Imagine asking an AI assistant to reorder packaging for a small Indian online seller. It finds the right size, but the cheapest option requires twice the quantity. Can it spend more, switch suppliers or enrol the business in a monthly delivery? A request to find a good deal leaves several purchasing decisions unanswered.
Walmart’s October 21 announcement makes those decisions timely. It plans Agent Profiles within the next month, separating agent purchases within a customer’s account. Customers would approve each purchase or set a spending limit, and could withdraw access. Its Agent Gateway is an early, invitation-only beta. These are announced arrangements, not evidence that every customer can use them today.
What has the shopper actually approved?
Walmart’s terms distinguish two permissions: the customer’s authority for an agent to act and Walmart’s permission for that agent to use its platform. One cannot replace the other. The terms treat authorised actions within the customer’s granted scope as the customer’s own actions and require monitoring and prompt reporting of suspected overreach.
That makes the wording of an instruction commercially important. In our hypothetical packaging order, “buy 100 envelopes for no more than ₹800 including delivery, once” defines a much narrower task than “keep packaging stocked”. Neither instruction, on its own, proves that a software system will enforce it.
For this seller, an approval screen should make the quantity, final charge, delivery date and any recurring commitment visible together. A change from 100 to 200 envelopes should be presented as a changed order, even if the price per envelope falls. A cheaper unit price can still tie up more cash and storage space.
Connect the instruction to the charge
Google’s March 18, 2026 developer guide describes a useful payment-design distinction. Under Agent Payments Protocol, or AP2, an intent mandate records purchasing constraints; a payment mandate binds authorisation to a particular cart and amount; a receipt completes the record. The guide illustrates a design, including a simulated signature. It is not an independent test of production security. It does not establish widespread adoption.
Applied to our example, the useful record would let the seller compare the approved envelope order with what was charged. An order history reveals what happened afterwards. An approval boundary determines what may happen beforehand. Both deserve attention when choosing an assistant.
The India connection
Delegated spending already has a human counterpart in India. In November 2025, NPCI BHIM Services announced UPI Circle Full Delegation, allowing trusted contacts to pay within a monthly limit of up to ₹15,000. That was human delegation, not an AI shopping launch. It offers a comparison about bounded authority, not evidence of a Walmart or Flipkart rollout in India.
For an Indian merchant considering agent-led orders, the practical opportunity is to serve a customer who wants routine buying handled with less effort. The operational test is whether the merchant can show the exact offer clearly enough for the customer’s purchasing rules to be applied. Delivery charges, pack sizes and repeat-order terms can decide whether an apparently suitable offer actually fits.
The business case should therefore count the work left behind: clarification requests, changed orders and customer support. Faster order placement alone cannot establish that a merchant earns more or that a buyer saves money. Those outcomes would need evidence from actual use.
The Crunch
Before handing over a recurring purchase, settle one concrete question: which changes must come back to the buyer? In our packaging example, that answer covers quantity, total cost and future deliveries. The useful assistant is one whose purchasing authority the customer can understand before the order is placed.