Five duties in India's data protection law, the DPDP Act, reach an AI build. Consent is tied to a stated purpose. People can withdraw. Data behind a decision must be accurate. Data is erased when its purpose ends. Breaches are reported promptly. Each one is easier to meet if the system records where its data came from. This guide takes the duties in turn and says what to build for each.
When the duties apply
The Rules were published in the Gazette dated 13 November 2025. Under Rule 1 of the DPDP Rules, 2025, Rule 4 on consent managers applies after one year, and most other rules, including Rules 3 and 5 to 16, apply eighteen months after publication, which is mid-May 2027. Check the notifications for the sections of the Act you depend on, since the Act itself came into force by notification.
Consent is tied to a purpose
Section 6(1) of the Act says consent must signify agreement to processing for the specified purpose and be limited to the data necessary for that purpose. Customer data gathered to service a loan was not gathered to train a model, and having it on file does not make it usable for that.
That leaves four routes for training and testing data:
| Route | What it needs | Watch for |
|---|---|---|
| Consent that names the training purpose | A notice and a clear action from the person | Old customers who agreed to something narrower |
| A lawful use without consent under section 7 | A reading of that section by counsel | On one reading of the text, which counsel should confirm, section 7 lists specific uses and training is not among them by name |
| Data outside the Act | The data is not about an identifiable person (section 2(t)) | Re-identification; see below |
| The research exemption | Section 17(2)(b) and Rule 16 with the Second Schedule standards | The data must not be used to take a decision specific to a person |
The last row matters for lenders and insurers. A model that scores applicants takes decisions about individuals, so the research exemption in section 17(2)(b) does not appear to fit it. That is one reading of the text, which counsel should confirm. A model built for aggregate statistics may.
Withdrawal has to reach the model
A person can withdraw consent at any time (section 6(4)). The company must then stop processing and have its processors do the same (section 6(6)), and erase the data unless a law requires retention (section 8(7)).
Deleting a row is easy. Taking one person's influence out of a trained model is hard to do and hard to prove. Design around the need instead:
- Trace every training record to its consent and its source. A company that cannot list what a model was trained on cannot act on a withdrawal.
- Make retraining routine. If retraining takes half a year, withdrawal requests will wait for it.
- Prefer retrieval where it meets the need. When an answer comes from a document in an index, deleting the document removes it from later answers. A fine-tuned model offers no equivalent. RAG beyond the demo explains how retrieval systems are built.
- Count the side copies. Caches, embeddings, logs, evaluation sets and backups hold personal data too.
Plan erasure against Rule 8(3), which sets a minimum retention of one year for personal data, traffic data and logs of processing, for the purposes in the Seventh Schedule. The Rules' own illustration of a cloud provider hosting customer records says the provider must keep data and logs for that year as well.
Decisions need accurate data, and correction needs a re-run
Where personal data is likely to be used to make a decision about a person, section 8(3) requires the company to ensure its completeness, accuracy and consistency. Section 12(2) lets the person ask for correction, completion or updating.
For a plain record, that is an edit. For a score or segment derived from the record, it raises the question of what else must change. If the derivation can be repeated on corrected inputs, correction means running it again. If it happened once and the working was not kept, correction is an investigation. The fields that make a repeat run possible are listed in the audit-trail field list.
Test anonymisation, do not assume it
Section 2(t) defines personal data as data about an individual who is identifiable by or in relation to the data. Swapping a name for an ID code still leaves a person identifiable by the code, so treat pseudonymised data as personal.
Whether a dataset is anonymous depends on whether people can be identified from it. Rich behavioural records can often be matched to someone from a handful of attributes. Run a re-identification test with realistic outside data, record the result and repeat it when the dataset changes.
Where the model runs
A hosted model API makes its provider a processor. Section 8(1) keeps the company responsible for processing done on its behalf, and section 8(2) allows a processor only under a valid contract. Rule 6(1)(f) asks for safeguard terms in that contract.
Retention terms, sub-processors and processing location are therefore contract points, not preferences. Rule 15 adds that transfers abroad are subject to requirements the Government may set for making data available to foreign States. For banks and insurers, sector rules narrow the choice further; see AI governance for regulated Indian companies.
Breach reporting needs a log
Under Rule 7, a company that becomes aware of a breach tells each affected person without delay and tells the Board without delay, with detailed information within seventy-two hours unless the Board allows longer. Rule 6(1)(c) and (e) asks for logs that allow detection and investigation, kept for one year. A model pipeline that sends prompts to several services needs a log that shows who received what.
The Act's Schedule sets the penalty ceilings:
| Breach | Ceiling |
|---|---|
| Failure to take reasonable security safeguards (section 8(5)) | Highest penalty band in the Schedule |
| Failure to notify a breach (section 8(6)) | Second penalty band in the Schedule |
| Any other provision of the Act or Rules | Lowest penalty band in the Schedule |
Larger companies may face more
The Government can notify a company as a Significant Data Fiduciary under section 10. Rule 13 then asks for a yearly impact assessment and audit, and due diligence that algorithmic software does not pose a risk to people's rights. Watch the notifications if you process data at scale.
What to build
- A data map listing each personal field, its consent basis and its stated purpose.
- A written data boundary: what may leave your environment, what must stay and what is masked.
- Retrieval first, fine-tuning only where measured results justify it.
- Lineage per training record, with a retraining routine.
- A replayable log of inputs, context, model version and output.
- A named owner for grievances and requests. The forms checklist covers the notice side.
Questions teams ask
Can we train on customer data we already hold?
Only if one of the four routes above covers the purpose. Holding the data is not enough.
What happens when a customer withdraws?
You stop processing, instruct processors to stop and erase the data unless a law requires retention. Your lineage record shows what to remove and what a retrain must exclude.
Do the duties differ for a chatbot with no accounts?
The Act covers any digital personal data you process. A chat that captures names or numbers brings the notice and consent duties with it.
Where Brihat fits
Brihat Infotech designs consent, purpose limits, retention and erasure into the data model of the platforms it builds, and puts a grievance officer on the launch checklist for any platform that holds personal data. See its compliance page. Start a conversation.