AI agents for operations leaders in e-commerce in India

Cash on delivery drives much of the work AI agents can do for e-commerce operations in India: confirming risky orders before dispatch, cutting return-to-origin losses, cleaning addresses, checking sellers and replying in several languages. Seller suspensions, fraud blocks, refund disputes and grievances stay with people. The DPDP Act 2023, e-commerce consumer rules and RBI payment rules apply.

What takes an Indian e-commerce operations leader's time?

Cash on delivery and its consequences. A large share of orders are still paid in cash at the door, and a meaningful number are refused on arrival or never reach the customer. Every such parcel travels back to the warehouse as return-to-origin, costing two shipping legs and tying up stock. Operations spends heavily on confirming intent before dispatch and on chasing the reasons afterwards.

Addresses. Addresses are often descriptive, with landmarks instead of street numbers, and the same locality is spelled several ways. A wrong pin code sends a parcel to the wrong hub. Cleaning and confirming addresses before dispatch is a daily job.

Sale events and sellers. Festive sales compress weeks of volume into days. Marketplaces onboard thousands of sellers whose tax registration, bank account and listings must be checked, and who generate their own stream of payout and dispute queries.

Where can AI agents help?

• COD order confirmation (depth 3). The agent contacts the customer through an approved channel, in their chosen language, to confirm a high-risk cash order before dispatch, and records the response. Orders that cannot be confirmed are held, not cancelled.
• Address cleaning (depth 2 and 3). The agent standardises addresses, checks the pin code against the locality, and corrects obvious errors within rules. Ambiguous corrections are drafted for confirmation by the customer or a person.
• Return-to-origin analysis (depth 1 and 2). The agent reads courier remarks, call logs and order history to classify why each parcel came back and drafts a weekly summary by courier, pin code and seller.
• Seller onboarding checks (depth 2). The agent compares the seller's tax registration, business name, bank account holder name and listings, and drafts a pass or a list of issues. A marketplace manager approves.
• Customer replies in regional languages (depth 2 and 3). Order status answers go out at depth 3 from approved templates; anything else is drafted.

Sale events are where held items pile up. An agent that holds uncertain cases is behaving correctly, but on a peak day the held queue can outrun the team. Plan the staffing and the triage rules before the event; our guide on handling a spike in an AI agent's held items sets out how.

What stays with people, and why?

Seller suspensions and delistings stay human. They affect a business's income, and the marketplace's terms usually require a reasoned decision and a route to appeal.

Fraud blocks and customer bans stay human. A pattern of refused COD orders may be fraud, a courier problem or a customer who moved. A person decides.

Refund disputes and grievances stay with people. Indian consumer protection rules expect e-commerce entities to have a grievance officer who acknowledges and resolves complaints within set times. An agent can draft and track; the officer answers.

Which rules apply in India?

The Digital Personal Data Protection Act 2023 governs customer and seller data. Processing needs consent or another permitted ground, notice must explain the purpose, data must be protected and deleted when no longer needed, and the business as data fiduciary is responsible for its processors, including the agent's supplier. Customer phone numbers collected for delivery should not be repurposed for marketing without a valid basis.

Consumer protection rules for e-commerce require marketplaces and inventory sellers to display seller details, handle grievances through a named officer, avoid manipulative pricing and not impose cancellation charges unfairly. Agents writing customer communications must stay inside those duties.

The RBI regulates payment systems and payment aggregators. In substance, its rules stop merchants from storing customers' full card details and require tokenisation; an agent should never see card numbers. Refunds to cards and account-to-account payments follow the payment provider's rules. Card data handling must meet PCI DSS v4.0. India has no AI-specific law.

This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.

How do you start in the first 90 days?

Weeks 1 to 4: start with return-to-origin analysis at depth 1. It reads data you already have, changes nothing, and tells you where the losses are by courier and pin code. Use it to pick the first live capability.

Weeks 5 to 8: shadow COD confirmation for one region. The agent prepares and records the contact; the existing team makes it. Compare refusal rates for confirmed and unconfirmed orders.

Weeks 9 to 13: certify COD confirmation at depth 3 for that region and languages, and start shadowing seller checks. Before the next sale event, set the held-item triage rules. Follow deploying AI agents: train, shadow, certify, widen. When a capability is no longer needed, for instance after a payment change reduces COD, retire the agent properly rather than leaving it running. For a contrast in returns culture, see our brief on operations agents in German e-commerce.

Where Pryme fits

In Pryme, the courier's delivery scan arrives as signed evidence and a customer's message saying the parcel never came is declared. The return-to-origin agent shows which it relied on for each classification. A seller's uploaded bank statement is external until the bank account is verified, so payouts to a new account wait for that check.