India’s Digital Personal Data Protection framework stopped being a planning exercise on 13–14 November 2025, when the Ministry of Electronics and Information Technology notified the DPDP Rules, 2025 and constituted the Data Protection Board of India. The Act itself received Presidential assent in August 2023; the Rules are what made it operable.
For teams building AI systems, this matters more than a general privacy law would, because one of the criteria for enhanced obligations explicitly contemplates AI-based profiling at scale. If you train on, retrieve, or reason over personal data belonging to people in India, the timeline below is your timeline.
And the awareness gap is real: EY India’s 2026 readiness survey found 71% of Indian enterprises still have limited understanding of the Act.
The enforcement timeline, with dates
| Milestone | Timing | What changes |
|---|---|---|
| Rules notified; Board constituted | 13–14 Nov 2025 | Act operable. Data Protection Board of India established in the NCR. Penalty framework live. |
| Consent Manager ecosystem | Around mid-2026 | Interoperable consent platforms operationalised; systems need to be compatible with Consent Manager APIs. |
| Consent Manager registration | Around Nov 2026 | Registration window; only India-incorporated entities meeting a net-worth threshold qualify. |
| Soft enforcement ends / legacy data | Nov 2026 | One year on, the Board is expected to shift from guidance toward active supervision. Historical data needs demonstrable lawful basis. |
| First SDF audit cycle | Q1 2027 | Independent data audits and DPIAs become operational for Significant Data Fiduciaries. |
| Full enforcement | 13–14 May 2027 | End of the 18-month transition. Full adjudicatory powers and penalties. |
Penalties reach ₹250 crore per violation and stack across categories. MeitY has floated compressing the 18-month window to 12 months, which would bring the hard deadline forward to around November 2026. Plan against the possibility.
Are you a Significant Data Fiduciary? The AI criterion matters
SDF status is notified by the central government rather than self-declared, but the indicative thresholds are public. An organisation is a likely candidate if it meets any of:
- Processing the personal data of roughly 5 million or more residents
- Annual turnover in the region of ₹250 crore or more
- Processing sensitive categories such as health or financial data, or using AI for sizeable decision-making or profiling
That third limb is the one AI teams tend to miss. A platform well below the user and revenue thresholds can still attract enhanced obligations purely because of how it uses AI to make or influence decisions about people.
SDF obligations are substantive: an India-based Data Protection Officer reporting to the board, an independent data auditor, and Data Protection Impact Assessments for high-risk processing — all expected to be operational by Q1 2027.
What this means for AI architecture specifically
Training data provenance
If personal data sits in a training set, you need a lawful basis for it and a record demonstrating that basis. Legacy datasets assembled before the framework are a particular exposure: from late 2026, organisations unable to show valid consent or another lawful ground for historical data face increased regulatory risk. Scraped or inherited datasets with unclear provenance are the hardest to defend.
Retrieval context and prompts
A RAG system that pulls customer records into an LLM prompt is processing personal data at inference time, not only at training time. The relevant questions are what the retriever can reach, whether access control is enforced during retrieval rather than after it, and whether prompts containing personal data are being logged.
Trace logs are a data store
This is the most commonly overlooked exposure in agent systems. Full trace capture — every step, tool call and intermediate output — is enormously useful for debugging and evaluation. It also silently creates a new repository of personal data with its own retention, access and erasure obligations. Instrument deliberately: decide what gets redacted before it lands in a trace, not after.
Erasure has to reach the model estate
When a data principal exercises erasure rights, deleting the row is the easy part. Ask where else that data now lives: in the vector index, in cached embeddings, in evaluation datasets, in trace logs, in fine-tuning corpora. Systems designed without this question tend to answer it expensively.
Notice, in the right language
Under Section 5(3), the privacy notice must be available in English or any of the 22 languages in the Eighth Schedule to the Constitution, where requested. For AI products serving users across India, machine-translating a legal notice is not the same as producing a legally accurate one.
How this stacks with sectoral AI regulation
DPDP is the horizontal layer. Financial-sector organisations sit under additional expectations:
- RBI FREE-AI framework (August 2025) — seven guiding principles and 26 recommendations across six pillars, covering explainability, disclosure, grievance redressal and drift monitoring.
- SEBI — Regulation 16C introduced by the Intermediaries (Amendment) Regulations in February 2025, plus a June 2025 consultation paper proposing guiding principles for responsible AI/ML use in the securities market, alongside the 2019 AI/ML reporting circulars.
- NITI Aayog — the national Responsible AI principles that set the vocabulary the sectoral regulators draw on.
The directions are consistent: human oversight of consequential decisions, transparency about AI involvement, accountability resting with the deploying entity, fairness across populations, and security of personal data. A team building for those five properties is broadly aligned across all of them.
A practical sequence for AI teams
- Inventory every system that touches personal data, including AI components inside vendor products. Most organisations find more than they expected.
- For each, record the lawful basis, the retention period, and where derived copies live — indexes, caches, evaluation sets, trace logs.
- Assess SDF likelihood honestly, paying attention to the AI-profiling limb rather than only user counts.
- Move access control into the retrieval layer so restricted content never reaches the application tier.
- Define redaction rules for trace logging before you scale instrumentation.
- Build an erasure path that reaches every derived store, and test it end to end.
- Fix legacy data provenance now — it is the item with the longest lead time and the least flexibility.
Frequently asked questions
When does India’s DPDP Act become fully enforceable?
The DPDP Rules were notified on 13–14 November 2025 with an 18-month transition. Full substantive compliance and the Data Protection Board’s full adjudicatory powers are expected from 13–14 May 2027. MeitY has proposed compressing that window to 12 months, which would move the deadline to around November 2026.
What are the penalties under the DPDP Act?
Penalties reach ₹250 crore per violation and can stack across categories, so a single incident touching several obligations can create substantially larger cumulative exposure.
Does using AI make my company a Significant Data Fiduciary?
It can contribute. SDF designation is notified by the central government, but indicative criteria include using AI for sizeable decision-making or profiling — alongside volume, turnover and sensitivity thresholds. An organisation below the user-count threshold can still qualify on the AI limb.
Do AI trace logs fall under the DPDP Act?
If they contain personal data, yes. Full-trace instrumentation for agent debugging creates a new store of personal data subject to the same retention, access and erasure obligations as any other. Redaction should be designed in before instrumentation scales.
What happens to personal data collected before the DPDP Rules?
Legacy data needs a demonstrable lawful basis. From around November 2026, organisations that cannot show valid consent or another lawful ground for historical datasets face increased regulatory exposure.
This article summarises publicly available regulatory material for general information and is not legal advice. Organisations should consult the primary MeitY and sectoral publications and their own counsel.
