Marketing teams across India are pasting customer data into AI tools to personalise copy. Very few have thought about what that does under DPDP, and the answer is uncomfortable.
The mechanism
When you send customer information to an external model to generate personalised content, you are disclosing personal data to a third party who processes it on your behalf. That third party is a Data Processor. Everything DPDP says about processors applies: a written contract, purpose limitation, security obligations, deletion terms.
Most teams using consumer AI tools for this have none of that. They have a subscription and a terms of service they did not read, which is not the same as a processing agreement.
Purpose limitation is the sharper issue
Consent under DPDP is tied to a stated purpose. Someone who agreed to receive product updates consented to product updates. They did not consent to their purchase history being sent to an external model to generate a personalised pitch.
You can argue that personalising a communication they consented to is within scope. That argument gets weaker the more data you send and the further the processing sits from what was described in the notice. It is weakest of all when data crosses borders to a processor the person was never told about.
Three practical positions
The conservative position is to never send identifiable customer data to an external model. Generate templates, personalise locally with merge fields. This is what most regulated Indian firms are settling on and it costs you very little in output quality.
The middle position is to pseudonymise before sending. Strip names, numbers and identifiers, send the behavioural signal, merge identity back in locally. Workable, more engineering than most teams expect.
The permissive position is to send real data under a proper processing agreement with an enterprise provider, with the purpose named in your notice. Defensible, but it means updating your notice and getting fresh consent for the population you collected under the old one.
What to do this month
Find out what is actually being pasted into AI tools. Ask the team, without blame, what they paste in a typical week. The answer is usually broader than management assumed. Then pick one of the three positions above and write it down as policy, because the current default of no policy is the worst of the three.
General information, not legal advice. Rules current as of July 2026.