Pasting a customer's complaint email into a chatbot to draft a reply feels like a private moment between you and a tool. Under the DPDP Act, it is a disclosure of personal data to a third party, and that carries real consequences for whoever made the call to paste it. Shadow AI is currently the largest blind spot in corporate data protection.

Prompts as personal data disclosures

When a generative AI tool is contracted by your organisation, it is typically acting as a Data Processor on your behalf. Pasting personal data into it for a purpose the original notice and consent under Section 6 never covered raises the same purpose-creep risk as repurposing a dataset for model training. The convenience of a chat window does not exempt it from the basic question of what basis you have to process that data in the first place.

The Data Processor trap

Many consumer-grade generative AI tools retain prompts to improve their own models by default, and a number of them run on infrastructure outside India. Section 8 keeps a Data Fiduciary responsible for its Data Processor's conduct, so an organisation that hands customer data to a tool without checking its retention and training settings carries that risk regardless of where the underlying server sits. Section 16 permits this kind of cross-border flow by default, but the responsibility for the data's security travels with it.

Securing the generative workflow

A workable policy does not have to be a total ban on AI tools, but it requires strict operational guardrails.

  • Classify what counts as personal data before anyone pastes it into a tool: names, identifiers, health or financial detail.
  • Default to enterprise tiers that contractually exclude prompts from training data, over free consumer tiers that do not.
  • Redact or tokenise identifiers where the task genuinely does not need them.
  • Put generative AI vendors through the same Data Processor due diligence you would apply to any other cloud vendor.

The risk in generative AI tools is rarely the model itself. It is the personal data an employee feeds it without thinking about where that data was supposed to stop.