When an AI Agent Clicks 'I Agree': Consent, Mandates and DPDP

Introduction

AI agents have moved from answering questions to running errands. Given a goal and a budget, they compare fares, fill passenger forms, tick boxes and reach the payment page. Each form asks a question the law has always assumed a human would answer: do you agree?

The Digital Personal Data Protection Act, 2023 (the DPDP Act) builds consent around the individual. A data fiduciary decides why and how personal data is processed; a data principal is the individual the data is about; a data processor acts for a fiduciary. The consent provisions commence on 13 May 2027 under G.S.R. 843(E). This article asks whether an agent's click can be her consent, what payment rules add, and how to build an agent whose bookings survive a dispute.

The legal position

Consent-manager registration under Rule 4 of the DPDP Rules, 2025 opens on 13 November 2026 but cannot yet work: no members of the Data Protection Board (the Board) had been appointed as of the latest official statement (12 August 2026). MeitY's voluntary India AI Governance Guidelines (5 November 2025) conclude that "a separate law to regulate AI is not needed given the current assessment of risks", but ask for audit trails and human checks at critical points for autonomous agents and accountability assigned by function, and list the role of consent managers in AI workflows as an open question.

Analysis

Contract can run on software; DPDP consent cannot

Contract law copes with agents. Section 11(c) of the IT Act attributes an automated system's messages to the person who set it running, and s.226 of the Contract Act enforces contracts made through an agent "as if" the principal had made them in person. An AI system is not "a person" under s.182, so in our view it is the user's instrument under s.11(c), though its provider may in some designs be her agent. Either way, a booking her agent makes will usually bind her.

DPDP consent is a different instrument, as Trayambic's article on notices and consent explains. Its test is personal: s.6(10) makes the fiduciary prove that consent was given "by such Data Principal". The Act names who may stand in her place: a parent or lawful guardian (s.2(j)), a registered consent manager accountable to her (s.6(7), 6(8)), and a nominee on her death or incapacity (s.14). A general agent is not among them.

Can consent be delegated?

In our view the s.6(1) question is authorship: did the data principal herself see this purpose and choose it?

• As a messenger, yes. Where the user has seen a purpose and approved it, an agent that carries her approval is a channel, like a browser. Section 11(c) supports attributing that message to her.
• As an author, no. Where the agent accepts a purpose the user never saw, such as a hotel's partner marketing, the consent is neither "informed" nor "specific" as to her. A standing instruction to "accept whatever is needed" cannot be a clear affirmative action for purposes not yet known.
• Often consent is not the ground. Passenger details given to complete a booking are data the user "voluntarily provided" for a specified purpose under s.7(a). In our view data she directs her agent to supply is supplied by her, so fulfilment rarely needs consent.

So the design has two tiers: the agent completes what the instruction covers and returns every optional purpose to the user.

Notices read by a machine

Section 5(1) requires the notice to be given "to the Data Principal", and Rule 3 requires it to be understandable independently of other information. In our view a notice an agent has parsed through an interface has not been given to her. Fiduciaries should publish notices an agent can pass on intact: itemised data, purposes, withdrawal link, language options. Agent builders should show the notice before relaying any consent and keep the version shown.

Payments keep a human checkpoint

From 1 April 2026 RBI's Authentication Directions require at least two distinct factors for digital payments, one dynamic and unique to the transaction unless the card is present (para 6). Issuers must follow the DPDP Act (para 9(3)) and compensate the customer "in full without demur" for losses from transactions made without complying (para 9(2)). The exemptions include later debits under an e-mandate and travel bookings on commercial or corporate cards, including GDS or IATA bookings. Cross-border payments are outside them, though card issuers needed a risk-based check on cross-border card-not-present payments by 1 October 2026 (para 10).

Nor is an e-mandate a general licence. RBI's Digital Payments E-mandate Framework, 2026 (21 April 2026) covers recurring transactions on cards, prepaid instruments and UPI. The customer registers once with an additional factor, the first debit needs one too, the issuer gives at least 24 hours' notice before each charge, and she can withdraw at any time (paras 4 to 6). Later debits may skip the additional factor up to ₹15,000, or ₹1,00,000 for mutual fund subscriptions, insurance premiums and credit card bills (para 8). In our view an agent that collects the user's one-time passwords to pay alone defeats the second factor.

Withdrawal, children and errors

Withdrawal. Withdrawal must be as easy as giving consent (s.6(4)), and the fiduciary must stop processing within a reasonable time (s.6(6)). In our view a merchant should honour a withdrawal its customer's agent relays, since a mistaken withdrawal costs little. Revoking the agent's own authority is a separate act that leaves consents given to merchants in place, so the agent should offer to send those withdrawals on.

Children. A minor cannot employ an agent (Contract Act s.183). For anyone under 18 the data principal includes the parent (s.2(j)), whose verifiable consent is needed (s.9(1); Rule 10). An agent that learns a teenager's habits risks the bar on tracking and behavioural monitoring of children in s.9(3), so in our view under-18 profiles should run without long-term memory.

Errors. A fiduciary must keep personal data complete, accurate and consistent where it is likely to be disclosed to another fiduciary (s.8(3)), which is the agent's whole job. A mistyped passport number is therefore a DPDP problem for the agent provider. Against the merchant, the user is likely bound: s.237 of the Contract Act holds a principal to an agent's unauthorised acts where her conduct led the third party to believe they were authorised. Each fiduciary stays responsible "irrespective of any agreement to the contrary" (s.8(1)), so terms of service cannot move DPDP duties, and data sent to the wrong merchant is, in our view, an accidental disclosure and so a personal data breach (s.2(u)).

The log is the evidence

In a dispute, the agent's log is the only record of what the user saw and approved. Section 6(10) puts the burden of proof on the fiduciary, Rule 6 requires logs to be kept for at least a year, and the AI Governance Guidelines ask for audit trails. Log the instruction, the plan shown, each notice version, each approval, each payment authentication and what went to whom.

Practical implications

Worked example (figures illustrative). An Indian AI start-up's travel assistant has two lakh users. One asks it to book Mumbai to Singapore for four nights, a hotel under ₹15,000 a night and travel insurance.

1. 1.Plan card. Before anything is sent, the agent shows each merchant, the data it will receive and any optional consents, all off by default.
2. 2.Airline. Name, passport and contact details go on her instruction under s.7(a). Passing passenger data to immigration is its own duty under rule 22 of the Immigration and Foreigners Rules, 2025. Its marketing consent is shown word for word; she declines.
3. 3.Hotel. The Singapore hotel prices in rupees for guests from India, so in our view s.3(b) brings its processing within the Act. It receives name, dates and contact details; the passport is shown at check-in.
4. 4.Insurer. The agent pre-fills identity and trip details but hands the health questions and the purchase confirmation to her. The answers are hers to vouch for, and rule 4(9) of the Consumer Protection (E-Commerce) Rules, 2020 requires an explicit, affirmative action before consent to a purchase is recorded.
5. 5.Payments. She authenticates the airline and insurer payments in her bank app; the hotel's cross-border card payment goes through her issuer's risk check. The agent never sees a one-time password.
6. 6.An error. The passport expiry date is sent wrongly. The log shows the agent read an outdated scan, so the start-up corrects the booking, bears the change fee and adds a validation check.
7. 7.Exit. When she closes her account, the passport vault is erased and withdrawals go to each merchant holding an optional consent.

Key takeaways

Contract law will usually bind a user to what her agent books; the DPDP Act will not let the agent author her consents. Build agents that deliver the user's choices instead of making them, leave the payment factor in her hands and keep the record that proves both.

• Under IT Act s.11(c) and Contract Act s.226, a booking made by a user's AI agent will usually bind her.
• DPDP consent must be given by the data principal with a clear affirmative action (s.6(1)); only parents and guardians, registered consent managers and nominees may act for her, so in our view an agent can deliver her consent but cannot author it.
• Fulfilment data the user directs her agent to send usually rests on s.7(a); every optional purpose should go back to her.
• RBI's Authentication Directions (from 1 April 2026) and E-mandate Framework, 2026 keep a dynamic second factor with the customer; e-mandates cover recurring payments only.
• The agent provider must keep data accurate when passing it to merchants (s.8(3)) and must be able to prove consent (s.6(10)), so its logs are its evidence.