AI & personal dataUpdated 22 November 20269 min read

Is 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.

By Suhas Bhairav · Reviewed against the DPDP Rules, 2025 (G.S.R. 846(E), 13 Nov 2025)

Your role
Data Fiduciary
Vendor role
Data Processor (if used only on your instructions)
Liability
Section 8(1) — stays with you
Contract
Required, s.8(2); security clauses, Rule 6(1)(f)
Log retention
At least 1 year (Rule 8(3))
Applies from
13 May 2027

Quick answer

ChatGPT or any hosted LLM can be used with customer data in India only if you treat the vendor as your Data Processor: use a business or enterprise tier with a data processing agreement, confirm training opt-out, retention and region, redact personal data where possible, and log usage. Under section 8(1) of the DPDP Act, you remain responsible for compliance from 13 May 2027.

  • Under DPDP, your company is the Data Fiduciary and the LLM vendor is your Data Processor for customer data you send.
  • Section 8(1) keeps liability with you regardless of what the vendor's terms say.
  • Consumer chatbot accounts used by staff are the highest-risk pattern; route usage through approved business tiers.
  • Verify current vendor terms yourself: training use, retention, region, sub-processors and breach notice.
  • Redact or tokenise customer identifiers before prompts wherever the task allows.
  • Staff pasting customer data into an unapproved tool can amount to a personal data breach.

Who this is for: Founders and CTOs, Customer support and sales leaders, CISOs and IT administrators, Data Protection Officers, Employees using AI assistants at work

Is it safe to put customer data into ChatGPT?

It can be, but "safe" is a property of your setup, not of the product name. The same model can be used in very different ways: an employee pasting a customer complaint into a personal consumer account, a company using a business workspace under a signed agreement, or an application calling a model API through a gateway that strips identifiers first. These carry very different risks.

Under the DPDP Act, three facts frame the answer:

  1. You are the Data Fiduciary. You decided to process the customer's data with an AI tool, for your purpose.
  2. The vendor is your Data Processor — if it processes the data only to serve your request and on your instructions.
  3. Section 8(1) keeps you responsible for compliance in respect of processing done on your behalf by a processor, irrespective of any agreement to the contrary.

So the question to ask is not "is ChatGPT DPDP compliant?" but "have we set up this processor relationship so that we can meet our own obligations?" That means a contract, appropriate security safeguards, clarity on the vendor's use of your data, retention you can live with, and logs.

This page does not describe any vendor's current settings. Terms, tiers and defaults change; verify them from the vendor's current documentation and your signed agreement.

The DPDP role analysis

ScenarioYour roleVendor roleKey risk
Employee uses a personal consumer account for customer dataData Fiduciary (the company still decided, or failed to prevent, the use)Unclear — no contract with you; vendor's own terms governNo processor contract, possible training use, no logs; may be an unauthorised disclosure
Company-managed business or enterprise workspace under a DPAData FiduciaryData ProcessorConfiguration, retention and access control
Your application calls a model API under API terms and a DPAData FiduciaryData ProcessorPrompt contents, logging, region, sub-processors
Vendor uses your customers' data to train its general modelsData FiduciaryActing for its own purpose for that use — fiduciary-likePurpose creep; your notice may not cover it
You provide an AI product to business customers using a model APIData Processor for your customers' data; Data Fiduciary for your ownYour sub-processorFlow-down terms, customer approvals

The fourth row is why "training on your data" matters so much. If customer data is used for the vendor's own model improvement, the vendor is no longer acting purely on your behalf, and your Rule 3 notice probably did not describe that purpose.

What to verify before allowing customer data

Check these from the vendor's current terms, trust documentation and your own signed agreement. Record the answers and the date checked.

  1. Tier. Is the usage on a business, enterprise or API tier with organisation-level controls, rather than individual consumer accounts?
  2. Training use. Is your data excluded from training the vendor's models by default, or only after an opt-out? Does that apply to all products you use, including file uploads and voice?
  3. Retention. How long are prompts, files and outputs retained by the vendor, including for abuse monitoring? Can you configure shorter retention, and does any exception apply?
  4. Data processing agreement. Is there a DPA covering processing on your instructions, confidentiality, security, breach notification, deletion, audit and sub-processors?
  5. Region. Where is data processed and stored? Is a region choice available? Check against any sector rules that apply to you — RBI payment data storage, for example.
  6. Sub-processors. Is there a published list, and will you be told of changes?
  7. Breach notice. How fast will the vendor tell you of an incident? You need enough detail to give the Board a report within 72 hours (Rule 7(2)(b)).
  8. Access and admin controls. SSO, role-based access, audit logs and the ability to disable features such as connectors or memory.
  9. Deletion. Can you delete workspaces, conversations and files, and receive confirmation?

Controls on your side of the boundary

Even with a good contract, the cheapest protection is sending less personal data.

Redact before you prompt

Put an AI gateway between staff or applications and the model. It should detect and replace personal data — names, phone numbers, emails, addresses, account and card numbers, Aadhaar-like and PAN-like patterns — with tokens, keeping the mapping in your own vault. Many tasks (summarising a complaint, drafting a reply, classifying intent) work as well on tokenised text. Rule 6(1)(a) names masking and virtual tokens as examples of appropriate data security measures.

Restrict and log

  • Allow only approved tools and tiers; block consumer AI domains on managed devices where policy requires.
  • Use SSO and role-based access so only staff who need customer data can send it.
  • Log each request: user, purpose, data classes detected, model, timestamp — enough for Rule 6(1)(c) visibility, retained for at least one year under Rule 8(3).

Scope features

Disable connectors, memory or file features you have not assessed. Each one is another data flow to map.

Purpose and notice

Confirm your Rule 3 notice describes AI-assisted processing for the relevant purpose, and that you have a lawful basis for it. See the AI pillar guide.

When staff paste customer data into unapproved tools

"Shadow AI" is the pattern most likely to cause a DPDP incident. An employee copies a customer's account details into a personal chatbot account to draft a reply. There is no processor contract, no company logging, and the vendor's consumer terms — not yours — apply.

Is that a personal data breach? The Act defines a breach to include unauthorised processing or accidental disclosure of personal data that compromises its confidentiality. Disclosure to a third party outside any authorised processor arrangement can fit that definition. If it does, Rule 7 requires intimation to each affected Data Principal and to the Board, with a detailed report within 72 hours. There is no materiality threshold.

Practical steps:

  1. Policy. Publish an AI acceptable-use policy that names approved tools and banned data types.
  2. Provide a good approved option. Staff use consumer tools when the sanctioned path is worse.
  3. Technical controls. Data-loss-prevention rules for clipboard and uploads to unapproved AI domains.
  4. Training. Short, scenario-based, repeated.
  5. Incident path. Make it easy and blame-free to report an accidental paste, so you can assess it quickly.

Sector rules that add to DPDP

Some organisations face extra layers beyond DPDP:

  • Banks, NBFCs and payment companies. RBI's 2018 payment data storage circular requires payment system data to be stored only in India; check whether any payment data could reach an AI vendor. RBI's FREE-AI committee report recommends board-approved AI policies, an AI inventory and customer disclosure; these bind only once RBI issues directions. See RBI FREE-AI.
  • SEBI-regulated entities. SEBI's amendment making regulated persons solely responsible for AI tool outputs and data security, regardless of vendor, applies here. See SEBI AI/ML rules.
  • All Indian service providers and body corporates. CERT-In Directions require specified incidents to be reported within 6 hours and ICT logs kept for 180 days within India. See CERT-In rules for AI systems.
  • Selling to EU customers. GDPR Article 28 processor terms and Chapter V transfer rules will apply to EU personal data you put into an AI tool. See selling AI SaaS to Europe.

Further reading on suhasbhairav.com

Before 13 May 2027

Compliance checklist

  1. Inventory current AI tool usage

    Survey teams and check network logs to find which AI tools and accounts handle customer data today.

    Section 8(1)

  2. Choose approved tiers and sign a DPA

    Move customer-data use to business, enterprise or API tiers under a data processing agreement.

    Section 8(2); Rule 6(1)(f)

  3. Verify training, retention and region terms

    Record the vendor's current answers on training use, retention periods, processing region and sub-processors, with the date checked.

    Rule 6(1)(f)

  4. Deploy redaction in an AI gateway

    Tokenise customer identifiers before prompts and keep the mapping vault inside your environment.

    Rule 6(1)(a)

  5. Enable SSO, roles and audit logs

    Restrict who can send customer data and keep request logs for at least one year.

    Rule 6(1)(b)–(c); Rule 8(3)

  6. Publish an AI acceptable-use policy

    Name approved tools, prohibited data types and how to report accidental disclosures.

    Rule 6(1)(g)

  7. Update the privacy notice

    Describe AI-assisted processing where customer data is used, with the purpose and data items.

    Rule 3

  8. Add AI disclosure to the breach runbook

    Define how to assess shadow-AI pastes and vendor incidents against Rule 7 and CERT-In timelines.

    Rule 7

Key dates

  1. DPDP Rules notified

  2. Processor contract, safeguards and breach duties apply to AI tool use

Answers

Frequently asked questions

Is ChatGPT DPDP compliant?

No tool is DPDP compliant on its own. Compliance depends on how you use it. Under the DPDP Act you are the Data Fiduciary and the vendor is your Data Processor, so you need a contract, appropriate safeguards, clarity on training use and retention, and logs. Section 8(1) keeps you responsible.

Can employees use ChatGPT with customer data in India?

Only through a company-approved setup with a processor contract, appropriate safeguards and logging, and only where your notice and lawful basis cover the purpose. Staff pasting customer data into personal consumer accounts risks unauthorised disclosure, which can be a reportable personal data breach under Rule 7.

Does OpenAI train on my company's data?

Check the vendor's current terms for the specific tier and product you use, since training defaults and opt-outs differ between consumer, business and API offerings and can change. Record the answer and the date you checked, and put the commitment in your data processing agreement where possible.

Who is responsible if an AI vendor leaks my customers' data?

You are, as the Data Fiduciary. Section 8(1) of the DPDP Act makes you responsible for processing done on your behalf by a Data Processor, irrespective of contract terms. You must notify the Board and affected customers under Rule 7, so require fast incident notice from the vendor.

Does customer data sent to ChatGPT leave India?

It may, depending on the vendor's processing locations and any region option. The DPDP Act allows transfers except to countries the Government restricts by notification, subject to Rule 15. Sector rules such as RBI payment data storage can be stricter, so check region terms before sending regulated data.

Should I redact personal data before using an LLM?

Yes, wherever the task allows. Replacing names, phone numbers, account numbers and similar identifiers with tokens before the prompt reduces exposure, and Rule 6(1)(a) lists masking and virtual tokens as security measures. Keep the token mapping inside your own environment.

Legal references and sources

  • Digital Personal Data Protection Act, 2023 — Section 8(1), 8(2) (official text)
  • Digital Personal Data Protection Rules, 2025 — Rule 6(1)(a), (c), (f) (official text)
  • Digital Personal Data Protection Rules, 2025 — Rule 7 (official text)
  • Digital Personal Data Protection Rules, 2025 — Rule 8(3) (official text)

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

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.