DPDP Act 2023 and AI Agents: What Your Engineering Team Must Know

Updated 2026-09-30. This post first said the DPDP Act’s obligations were active law. That was wrong. The DPDP Rules 2025 (G.S.R. 846(E), notified 13 November 2025) bring the Act into force in stages: the Board and institutional provisions from 13 November 2025, Consent Manager registration from 13 November 2026, and the notice, consent, security, breach, erasure, rights and penalty provisions from 13 May 2027. We have also corrected section numbers (withdrawal is s.6(4); grievances are s.8(10) and s.13; access is s.11 and erasure s.12), replaced code examples that showed an API Grantex does not have, and removed claims that any product makes you “compliant”. The EU AI Act dates now follow Regulation (EU) 2026/1744.

India’s Digital Personal Data Protection Act 2023 is on the statute book, and its main obligations apply from 13 May 2027. If your AI agents process personal data of people in India (reading emails, accessing calendars, analysing documents, managing contacts), your engineering team has specific obligations to prepare for. This post breaks down what the DPDP Act requires for AI agent deployments, why agents create unique compliance risks, and what you need to build.

What DPDP Requires

The DPDP Act establishes a framework around three roles:

Data Principal — the individual whose personal data is processed (your end user)
Data Fiduciary — the organization that determines how and why data is processed (usually you, the developer)
Data Processor — a person that processes data on behalf of the fiduciary, such as a hosting or service provider. Your AI agent is software, not a Data Processor; its processing is the fiduciary’s, or a processor’s where one runs it

The Act creates obligations in six areas that directly affect AI agent deployments:

Consent (s.6): Consent must be free, specific, informed, unconditional and unambiguous, given by a clear affirmative action. Bundling unrelated purposes into a single consent does not meet that standard.
Purpose limitation (s.4, s.6(1)): Data can only be processed for a lawful purpose the user consented to (or a legitimate use). An email-reading agent should not start accessing files without consent for that purpose.
Notice (s.5, DPDP Rules r.3): Before or when consent is asked for, the user receives a notice that itemises the data and the purposes and explains how to withdraw, exercise rights and complain to the Board.
Withdrawal (s.6(4)): The user must be able to withdraw consent at any time, and withdrawal must be as easy as granting consent. After withdrawal, processing must cease within a reasonable time (s.6(6)).
Data principal rights (s.11, s.12): Users can request a summary of their data and its processing (s.11), and correction and erasure (s.12).
Grievance mechanism (s.8(10), s.13): You must provide an effective way for users to raise grievances, and publish a response period of no more than 90 days (DPDP Rules r.14(3)).

Why AI Agents Are a Risk Vector

Traditional web applications have a relatively bounded compliance surface. The user fills out a form, the backend processes it, data is stored in a database. The data flow is predictable and auditable. AI agents are fundamentally different: Agents are autonomous. Once authorized, an agent makes decisions about what data to access and what actions to take. A calendar agent might read 500 events to find a free slot. An email summarizer processes every email in the inbox. The scope of data access is determined at runtime by the agent, not at design time by the developer. Agents delegate. Multi-agent pipelines mean one agent hands off tasks to sub-agents. The email summarizer might call a translation agent, which calls a formatting agent. Each delegation extends the data processing chain — and each link must maintain the original consent boundaries. Agents are opaque. LLM-based agents make non-deterministic decisions. You cannot predict exactly which data an agent will access or what actions it will take. This makes traditional “data processing inventory” approaches insufficient. Agents scale horizontally. A single deployment might serve thousands of users simultaneously, each with different consent profiles. Manual consent tracking does not work at this scale. These characteristics mean that consent boundaries are hard to bolt on after the agent is built. The authorization layer is a natural place to bound what an agent can do.

Four Things Your Engineering Team Should Build

Every agent authorization that relies on consent should have a structured consent record that captures:

Who gave consent (the data principal)
What was consented to (specific purposes, not vague descriptions)
When consent was given
How the user was informed (the exact notice version shown)
How long processing may continue

This is not a checkbox in your terms of service. The fiduciary bears the burden of proving consent (s.6(10)), so the record has to stand up on its own. In Grantex, a consent record binds a data principal to an existing, active grant and to a versioned consent notice, and carries a signed proof (a JWS) over both:

import { Grantex } from '@grantex/sdk';

const grantex = new Grantex({ apiKey: process.env.GRANTEX_API_KEY });

const consent = await grantex.dpdp.createConsentRecord({
  grantId: 'grnt_01HXYZ...',
  dataPrincipalId: 'user_abc123',
  purposes: [
    {
      code: 'email_summary',
      description: 'Read email subjects and bodies to generate daily summaries',
    },
  ],
  consentNoticeId: 'email-summarizer-notice',
  processingExpiresAt: '2027-12-31T23:59:59Z',
});

The record stores the purposes and the grant’s scopes side by side. Grantex does not check an agent’s later actions against the purposes; the grant’s scopes are what bound the agent, so choose scopes that match the purposes.

2. Purpose Enforcement

Consent without enforcement is just documentation. The DPDP Act allows processing only for a lawful purpose (s.4). For AI agents, a practical design is:

The agent’s grant token contains only the scopes needed for the consented purposes
Every API call verifies the token’s scopes before executing
If the agent tries to access data outside its scope, the request is denied and the denial is logged

This is where the authorization layer helps. The grant token is a signed JWT with a scp claim, and services verify the scope before executing any action. An email-reading agent cannot access files, calendar entries, or contacts: the token does not contain those scopes, and the check fails. Mapping purposes to scopes remains your design decision.

3. Right to Withdrawal

Section 6(4) requires that withdrawal of consent is as easy as giving consent. If granting consent takes one click, withdrawal should not take more. You cannot bury the withdrawal mechanism in a settings page behind three navigation levels. The withdrawal control lives in your application; Grantex records the withdrawal and, if you ask, revokes the grant:

const result = await grantex.dpdp.withdrawConsent(consent.recordId, {
  reason: 'User withdrew consent in account settings',
  revokeGrant: true,
});
// The grant and every grant delegated from it are revoked; an audit entry is written.

Without revokeGrant: true (or the server flag DPDP_WITHDRAWAL_REVOKES_GRANT=true), the grant stays active. Stopping processing in your own systems after a withdrawal is still your step (s.6(6)).

4. Audit-Ready Exports

When the Data Protection Board inquires into a complaint or a breach, you will need to produce evidence. Useful material includes:

Consent records and the notice versions they were given against
Withdrawal history with timestamps
Authorization and audit logs (what was each agent allowed to do, and what happened?)
Grievance records with the response period you published
Your breach register

Grantex’s export endpoint produces a JSON package of what it holds for a period:

curl -X POST https://api.grantex.dev/v1/dpdp/exports \
  -H "Authorization: Bearer $GRANTEX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{ "type": "dpdp-audit", "dateFrom": "2027-05-13T00:00:00Z", "dateTo": "2027-06-30T23:59:59Z" }'

It covers what Grantex records, not the data your own systems process.

The EU AI Act Is Next

While building DPDP controls, assess whether the EU AI Act applies to your deployment. Under Regulation (EU) 2024/1689 as amended by Regulation (EU) 2026/1744, the Art. 50 transparency obligations apply from 2 August 2026, high-risk obligations for Annex III systems from 2 December 2027, and for product-embedded Annex I systems from 2 August 2028. Depending on your role and risk classification, relevant supporting controls may include:

Risk management documentation (Art. 9) — what controls limit what your agents can do?
Record-keeping (Art. 12) — can you show what your agents were authorised to do and when?
Human oversight mechanisms (Art. 14) — can operators see what agents are doing and stop them?

Some of the same controls help with both laws: consent and grant records support human oversight evidence, and the audit log supports record-keeping. They do not cover the EU AI Act on their own; its Art. 50 disclosures, risk management, technical documentation and conformity assessment are separate work. Grantex’s export endpoint supports dpdp-audit, gdpr-article-15 and eu-ai-act-evidence exports from the same underlying records.

What To Do Now

Audit your current agent authorization. Are your agents using shared API keys or structured, scoped, consent-backed grants? Shared keys make it hard to show what each agent was allowed to do for whom.
Map your data processing purposes. For each agent, document exactly what personal data it accesses and why. This becomes your consent record schema.
Implement structured consent. Create consent records for every agent authorization that relies on consent, with the TypeScript, Python or Go SDK, the CLI, or the REST API. This is the foundational step.
Build the data-principal experience. Use the authenticated DPDP APIs or SDKs to list consent records, withdraw consent, request exports/erasure, and file grievances from your own application. Grantex does not currently provide a separate or embeddable DPDP end-user portal, and it is not a Consent Manager.
Set up audit exports. Run a test export now, and check it against what your counsel says you will need to produce.

The main DPDP obligations apply from 13 May 2027. Penalties are set in the Schedule to the Act (up to Rs 250 crore for failing to take reasonable security safeguards). The engineering work is straightforward if you start early with the right infrastructure. This post is not legal advice.

Learn More

DPDP Act 2023 Compliance Guide — detailed section-by-section mapping
EU AI Act Compliance Guide — article-by-article mapping
DPDP Compliance Module — SDK reference and API
Compliance Matrix — cross-framework mapping table
OWASP Agentic Top 10 and Compliance — security framework mapping

Ownership

Grantex is owned by Orchestrum Technologies LLP. Inventor and owner: Sanjeev Kumar. Ownership contact: sanjeev@orchestrum.in or mishra.sanjeev@gmail.com.

​What DPDP Requires

​Why AI Agents Are a Risk Vector

​Four Things Your Engineering Team Should Build

​1. Structured Consent Records

​2. Purpose Enforcement

​3. Right to Withdrawal

​4. Audit-Ready Exports

​The EU AI Act Is Next

​What To Do Now

​Learn More

​Ownership