FUTUREISNOW · STRATEGIC SIGNAL
AI agents do not automatically violate India’s DPDP Act. But autonomy increases the importance of purpose limitation, access control, security, transparency and accountable oversight. The practical risk is not that an agent is inherently unlawful; it is that an organization may let it access or use personal data without adequate authorization, traceability or safeguards.
The signal: autonomy changes the control problem
Traditional applications usually follow predefined workflows: a user requests a record, a service checks permissions, and an action is logged. An AI agent can plan multiple steps, call tools, retrieve information and adapt its next action based on intermediate results. That flexibility can improve productivity—but it also creates more paths through which data can be accessed, combined, disclosed or changed.
The governance challenge is to ensure that each tool call and downstream action stays within an approved purpose and authority boundary. A company cannot rely only on a model’s instructions to behave responsibly. It needs technical controls around the model, plus policies and accountable people who determine what the system is allowed to do.
What the DPDP framework means for agentic workflows
Timing matters: the DPDP Act is being brought into force in phases, not all at once. The next scheduled tranche is 13 November 2026, when section 6(9) and section 27(1)(d) are due to commence. The broader set of core processing obligations—including sections 3–17, which contain the main processing duties and data-principal rights—is scheduled for 13 May 2027. Teams should map each control to the provision and commencement date that actually applies, rather than describing the entire Act as already fully operative. See the officially reproduced DPDP commencement notification and MeitY’s DPDP Rules and commencement documents.
India’s Digital Personal Data Protection Act, 2023 places obligations on Data Fiduciaries for processing digital personal data, including requirements related to lawful processing, notice and consent where applicable, security safeguards, breach response and rights of Data Principals. The precise obligations depend on the circumstances, statutory provisions, applicable exemptions and commencement timeline of the DPDP Rules, 2025.
There is no blanket rule that every AI agent is unlawful or that every processing step requires a separate consent prompt. Instead, organizations need to establish the purpose and applicable basis for processing, provide required notices, honor relevant rights, and ensure that vendors and systems operate within those boundaries. Obtain legal advice for high-impact or ambiguous use cases.
The official DPDP Act, 2023 and DPDP Rules and commencement documents should anchor legal interpretation. Teams should not infer requirements from dramatic headlines or assume every provision became effective on the same date.
Where agentic systems create exposure
- Over-broad access: An agent inherits credentials that expose more records or systems than its task requires.
- Purpose drift: Information collected for one business process is reused for a different objective without an appropriate assessment.
- Prompt injection and untrusted content: A document, email or webpage attempts to redirect an agent into disclosing data or invoking unauthorized tools.
- Uncontrolled onward sharing: Prompts, retrieved passages or outputs are sent to external model providers, plug-ins or logging systems without suitable review.
- Weak traceability: Logs record that an agent ran, but not which data it accessed, which policy allowed the access or what action followed.
- Unreviewed high-impact actions: The agent changes a customer record, initiates a payment or sends a consequential communication without a proportionate approval gate.
A safer architecture: bounded autonomy
Enterprises should treat an agent as an untrusted decision-making component operating inside a controlled system. The model can propose a plan, but separate services should enforce identity, permissions, data boundaries and action policies. The controls should be independent of whether the model follows its prompt correctly.
| Control layer | What to implement | Evidence to retain |
|---|---|---|
| Identity and access | Dedicated service identities, least privilege, scoped tokens and short-lived credentials | Identity, requested scope, policy decision and access events |
| Purpose and data access | Approved use cases, data minimization, field-level restrictions and retrieval filters | Purpose mapping, data source, access rationale and policy version |
| Tool execution | Allowlisted tools, validated parameters, transaction limits and separate authorization checks | Proposed action, validation result and executed action |
| Human oversight | Approval gates for sensitive, irreversible or high-impact actions | Reviewer identity, decision, timestamp and exception reason |
| Monitoring and response | Redaction, anomaly detection, incident procedures, kill switches and rollback paths | Correlated logs, incident record and remediation evidence |
These controls are useful security engineering practices, not a guarantee of legal compliance. The organization still needs to map them to its actual data, purposes, applicable rules and risk profile. For more on operational controls, see our guide to agentic AI governance for production workflows.
A practical example: customer-retention analysis
Suppose a business asks an agent to identify customers who may be at risk of leaving. A poorly controlled agent might combine purchase history, support tickets, location data and unrelated personal information simply because those sources are technically available. A safer design defines the approved purpose, limits the fields and sources that can be queried, tests whether the analysis is appropriate, and prevents the agent from taking consequential actions without authorization.
The system should record which approved data sources were used, which policy authorized access, what the model recommended, what validation occurred and whether a human approved the next step. If the request falls outside the approved purpose or available permissions, the agent should stop and escalate—not improvise a workaround.
A 30-day readiness checklist
- Inventory agents and tools. List production and pilot agents, service identities, connectors, data sources and external providers.
- Classify the use cases. Document purpose, data categories, expected actions, affected people and consequences of failure.
- Review the legal basis and notices. Map the processing to applicable DPDP requirements, sector rules, contractual terms and relevant commencement dates.
- Reduce permissions. Remove shared credentials, limit tool scopes and separate read access from write or transaction authority.
- Test adversarial scenarios. Simulate prompt injection, data exfiltration, conflicting instructions, stale permissions and repeated actions.
- Define approval thresholds. Specify which actions are prohibited, which require human approval and how exceptions are handled.
- Verify the audit trail. Ensure logs can connect user requests, tool calls, data access, policy decisions and final actions without unnecessarily duplicating personal data.
- Assign accountable owners. Establish ownership across product, security, privacy, engineering and the business process.
The decision for leaders
The right response to agentic AI is neither blanket prohibition nor unrestricted autonomy. It is to match the system’s authority to the risk of the task. Low-impact, reversible work may be automated more broadly; access to sensitive records, external disclosure and consequential actions require tighter boundaries and stronger evidence.
For Indian enterprises, compliance should be designed into the operating model rather than added after an agent reaches production. Clear purposes, least-privilege access, independent tool controls, reliable logs and accountable human oversight turn autonomy from an uncontrolled capability into a governed business process.
Signals over Noise. FutureIsNow tracks the governance choices shaping enterprise AI adoption.



