DPDP 2027 and enterprise AI: where the ₹250 crore exposure actually sits

On 13 May 2027, the substantive duties of India's Digital Personal Data Protection Act become enforceable. That is about 200 days away. Notice and consent, security safeguards, breach reporting, children's data, cross-border transfers and the extra duties of Significant Data Fiduciaries all switch on together.

Most Indian enterprises are treating this as a privacy project: a consent banner, a revised policy, a grievance email address. That work is necessary, but it misses where the exposure has moved. Since 2024, personal data has been flowing into places no privacy register was designed for: copilots, chat assistants, retrieval indexes, agent logs and model vendors.

The Act does not mention artificial intelligence once. It does not need to. Every AI system that touches the personal data of an Indian customer, employee or borrower is processing under the Act. On 13 May 2027, the same law that governs your CRM will govern your prompts.

This article covers what the law actually says about AI, where the ₹250 crore figure really applies, the seven things leadership teams most often miss, and a 200-day plan.

What the law actually says about AI

The clock is fixed, for now. MeitY notified the DPDP Rules, 2025 in November 2025 with three commencement dates: November 2025 for the Board and procedural provisions, November 2026 for Consent Managers, and May 2027 for everything substantive (AZB & Partners sets out the phasing clearly). In January 2026, MeitY was reported to be considering a 12-month window for Significant Data Fiduciaries, which would have pulled their deadline into November 2026. No amendment has been notified. The only 2026 change is the Removal of Difficulties Order of 5 October, which corrects wording in sections 9 and 10 and moves no date.

There is no AI law, so DPDP is the AI law. MeitY's India AI Governance Guidelines of November 2025 are voluntary and leave rules to sector regulators. They expect DPDP implementation to settle AI's open questions, including training on public data and how consent fits AI workflows. Until that happens, DPDP is the enforceable baseline for any AI system that handles personal data.

"Publicly available" is narrower than most AI teams assume. Section 3(c)(ii) excludes only personal data that the individual made public, or that someone published because a law required it. It is not a general licence to scrape. There is also no broad legitimate-interest ground for commercial model training, so consent and purpose limitation apply (Lakshmikumaran & Sridharan explains why GDPR playbooks do not carry over).

Algorithmic due diligence is a real duty, but undefined. Rule 13(3) requires a Significant Data Fiduciary to verify that its technical measures, including algorithmic software, are not likely to pose a risk to individuals' rights. The Rules do not say what that verification involves. Model inventories, bias testing, explainability evidence and vendor assessments are the obvious candidates. Firms that define it for themselves now will set the standard their auditors use.

The Board is not yet staffed. MeitY invited applications for the Data Protection Board's Chairperson and Members in May 2026. That is a reason to use the time, not a reason to wait: the obligations bind from 13 May 2027 whether or not the first inquiry has opened.

Where the ₹250 crore exposure actually sits

The ₹250 crore figure is the most quoted and least understood number in Indian privacy. The Schedule to the Act sets seven penalty bands. The top band does not apply to every failure. It applies to one: failing to take reasonable security safeguards to prevent a personal data breach. Mapped against how AI systems actually fail, the bands look like this:

• Up to ₹250 crore (roughly $28–30 million), security safeguards (s.8(5)). An agent tricked by prompt injection into revealing customer records. An exposed vector index. A model vendor whose prompt logs are breached. These are AI's most common incident types, and they land in the top band.
• Up to ₹200 crore (roughly $22–24 million), breach notification (s.8(6)). The Rules require prompt notice to affected individuals and a detailed report to the Board within 72 hours. An AI leak that nobody classifies as a personal data breach becomes a second violation.
• Up to ₹200 crore, children's data (s.9). Consumer, edtech and gaming AI that profiles or targets minors.
• Up to ₹150 crore, Significant Data Fiduciary duties (s.10). Missing data protection impact assessments, audits or algorithmic due diligence.
• Up to ₹50 crore, everything else. Training on data collected for another purpose, or ignoring an erasure request that never reached the retrieval layer.

Two details matter to a CFO. First, these are ceilings per instance, set by the Board after an inquiry that weighs gravity, duration and mitigation. Second, they are not linked to turnover the way GDPR fines are, so a mid-market firm faces the same ceilings as a bank. One poorly governed AI incident can engage the top two bands at once. Global firms are not outside this: the Act also applies to processing outside India when it is connected to offering goods or services to people in India.

Seven things CEOs and CFOs miss

1. AI creates personal data stores you are not tracking. Prompts, chat transcripts, call summaries, embeddings in vector databases, agent memory and evaluation sets all contain personal data. Most DPDP data maps cover the CRM, core systems and HR. Few cover the copilot that has read three years of customer email.

2. Old consent does not stretch to new AI uses. Data collected to service a loan or answer a complaint was collected for that purpose. Using it to train a model or power a new AI product is a different purpose. Section 5(2) also requires fresh notice for data consented to before the Act began. Historic data lakes are where most firms will find their largest gap.

3. Erasure has to reach the AI layer. When someone withdraws consent or the purpose is complete, their data must be erased. Data in a retrieval index can be deleted. Data learned into a fine-tuned model's weights cannot be cleanly removed. Keeping personal data in governed retrieval layers rather than in weights is now a compliance decision, not only an engineering one. We covered that architecture in the 2027 context layer.

4. An AI leak is a reportable breach. Under the Act, a personal data breach includes unauthorised disclosure. If an agent reveals one customer's details to another, that is a breach that triggers notification. Most AI incident runbooks stop at "roll back the model". They need a path to the breach process, which is what the 2027 AI agent control layer is designed to provide.

5. You stay liable for your AI vendors. The data fiduciary remains responsible for what its processors do, and may engage them only under a valid contract. Pre-2025 contracts with model and SaaS vendors rarely cover prompt retention, use of your data for training, sub-processors or hosting region. Cross-border transfers are allowed by default, but stricter sector rules, such as RBI data localisation, still apply.

6. Shadow AI is processing without a contract. An employee pasting a customer complaint into a public chatbot is processing personal data through an unapproved processor, with no safeguards and no record. In most mid-market firms this is a daily habit, not an edge case.

7. AI output can break the accuracy duty. Section 8(3) requires personal data to be complete, accurate and consistent when it is used for a decision about the individual. A hallucinated summary written into a customer file, then used for a credit, claims or hiring decision, is a data quality failure under the Act, not just a model error.

Banks, NBFCs and fintechs: two frameworks, one build

Regulated financial firms are the likeliest Significant Data Fiduciaries, and they were the main target of the reported proposal to shorten the timeline. They also answer to a second reference point: the RBI's FREE-AI Committee report of August 2025 (direct PDF download). It sets out seven sutras, six pillars and 26 recommendations, and MeitY adapted the same seven sutras for its national AI guidelines.

The seven sutras are Trust is the Foundation; People First; Innovation over Restraint; Fairness and Equity; Accountability; Understandable by Design; and Safety, Resilience and Sustainability. Several map directly onto DPDP duties. Accountability resting with the deploying entity mirrors the fiduciary's liability for its processors. Understandable by Design is the evidence an SDF will need for Rule 13 algorithmic due diligence. Safety and Resilience is the security-safeguards band in practical form.

The RBI's own survey shows how far the sector has to go. Of the 127 supervised entities using AI, only 15% used interpretation tools, 18% kept audit logs and 21% monitored for data or model drift. Only about a quarter of respondents had formal processes for AI incidents. Those are exactly the records a DPDP auditor will ask for.

The practical move is to build one control set that satisfies DPDP, FREE-AI and existing RBI directions together, rather than three parallel programmes. The report itself helps: Annexure V suggests an outline for a board AI policy and Annexure VI gives an indicative AI incident reporting form. Our TRUST framework and CLEAR framework were built for the same problem in credit decisioning.

The 200-day plan

There is enough time to do this properly, but not enough to start in March. Work in three phases.

Phase 1: Find (now to 31 December 2026). Build an inventory of every AI system, including the tools teams adopted without approval. For each one, trace where personal data enters: prompts, uploaded files, retrieval indexes, logs, fine-tuning sets and vendor pipelines. Record the purpose and the lawful basis behind each use. Decide honestly whether you are likely to be designated a Significant Data Fiduciary.

Phase 2: Fix (January to March 2027). Update notices and consent for AI uses, and plan the fresh notice owed on legacy data. Add AI clauses to vendor contracts covering retention, training use, sub-processors and region. Make erasure reach indexes, caches and logs, and set retention periods for prompts and transcripts. Replace shadow AI with approved tools and a clear policy. If you are a likely SDF, define your algorithmic due diligence method and start the impact assessment.

Phase 3: Prove (April to 12 May 2027). Run a tabletop exercise in which an agent leaks customer data, and time the path to a 72-hour Board report. Test one erasure request end to end, through to the vector store. Assemble an evidence pack the board can sign off. The aim is not perfection on 13 May. It is being able to show what you did and why.

The 2027 enterprise AI readiness checklist ranks governance and data sovereignty among the top three priorities for the year. DPDP is where both become enforceable.

The CFO line

Most DPDP budgets will fund the visible parts: a consent platform, a policy rewrite, legal review. The larger cost sits underneath. It is in re-plumbing data pipelines so that purpose and erasure reach the AI layer, renegotiating vendor contracts, and building the logging and evidence that safeguards exist. That work is cheapest when it is designed into AI systems being built now, and most expensive when it is retrofitted after an incident.

Treat DPDP readiness for AI as part of the governance line in your 2027 AI budget, not as a separate compliance cost. Our 2027 AI budget guide shows how to size that line and defend it to the board.

Where to start

If you want a structured starting point, the AI × DPDP Readiness Kit packages this article into working tools. It includes an AI personal-data inventory template, a consent and purpose map for each AI use case, an AI vendor clause checklist, an erasure runbook for retrieval and fine-tuned systems, an AI-incident-to-breach playbook and a board one-pager. Sector editions for banks, NBFCs and fintechs add a Rule 13 due-diligence worksheet and a FREE-AI crosswalk.

For firms that want it done with them, our AI Production Readiness & Governance Review maps your live AI systems against DPDP in two to three weeks and leaves you with a gap report and a board summary. Book a 20-minute fit call.

Read the originals

• Digital Personal Data Protection Rules, 2025 (MeitY, full text)
• MeitY data protection framework page (Act, notifications, Board)
• PIB backgrounder on the notified Rules (November 2025)
• RBI FREE-AI Committee report (PDF download)
• Law-firm notes: AZB & Partners on the phased rollout, Hogan Lovells on the Rules, Lakshmikumaran & Sridharan on DPDP and AI

Regulatory position as of 20 October 2026. This article is analysis, not legal advice. Confirm your obligations with qualified counsel.