The DPDP Act and your AI chatbot, what to do before May 2027.
India’s Digital Personal Data Protection Act never mentions AI, but every chatbot and voice agent that handles customer data is covered by it. Here is what it means for notices, consent, chat transcripts, training data, leaks and children, with a checklist to work through before the main duties apply on 13 May 2027.
- A chatbot or voice agent that handles names, phone numbers, orders, recordings or transcripts is processing personal data, so the DPDP Act applies.
- People need a clear notice, and a real choice where consent is the basis. Withdrawing consent must be as easy as giving it.
- Transcripts and recordings must be kept safe and deleted when their purpose is done. Training AI on them usually needs a separate basis.
- A chatbot that shows one customer’s data to another, for example after a prompt injection, has caused a personal data breach that must be reported.
- The main duties and penalties apply from 13 May 2027. The highest penalty, ₹250 crore, is for failing to keep data safe.
Your chatbot is covered
The Digital Personal Data Protection Act, 2023 (the DPDP Act) applies to digital personal data: data about a person who can be identified from it. A support chatbot that asks for a phone number, looks up an order, reads a complaint or records a call is processing personal data. So is an HR assistant that answers questions about leave, and a voice agent that calls sales leads.
The company that decides why and how the data is used is the Data Fiduciary, and it carries the duties, even when the chatbot runs on another company’s AI model. The model provider, the chatbot platform and the cloud host are usually Data Processors working for you. Our DPDP Act guide explains the roles and dates in full; this article is about what they mean for conversational AI.
The timing: the DPDP Rules were notified on 13 November 2025, and the main duties for businesses, with penalties, apply from 13 May 2027. Rebuilding consent flows, retention rules and vendor contracts for a live chatbot takes months, so the work starts now.
Notice and consent, inside the chat
A chatbot is a good place to get notice and consent right, because the conversation itself can carry them:
- Notice (Section 5). When you ask for personal data, say what you will collect and why, in plain words. The Rules ask for a notice that stands on its own, lists the data and the purposes, and links to how people can withdraw consent, use their rights and complain to the Data Protection Board. People can ask for it in English or any of the 22 languages of the Constitution’s Eighth Schedule, so a bot that chats in Hindi should offer its notice in Hindi too.
- Consent (Section 6). Where consent is your basis, it must be free, specific, informed, unconditional and unambiguous, given by a clear action, and limited to the data the purpose needs. No pre-ticked boxes, and no “by continuing to chat you agree to everything”.
- Legitimate uses (Section 7). Some processing needs no consent, for example when a person voluntarily gives data for a specific purpose and has not objected to its use. A customer typing an order number to track a parcel may fall here. Your lawyer should confirm the basis for each purpose.
- An easy exit (Section 6(4)). Withdrawing must be as easy as agreeing. If “yes” is one tap in the chat, “stop” should be too.
Transcripts and voice recordings
Every conversation leaves a record: the transcript, any uploaded files, and for voice agents, the recording and its text. These are personal data like any other:
- Keep them safe (Section 8(5)). The Rules list minimum safeguards: encryption, masking or tokens; access control; logs and monitoring of who accessed what, kept for a year; backups; and security terms in contracts with processors.
- Delete them when the purpose is done (Section 8(7)). Set a retention period for transcripts and recordings before launch, and enforce it everywhere, including your chatbot vendor’s systems and any analytics copies.
- Limit who sees them. When a human agent takes over a chat, they see its history. Give them only what the task needs.
Training AI on customer conversations
Conversations make valuable training data, which is exactly why the Act limits their use. Data collected to answer a support question was collected for that purpose. Using it to train or fine-tune a model is a new purpose, and usually needs fresh consent or another valid basis. In practice:
- Decide, and write down, whether conversations will ever be used for training, before you collect them.
- If they will, ask separately and clearly, and respect a “no”.
- Remove or mask personal data in anything used for training or evaluation.
- Check your AI model provider’s terms, so that your customers’ data is not used to train its models without your agreement and your customers’ knowledge.
When a chatbot leaks, it is a breach
The Act defines a personal data breach broadly: any unauthorised processing, or accidental disclosure, sharing, change, loss or destruction of personal data, or loss of access to it. A chatbot that reveals one customer’s address to another fits that definition, however it happened.
The usual causes are well known: a prompt injection that talks the bot past its rules, a document search that ignores who is asking, a tool that can look up any customer rather than the one in the chat, or a memory feature that carries data between users. When it happens, you must tell the Data Protection Board and each affected person without delay, and send the Board a detailed report within 72 hours. If it is also a cyber incident, CERT-In expects a report within six hours.
The highest penalty in the Act, up to ₹250 crore, is for failing to take reasonable security safeguards to prevent a breach. For a chatbot, testing whether it can be tricked into leaking data, and fixing what you find, is one of the clearest ways to show your safeguards were reasonable.
Children
Anyone under 18 is a child under the Act. Processing a child’s data needs verifiable consent from a parent or lawful guardian, and you must not track, behaviourally monitor or target advertising at children. Edtech, gaming and consumer apps with chatbots should assume children will use them and design for it: age checks, parental consent and no profiling of young users. The Rules exempt some classes of organisation and purpose from parts of these duties, so check what applies to you.
People’s rights, through the chat
People can ask for a summary of their data, have it corrected or erased, and name someone to act for them. They can complain, first to you and then to the Board, and your grievance process must answer within 90 days. A chatbot can make this easy by taking requests in the conversation, as long as it confirms who is asking and passes each request to people and systems that can act on it.
Your AI vendors
You stay responsible for personal data your processors handle, including your AI model provider, chatbot platform and cloud host. Put a contract in place with each, covering security safeguards, breach notice to you, deletion at the end of the contract and limits on using your data. The Act allows personal data to be transferred outside India except to countries the government restricts, but sector rules, such as the RBI’s for payment data, can still require data to stay in India.
A checklist for your chatbot
- Map what personal data each chatbot and voice agent collects, sees and stores, including files and recordings.
- Choose and record the basis for each purpose: consent or a legitimate use.
- Write a plain notice for the chat, with the links the Rules require, in the languages your users speak.
- Make consent a real choice, and make “stop” as easy as “yes”.
- Ask only for the data the task needs.
- Limit the bot’s tools and searches to the person in the chat.
- Encrypt transcripts and recordings, log access, and set retention periods.
- Decide whether conversations are ever used for training, and get the right basis if so.
- Sign data processing terms with every AI vendor.
- Red-team the bot for data leaks, fix what you find, and rehearse the 72-hour breach report.