Agentic commerce: AI that shops and pays for you within set limits

It’s a busy Tuesday evening. You’re stuck in a late work call and suddenly realise that you’re out of milk, fruits, and bread. Instead of opening an app, searching, adding items to cart, applying coupons, and making a payment… you simply tell all this to an artificial intelligence (AI) assistant. 

Within seconds, the order gets placed and paid for automatically, saving you all the hassle. 

That’s the future of agentic commerce — where AI not only shops for you, but also pays for you within limits you set. 

As digital commerce shifts to agentic rails, enabling agentic payments closes the loop — creating a truly end-to-end, autonomous shopping experience. 

Recently, Tata-backed digital commerce platform BigBasket, along with OpenAI, the National Payments Corporation of India (NPCI), and fintech Razorpay have partnered to enable this end-to-end flow, in one of the earliest use cases to implement the agentic experience at scale.  

While still under pilot stage, users can now tell AI agents like ChatGPT what they want, get personalised orders, and complete payments via UPI Reserve Pay — securely pre-authorising the order within set spending limits. 

“The payment flow needs to be fully ironed out. Some of the use cases we’re anchoring on, like reordering usual items, require training and updating the model to ensure a seamless, low-edit customer experience. Building, testing, and refining those models takes time,” said Preeti Jain, head of product and design, BigBasket.  

The focus, Jain added, was on small ticket items to gauge how users interact with agentic shopping and payments experience.

The secret sauce of a refined agentic payment experience hinges on NPCI’s latest launch in the form of UPI Reserve Pay. 

A user-defined amount is first blocked for a specific merchant, such as BigBasket. When a payment is triggered on ChatGPT via UPI Reserve Pay, the amount is debited from those pre-blocked funds. 

For NPCI, the convergence of AI and UPI is to make transactions more intuitive, intelligent and inclusive. 

During the launch of this pilot at the recently-concluded Global Fintech Fest, Sohini Rajola, ED-Growth, NPCI, said, “Agentic Payments mark an important step in India’s digital payments journey. By enabling user-authorised AI agents to initiate secure payments, we are moving closer to a future where technology anticipates needs and simplifies experiences.” 

‘Model context protocol and integrations’   Let’s say a user orders a fresh fruit bowl through an AI agent. The agent checks for availability, matches inventory, and confirms the order on the digital commerce platform.

 

Once confirmed, a transaction request is sent to the payment partner — such as Razorpay or Cashfree Payments — which then uses UPI Reserve Pay to process the payment. This flow is enabled through the payment partner’s model context protocol (MCP), linking the AI agent with the payment system. 

“Think of it like, while everyone interacted with application programming interfaces (APIs) initially, in the AI world, you would interact with MCPs so that the model can understand or access whatever functionalities one (for example, a merchant) may have,” said Reeju Datta, co-founder, Cashfree Payments. 

 

The MCP is like a “universal connector” — in other words like a USB-C port — that allows different software systems to work together easily.  This enables AI agents and assistants to interface directly with core APIs, streamlining integration for merchants of all sizes.  

Without it, businesses would require manual and complex integration of compatible AI systems with every element of their fintech infrastructure, which includes applications such as payments, verification, payouts, among others. 

“Ideally, if we are able to have integrated play then that simplifies our development cycle as well as the future maintenance,” said Jain from BigBasket.  She explained that costs to implement agentic interfaces are not prohibitive and that the industry at large is still in trial mode with the new technology.  

Just as companies today account for infrastructure costs like cloud services, they may soon factor in LLM-related expenses as part of their overall service or infrastructure costs. 

“It’s part of the infrastructure, like AWS, now LLM costs will be included too, which are already there anyway since companies are working on AI systems,” Datta said.  

Evolving use cases   Making the experience agentic for users can reduce drop offs on the back of data points such as browsing history, preferred payment methods, and even potential offers for a customer. 

 

“Once you have insights like browsing journeys, order history, and preferred payment methods, there’s a lot a shopping agent can do to improve conversions,” said Lavika Aggarwal, group product manager at Cashfree. 

Jain added that AI agents capable of analysing handwritten notes or generating recipes based on available ingredients represent emerging use cases where a fully agentic, end-to-end experience can make a significant impact. 

“The bet is that all e-commerce companies will essentially have their own AI agents within their websites where the commerce happens. Additional things like personalisation, recommendations, all of that can be done by that shopping agent,” Datta added.  

This may again encourage tech service providers (TSPs) to build AI-first infrastructure for other companies to plug-and-play, which today is being built by e-commerce firms themselves.  What remains to be seen, however, is the adoption of agentic use cases by customers. 

The security aspect   When it comes to security of transactions, Datta explained that the core payment processing remains unchanged, and guided by all existing compliance requirements. 

 

“It will not happen without your consent. Even if you look at Reserve Pay, you are specifying block limits, which you can change or cancel any time you want,” he said.  

While an autonomous payments experience removes the friction of browsing and manual authorisation for each transaction, experts said it must also reassure users about payment safety as the technology continues to evolve. 

“Managing customer anxiety by providing clear, realistic updates is key. Informing customers on a timely basis about reversing transaction amounts in case of failed payments and an update by when that can happen is important,” Jain said.  

Experts added that with the right user-defined intent captured by the model, purchases are executed only after the user confirms the transaction, minimising concerns of hallucinated transactions.  

“LLMs by design are probabilistic in nature, but when the protocols are coming, they’re being cryptographically signed for intent specifically. If a user’s request is vague, the agent prompts for clarification, turning it into straightforward intent-based prompts,” Aggarwal explained.