AI agents may soon shop and pay. Are brands ready?

AI agents can do a lot of things, possess a growing number of capabilities, and we are repeatedly assured they evolve every day. 

No one is denying that, least of all Reels, Shorts and future-forward newsletters promising that you (and I) too can be millionaires sitting at home, running multi-million enterprises “alone” through the efforts of legions of AI agents (or agentic AI if you want to be fancy) building product, running operations, crunching numbers, doing things, and watching profits presumably prosper. And you can sit back, enjoy the show, and then, in your own richly remunerated free time, start your own newsletter or your tube channel or Insta-fame account.

While we buy into that dream, India’s payments ecosystem may soon offer technology companies a more immediate one: allowing AI agents to help consumers choose purchases and then pay for them through UPI.

Not unrestricted access to our money, at least not under the reported proposal. But something still consequential: delegated authority to move it, within limits, after an algorithm has helped decide where it should go.

AI assistants can already search, compare and recommend. Payment is the point at which most still have to hand control back to a human. NPCI’s proposed Unified Agent Protocol could begin removing that handover.

According to a recent report, the National Payments Corporation of India is developing a common protocol through which verified AI agents could initiate UPI transactions within limits set by users. The system would reportedly create a registry of trusted agents and retain transaction logs for audits and disputes.

The infrastructure is arriving as agents move rapidly from demonstration to deployment. McKinsey’s latest global survey found that 62% of organisations were already experimenting with or scaling AI agents, including 23% that had begun scaling them in at least one business function.

India may be particularly receptive to the next stage. Adobe’s 2026 AI and Digital Trends research found that 60% of Indian consumers were interested in creating a personal AI agent, 55% would interact with an agent offered by a brand, and 58% were comfortable with agents interacting with one another.

Yet deployment remains shallow: in no individual function had more than 10% of respondents scaled agentic systems, a statistic common among studies, suggesting that the technology is advancing faster than the operating models and controls surrounding it.

Also Read: Google-Hindware ruling could push brands from rented intent to owned demand

The protocol remains under development, and NPCI has not publicly released detailed technical specifications. But its reported design points towards a future in which a consumer could ask an AI assistant to find, compare and purchase a product without navigating a marketplace or completing the payment themselves.

That may solve one trust problem: whether the software moving money is genuine, authorised and operating within the consumer’s mandate. It may not solve another: why the agent chose what it bought.

Whose game?

“Conventional attribution dies the moment the click disappears,” said veteran marketer Shubhranshu Singh. “Last click logic assumed a human journey with visible touchpoints. Whereas an agent compresses discovery, comparison and purchase into a single invisible transaction.”

Also Read: When AI does the shopping, who sees the advertising?

That leaves several participants with plausible claims over the same conversion. The AI platform executed the decision. The marketplace supplied the catalogue. A publisher’s content may have shaped the shortlist. The brand may have influenced the consumer’s original request.

“Everyone will claim that sale,” Singh said. “Credit will migrate to whoever can prove influence over the agent’s consideration set. The industry must shift from measuring clicks to measuring presence and ask: was the brand in the shortlist that the agent assembled, and why?”

The conflict is obvious.

“The irony is that the referee is an algorithm owned by one of the claimants,” he added.

This means NPCI’s proposed trust layer could provide a relatively neutral transaction record. But a transaction log is not necessarily a record of commercial influence. Unless the wider ecosystem preserves the sources consulted and factors used by the agent, marketers may know that a sale occurred without knowing what contributed to it.

The shift could also change what it means for a brand to be discoverable.

“Agents do not browse shelves; they resolve constraints,” said Siddhant Sethi, AI professional and specialist at White Rivers Media. “So, the real optimisation target will not be consumer attention, but the agent’s decision logic.”

Sellers, he said, will work backwards from what the agent can parse and compare: identifiers, landed price, live inventory, delivery certainty, returns, compatibility and warranty. “The new shelf is not visual. It is a schema.”

Brands may therefore begin optimising not simply for attention, but for machine-readable information. Product data, inventory, delivery certainty and service records could become as commercially important as conventional creative.

That would not necessarily make brand building irrelevant. It may split advertising’s audience between the human who frames the request and the machine that executes it.

“Agentic commerce splits advertising’s audience in two,” Singh said. “The human maybe still writes the brief but the machine executes it. The paradox is that with agentic commerce, brand building becomes more valuable precisely when brand advertising becomes harder to deliver”.

Gaming the system

Any system that rewards machine-readable signals also creates incentives to manipulate them.

If an agent rewards low listed prices, sellers may shift costs into fees it reads poorly. If it rewards speed, delivery estimates may become aggressively optimistic. If it seeks exact matches, sellers may split catalogues into hyper-specific variants to occupy more of its consideration set.

“This is not classic advertising. It is choice architecture embedded in data,” Sethi said. “In agentic commerce, what gets parsed gets purchased.”

That expands the meaning of brand safety.

Also Read: AI is rewriting Indian marketing's OS. Is the industry ready?

In conventional digital advertising, brand safety concerns the content beside which an advertisement appears. In agentic commerce, the central risk may be whether the information used to make the purchasing decision is trustworthy.

Fake reviews could manufacture approval. Misleading structured data could make products appear to meet requirements they do not. Counterfeit listings could borrow the reputation of established brands. Sponsored placements could be presented as independent recommendations.

External webpages may also contain instructions intended to influence an autonomous agent, commonly described as prompt injection. Whether such attempts succeed would depend on the agent’s safeguards.

As currently reported, the proposed protocol would verify the agent and payment request, while NPCI may not necessarily see the underlying product. That may establish whether the payment was authorised, but not whether the recommendation was trustworthy.

“Authentication can prove that the right agent paid; it cannot prove that the product deserved to be chosen,” Sethi said. “A valid payment can still be a manipulated purchase.”

Platforms would therefore require what he called a “chain of custody for commercial truth”, showing who supplied each claim and whether it was supported by other sources and post-purchase outcomes. “Brand safety here is no longer adjacency. It is decision integrity.”

When that integrity fails, liability may depend on where the failure occurred.

Blame game?

“The mere fact that a user authorised an AI agent to make a payment should not amount to an acceptance of every decision taken by that agent,” said Ankit Sahni, partner at Ajay Sahni & Associates.

If the agent exceeds the user’s mandate, misunderstands an instruction or fails to apply promised safeguards, the AI provider may face liability. If the product is counterfeit or the listing misleading, responsibility may lie with the merchant and potentially the marketplace.

The payment intermediary’s responsibility, Sahni said, should ordinarily remain confined to failures in the payment layer, such as executing a transaction beyond the mandate or authenticating an invalid instruction.

The user may bear some responsibility where the transaction remained within a clearly defined mandate. But, Sahni cautioned, “authorisation to spend cannot become a blanket waiver of consumer rights”.

The emergence of agentic payments requires a distinction between payment authorisation and purchase accountability, said Anirban Mohapatra, partner at Cyril Amarchand Mangaldas.

Marketplaces already have obligations concerning counterfeit listings and misleading merchant information. “As agentic commerce evolves,” said Mohapatra, “the legal framework may need clearer transparency requirements around the agent’s instructions, information sources and reasons for choosing a product.”

Indian law does not presently provide a general right to an explanation for every AI decision, Sahni noted. Existing consumer-protection principles could nevertheless require disclosure where a recommendation is sponsored or commercially influenced.

“An AI agent should not be permitted to present an advertisement or preferential placement as though it were an independent recommendation,” he said.

Ad models behind AI wars: How Google, Meta, OpenAI are approaching advertisingAlso Read:

Mohapatra similarly argued that verifying the agent’s authority answers only one part of the equation, saying, “It does not explain why a particular product was selected, whether the recommendation was influenced by commercial arrangements, or what information the agent relied upon in reaching its decision.”

Regulatory attention may therefore move beyond payment authentication towards the decision-making layer, including transparency around sponsored recommendations, commercial affiliations and the factors that influenced a purchase.

India’s proposed system may eventually create a clear audit trail showing who authorised an AI agent, how much it could spend and where the money went.

For advertisers, consumers and courts, the harder trail may be the one preceding the transaction: what the agent saw, what influenced it and why it chose one product over another. Maybe they could use a newsletter.