is no longer served, and the fiduciary must cause its processors to erase too. Section 12 gives Data Principals a right to erasure. In AI systems, personal data spreads to places traditional deletion scripts miss:

  1. Source stores — CRM, ticketing, file storage.
  2. Chunk and embedding stores — every vector derived from the source. Key vectors by source_doc_id and data_principal_id so a delete event can remove them all.
  3. Caches — semantic caches and response caches holding old answers.
  4. Conversation memory for assistants and agents.
  5. Evaluation and fine-tuning datasets — remove the records and record the dataset version.
  6. Model weights — whether weights "contain" personal data is not settled. Reduce the risk by scrubbing PII before training, deduplicating, testing for memorisation, and retraining on a schedule from cleaned data.
  7. Processors — send deletion requests to model, vector database and observability vendors, and keep acknowledgements.

Build erasure as an event (erasure.requested) consumed by every store, with a completion report. See erasure requests and LLMs.

AI use cases: where to go next

Each AI pattern has its own DPDP pressure points. The detailed guides in this section:

If you sell into Europe, also read EU AI Act for Indian companies. Regulated entities should check sector rules: RBI FREE-AI and SEBI AI/ML rules.

Further reading on suhasbhairav.com

Before 13 May 2027

Compliance checklist

  1. Inventory every AI feature and data flow

    List each model call, index, agent tool and log store with data elements, purpose, basis, processor and region.

    Section 8(1); Rule 6(1)

  2. Assign a lawful basis per AI purpose

    Separate serving the user from improving models; obtain specific consent for training uses where no legitimate use applies.

    Sections 4, 6, 7

  3. Update notices for AI processing

    Itemise data used by AI features and describe the purposes in clear language, with withdrawal and complaint links.

    Rule 3

  4. Redact or tokenise PII before external model calls

    Insert a redaction service in the AI gateway and keep the token vault inside your boundary.

    Rule 6(1)(a)

  5. Enforce access control at retrieval and tool level

    Apply document ACLs during RAG retrieval and least-privilege credentials for agents.

    Rule 6(1)(b)

  6. Log model calls, retrievals and tool actions

    Capture request IDs, pseudonymous user IDs, purpose, model, documents and tools; review for unauthorised access.

    Rule 6(1)(c)

  7. Set tiered retention for prompts, outputs and logs

    Keep processing logs and required personal data for at least one year in restricted storage, then erase.

    Rule 8(3); Rule 6(1)(e)

  8. Sign processor contracts with AI vendors

    Cover no-training-on-your-data, security, breach notice, retention, erasure, sub-processors and region.

    Section 8(2); Rule 6(1)(f)

  9. Build erasure propagation

    Emit erasure events to source stores, vector indexes, caches, memory, datasets and vendors, with completion reports.

    Sections 8(7), 12

  10. Add AI incidents to the breach runbook

    Treat prompt-injection leaks and cross-tenant retrieval as personal data breaches with Board and CERT-In timelines.

    Rule 7

Key dates

  1. DPDP Rules notified

  2. Consent Manager registration opens

  3. Fiduciary obligations apply to AI systems processing personal data

Answers

Frequently asked questions

Does the DPDP Act apply to AI and LLM applications?

Yes. Any AI system that processes digital personal data — prompts, uploaded files, embeddings, transcripts or outputs about identifiable people — is subject to the DPDP Act. There is no AI exemption. Core obligations such as notice, security safeguards, breach intimation and retention apply from 13 May 2027.

Is there an AI law in India?

There is no standalone AI law in India yet. MeitY's India AI Governance Guidelines of November 2025 rely on existing laws such as the IT Act and DPDP Act, plus sectoral rules, and recommend voluntary frameworks. The DPDP Act is the main binding law for AI systems using personal data.

Do I need consent to use customer data in an AI feature?

You need a lawful basis for each purpose. Using an LLM to answer a customer's own query is usually covered by the basis for the service, if your notice describes it. Using her data to train or fine-tune models is a separate purpose that generally needs specific consent.

How long should I keep LLM prompts and logs in India?

Rule 8(3) requires personal data, associated traffic data and processing logs to be retained for at least one year from processing, then erased unless another law requires longer. Keep detailed logs in a restricted, encrypted tier and short-lived debugging traces separately.

Is a model API provider a Data Processor under DPDP?

Usually yes, when it processes prompts only to return outputs on your instructions. You remain the Data Fiduciary and responsible under section 8(1). If the provider uses your data for its own purposes, such as training, it starts acting as a fiduciary for that use.

Is a prompt injection data leak a personal data breach?

Yes, if it causes unauthorised disclosure or use of personal data. Rule 7 requires intimation to each affected Data Principal and to the Board without delay, with a detailed report within 72 hours. CERT-In's 6-hour reporting may also apply.

Legal references and sources

This guide explains the law for engineering and business teams. It is not legal advice; confirm how it applies to your organisation with qualified counsel, and check the Gazette for later notifications or amendments.

About the author

Suhas Bhairav

Suhas Bhairav

Systems architect and applied AI practitioner who designs production AI systems for Indian and global companies, with a focus on privacy-by-design under the DPDP Act and the EU AI Act.