Quick answer
AI customer support under DPDP must process only the personal data needed to resolve the query, protect transcripts with Rule 6 safeguards, retain logs for at least a year under Rule 8(3) and then erase them, and route rights requests correctly. Every response to a rights request must include the Rule 9 contact person, and grievances must be answered within the published period, at most 90 days.
- Resolving a customer's own query is usually covered by the s. 7(a) voluntary-provision legitimate use; reusing transcripts for training is a different purpose.
- Mask identifiers before transcripts reach the LLM and the trace log.
- Support bots must recognise access, correction, erasure and grievance requests and route them to the formal Rule 14 process.
- Rule 9 requires the DPO or contact person's details in every response to a rights request, including bot responses.
- Regulated entities should expect AI disclosure and human escalation expectations from sector regulators; check the latest circulars.
Who this is for: Customer support and CX leaders, Conversational AI engineers, Contact centre and BPO operators, Data Protection Officers, Fintech, telecom and e-commerce support teams
Which DPDP duties apply to AI customer support?
A support conversation is dense with personal data: name, phone, email, order or account numbers, addresses, payment references, and often complaints that reveal health, financial or family details. Adding AI multiplies the copies — the LLM call, the retrieval index over past tickets, agent-assist summaries, QA scoring and analytics.
The core duties, all applying from 13 May 2027 for the Rules:
- Lawful basis. When a customer contacts you and shares details to resolve her issue, section 7(a) — data voluntarily provided for a specified purpose where she has not indicated non-consent — usually covers handling that query. Using the same transcripts to train models or for marketing is a separate purpose.
- Security safeguards. Rule 6 applies to transcripts, recordings and AI logs alike.
- Retention. Rule 8(3) requires personal data, traffic data and processing logs to be kept for at least one year from processing, then erased.
- Breach. Rule 7 requires notice of every personal data breach to affected users and the Board.
- Rights and grievances. Rule 14 requires published means to exercise rights and a grievance response period of not more than 90 days; Rule 9 requires contact details in every response to a rights request.
A privacy-aware architecture for support AI
Design the pipeline so the model sees the minimum it needs.
- Authenticate before personalising. The bot answers generic questions anonymously. It fetches account data only after the customer verifies identity through OTP or a logged-in session.
- Fetch by tool, not by context dump. Give the agent narrow tools —
get_order_status(order_id)— that return only required fields, rather than injecting the full customer record into the prompt. - Mask at the gateway. Tokenise phone numbers, emails, card fragments and addresses before the LLM call; re-hydrate only in the customer-facing reply where needed.
- Ticket RAG on masked history. If the bot retrieves past tickets for answers, index masked, de-identified resolutions — not raw transcripts that contain other customers' data.
- Agent-assist boundaries. Summaries and suggested replies shown to human agents should respect the same role-based access as the CRM.
- Handoff with context, not copies. Escalation passes a ticket ID and a masked summary; the human agent opens the record in the CRM under their own permissions.
| Component | Personal data | Key control |
|---|---|---|
| Bot conversation | Full transcript | Masking, Rule 8(3) retention |
| Tool calls | Account fields | Field-level scoping |
| Ticket RAG index | Past resolutions | De-identification |
| QA and analytics | Scores, samples | Pseudonymous IDs |
Routing rights requests and grievances from the bot
Customers will type "delete my data", "what do you know about me" or "this is wrong, fix my address" into the support bot. Under DPDP these are requests to exercise rights under sections 11 to 13, and the bot must not treat them as ordinary chat.
Build an intent classifier for:
- Access — summary of personal data and processing (section 11).
- Correction, completion, updating (section 12).
- Erasure (section 12).
- Withdrawal of consent (section 6).
- Grievance — a complaint about how data was handled (section 13).
- Nomination (section 14; Rule 14(4)).
On detection, the bot should open a formal request through the channel published under Rule 14(1), collect the identifiers your terms of service require (Rule 14(1)(b) and 14(5) list examples such as customer ID, email and mobile), give a reference number and state the response timeline.
Rule 9 requires the business contact information of the DPO or designated contact person in every response to a communication exercising rights, including automated acknowledgements. Section 13(3) requires Data Principals to exhaust your grievance redressal before approaching the Board, so a support bot that drops these requests weakens your position. See grievance redressal.
Voicebots and call recordings
Voice support adds audio, which carries more than words: voice characteristics, background speakers and emotional cues.
- Notice at call start. State that the call is handled or assisted by an automated system and may be recorded, with a short path to the full notice.
- Transcribe, then minimise. Many teams keep the masked transcript for the retention period and expire raw audio earlier where no law or dispute requires it. Rule 8(3)'s one-year minimum applies to personal data and processing logs; decide and document whether audio is needed to meet it in your context.
- Do not build voiceprints casually. Using voice for authentication creates biometric-style data; treat it as a distinct purpose with explicit consent.
- Speech vendors are processors. Check retention and training-use terms for speech-to-text and text-to-speech APIs.
- Background voices. Recordings can capture other people near the caller; keep retention short and never use such audio for training.
The voice-specific issues are covered in voice AI and call recordings.
Sector expectations and EU users
Support AI in regulated sectors carries extra expectations.
- Banks and NBFCs. The RBI FREE-AI committee report (August 2025) recommends disclosing to customers when they interact with AI, and providing consumer grievance mechanisms with human escalation. These recommendations bind only once RBI issues directions or circulars; check the latest position.
- Securities market. SEBI's 2025 amendment makes a regulated person using AI tools solely responsible for data privacy, security and outputs, regardless of vendor.
- Insurers, telecom, health. Sector regulators and laws may impose record-keeping and complaint-handling requirements that are stricter than DPDP; those continue to apply.
If the bot serves people in the EU, Article 50 of the EU AI Act requires providers to design AI systems that interact with people so they are informed they are interacting with AI, unless obvious; this applies from 2 August 2026. GDPR also applies to EU customers' data. See AI Act Article 50 checklist.
A clear "you are chatting with an AI assistant — type 'agent' for a person" banner satisfies both the EU rule and the direction of Indian sector guidance at almost no cost.
Common mistakes
- Full CRM record in the system prompt. The model sees far more than the query needs, and every trace stores it.
- Training on raw transcripts. Fine-tuning a support model on unmasked tickets copies thousands of customers' data into weights you cannot erase.
- Bot ignores rights requests. "I can help with orders only" in response to "delete my account" is a failed rights request.
- No Rule 9 contact in acknowledgements. Automated replies to rights requests still need the contact person's details.
- Agents pasting into public chatbots. Human agents copying customer data into consumer AI tools outside your contract. Provide an approved, gateway-backed assistant instead — see is ChatGPT safe with customer data.
- Indefinite transcript storage. Rule 8(3) is a one-year minimum followed by erasure, not a licence to keep everything.
- Outsourced support without contract terms. BPO partners are processors; your contract must cover safeguards under Rule 6(1)(f). See DPDP for IT services and BPO.
Further reading on suhasbhairav.com
- BlogPersistent Agent Memory and Privacy: Balancing Personalization with Data Retention Risks
- BlogVoice Agent Compliance: Consent, Audit Trails, Escalation
- BlogGDPR Right-to-Erasure in Vector Stores with AI Agents
- AI TemplatesNext.js AI Chatbot Templates and ChatGPT Clone Starters
- AI Use CasesAI Agent Use Case: Call Centers Using Conversation Transcripts to Monitor Service Quality
Before 13 May 2027
Compliance checklist
Authenticate before fetching account data
Answer generic queries anonymously and fetch personal data only after verification.
DPDP Act s. 8(5)
Use narrowly scoped tools
Return only the fields needed for each support action instead of injecting full customer records.
DPDP Act s. 6(1)
Mask transcripts before LLM calls and logs
Tokenise identifiers at the gateway and log only masked text.
Rule 6(1)(a)
Detect and route rights requests
Classify access, correction, erasure, withdrawal, grievance and nomination intents and open formal requests.
Rule 14(1)
Add Rule 9 contact details to rights responses
Include the DPO or contact person's business details in every acknowledgement and reply.
Rule 9
Publish and meet a grievance timeline
Set a response period well under the 90-day ceiling and track it.
Rule 14(3)
Set transcript and recording retention
Keep processing logs at least one year, then erase on schedule.
Rule 8(3)
Disclose AI and offer human escalation
Tell customers they are talking to an AI and provide an easy route to a person.
EU AI Act Art. 50; RBI FREE-AI (recommendation)
Bind BPO and AI vendors by contract
Cover safeguards, retention, training use and breach notice.
Rule 6(1)(f)
Key dates
DPDP Rules notified
EU AI Act Article 50 transparency obligations apply
Rules 6–9 and 14 apply to support operations
Answers
Frequently asked questions
Do I need consent to use an AI chatbot for customer support in India?
Do I need consent to use an AI chatbot for customer support in India?
Usually not for handling the customer's own query. When she contacts you and provides details to resolve an issue, the section 7(a) legitimate use generally applies. Consent becomes necessary for different purposes, such as reusing transcripts to train models, marketing, or building voiceprints, and those should be presented separately.
How long can I keep chatbot transcripts under DPDP?
How long can I keep chatbot transcripts under DPDP?
From 13 May 2027, Rule 8(3) requires personal data and processing logs to be kept for at least one year from processing, then erased unless another law requires longer. Sector rules may set longer record-keeping periods. Do not keep transcripts indefinitely, and mask them so the retained copy exposes as little as possible.
What should a support bot do when a user asks to delete their data?
What should a support bot do when a user asks to delete their data?
Treat it as a formal erasure request, not a chat message. Open a request through your published Rule 14 channel, collect the identifiers your terms require, give a reference number and timeline, and include the Rule 9 contact person's details. Then run your erasure workflow across CRM, transcripts and AI stores.
Can I train a support model on past tickets?
Can I train a support model on past tickets?
Training is a different purpose from resolving the original query, so plan for consent or robust de-identification. Mask or tokenise identifiers, remove free-text personal details, keep lineage so erased customers can be excluded, and test the fine-tuned model for memorisation of real customer data.
Do I have to tell customers they are talking to an AI?
Do I have to tell customers they are talking to an AI?
The DPDP Act has no explicit AI-disclosure rule, but EU AI Act Article 50 requires it for EU users from 2 August 2026 unless obvious, and the RBI FREE-AI report recommends it for regulated entities. A clear AI banner with a human-agent option is a low-cost default.
Is my BPO or contact-centre partner liable under DPDP?
Is my BPO or contact-centre partner liable under DPDP?
The BPO is usually your Data Processor, and section 8(1) keeps you responsible as the Data Fiduciary. You may engage it only under a valid contract, which should provide for reasonable security safeguards under Rule 6(1)(f), retention, erasure and prompt breach notice so you can meet Rule 7.
Legal references and sources
- Digital Personal Data Protection Act, 2023 — Sections 7(a), 8 and 13 (official text)
- Digital Personal Data Protection Rules, 2025 — Rules 9 and 14 (official text)
- Digital Personal Data Protection Rules, 2025 — Rules 6 and 8(3) (official text)
- Digital Personal Data Protection Rules, 2025 — Rule 7 (official text)
- DPDP Rules, 2025 (G.S.R. 846(E)) ↗
- MeitY data protection framework ↗
- Reserve Bank of India ↗
- EU AI Act (Regulation (EU) 2024/1689) ↗
This guide explains the law for engineering and business teams. It is not legal advice; confirm how it applies to your organisation with qualified counsel, and check the Gazette for later notifications or amendments.
About the author

Suhas Bhairav
Systems architect and applied AI practitioner who designs production AI systems for Indian and global companies, with a focus on privacy-by-design under the DPDP Act and the EU AI Act.
Related guides
- AI & personal dataVoice AI and Call Recordings Under the DPDP ActCall recordings, transcripts, voice bots and voiceprints all process personal data. How to handle notice, consent, redaction, retention and vendors under DPDP.
- AI & personal dataDPDP Compliance for RAG Chatbots and Knowledge AssistantsRAG chatbots copy personal data into chunks, embeddings and chat logs. This guide maps DPDP duties to ingestion, retrieval, generation and deletion.
- DPDP Rules 2025DPDP Grievance Redressal: The 90-Day Response RuleRule 14(3) requires every Data Fiduciary and Consent Manager to publish a grievance response period of no more than 90 days and to build systems that meet it, from 13 May 2027.
- AI & personal dataLLM Logging and Observability Under DPDP's One-Year RuleDPDP asks you to minimise data yet keep processing logs for a year. A tiered design for LLM traces resolves the tension: log less, mask early, lock access, then erase on schedule.
- AI & personal dataIs ChatGPT Safe to Use With Customer Data in India? (DPDP View)Whether ChatGPT or any hosted LLM is safe for customer data depends on the tier, contract and your controls. Under DPDP you stay the Data Fiduciary and liable.
- Sector guidesDPDP for IT Services, BPO and Outsourcing FirmsWhat the DPDP Act's section 17(1)(d) outsourcing exemption covers, what still applies to Indian IT services and BPO firms, and how to treat domestic and employee data.