How Indian D2C brands use AI agents, and what to keep human

Can an agent answer product questions at 2 AM?

The job: “Does it last all day?” “Is it okay for oily skin?” “When will it reach Lucknow?” These questions arrive late at night, and the same 30 come back again and again.

Some brands do not even give the question a place to land. When we opened 100 Indian D2C homepages on a phone on 19 October 2026, 17 had no WhatsApp link, no chat window and no signup field.

How an agent helps: you write a facts sheet once. Notes for each product, delivery days by region, return rules, COD charges. The agent answers from that sheet and nothing else. It says it is an AI assistant. When a question is not on the sheet, it says so and tells the buyer when a person will reply.

Keep human: anything about health, allergies or skin reactions. A wrong answer there is not a lost sale. It is a hurt customer.

Copy this (instruction for the agent):

“Answer from the facts sheet below and nothing else. Say you are an AI assistant in your opening line. If the answer is not on the sheet, say a person will reply by [time]. Never guess about skin, health or allergies.”

Copy this (reply the buyer sees):

“Hi! I am the [Brand] assistant, an AI. [Product] lasts about [hours from your sheet] on skin. Want me to share the travel size so you can try it?”

Up next That answer brings the order in. The next job decides whether the parcel comes back to you.

Should an agent confirm cash on delivery orders?

The job: a COD order arrives with half an address and a number that may be wrong. You ship it, it comes back, and you pay the courier both ways.

The numbers are not small. Unicommerce’s India D2C Report 2026, built on 410M+ shipments across 6,000+ brands, puts return to origin at 39.2% at the festive peak. The brands that brought it down made a few changes together, and one of them was checking the address before dispatch.

How an agent helps: it messages every COD buyer soon after the order, reads back the item, the amount and the pin code, and asks for one reply. A clear yes goes to packing. A changed address or a vague answer goes to a person.

Keep human: cancelling an order or marking a buyer as fake. The agent flags it. A person decides.

Copy this:

“Hi [name], this is the [Brand] order assistant, an AI. You ordered [product] for ₹[amount], cash on delivery, to pin code [pin]. Reply 1 to confirm or 2 to change the address.”

Up next A confirmed order is half the story. Seven in ten carts never reach the order button.

How should an agent handle abandoned carts?

The job: Baymard Institute averages 50 studies and puts the cart abandonment rate at 70.22%. Seven in ten carts are left behind.

Many brands answer with a discount blast. That teaches buyers to wait for the coupon.

How an agent helps: one reminder, a few hours later, that answers the doubt buyers usually have at that step (size, delivery days, COD, returns) and offers help instead of a discount. If the buyer replies, the agent answers from the facts sheet.

Keep human: discounts. Any coupon, offer or price change waits for your yes. Our own THA Agent Desk runs on the same rule: every price waits for a person.

Copy this:

“Hi [name], you left [product] in your cart. Most buyers ask about [size, fit or delivery] at this step, so here it is: [one line answer]. Need help choosing? Just reply here.”

Up next The sale is not the end. One message after delivery decides whether she buys again.

Can an agent ask for reviews and reorders?

The job: after delivery, nobody asks how it went. Happy buyers forget to leave a review. Unhappy ones go to your Instagram comments.

How an agent helps: a few days after delivery, one short message asks how it is going. A happy reply gets the review link. When a pack is likely to run out, based on its size, one refill reminder goes out.

Keep human: every unhappy reply. The handoff is what makes people willing to use the agent at all. Salesforce found that 45% of consumers are more likely to use an AI agent when there is a clear way to reach a person.

Copy this:

“Hi [name], your [product] reached you on [day]. How is it so far? Reply 1 if you love it or 2 if something is off. If it is 2, a person from our team replies today.”

Up next Four jobs talk to your buyers. The fifth talks to you, every Monday morning.

Can an agent run your weekly numbers?

The job: every Monday someone opens the store dashboard, the courier panel, the ads account and the WhatsApp inbox, and copies numbers into a sheet. It takes an hour and is out of date by Tuesday.

How an agent helps: it pulls the numbers into one short note: orders, COD share, RTO, ad spend and the questions customers asked most. Then it points out what changed. You decide what to do about it.

Keep human: the decision. An agent can tell you RTO went up in one city. It cannot tell you whether to stop COD there, because that depends on what you know about your buyers there.

Copy this (instruction for the agent):

“From this week’s order and ad sheets, list: total orders, COD share, RTO count, ad spend, the 5 questions customers asked most, and the 3 numbers that changed most since last week. Do not guess. If a number is missing, write ‘missing’.”

Up next Five jobs an agent can take. Now the harder list: what you should never hand over.

What should a D2C brand always keep human?

Here is my list. Keep it next to your yes rules.

1. Refunds, discounts and price exceptions. Money going out needs a person’s yes.
2. Complaints and angry customers. They want to be heard before they want a fix.
3. Health, skin and allergy questions. A guess here can hurt someone.
4. New offers and your brand voice. The agent copies your voice. It should not invent it.
5. Deals with creators, wholesalers and partners. Relationships are not tickets.
6. Cancelling an order or blocking a buyer. The agent flags. A person decides.

Two more rules sit on top of the list. Every agent message says it is AI, because nearly 75% of consumers told Salesforce they want to know when they are talking to an AI agent. And every chat shows a way to reach a person. Gartner’s 2026 survey found customers “will push back when AI acts as a barrier to reaching a human.”

An agent should take work off your team, not put a wall between you and your buyers.

This is the same lens behind our work with a D2C fragrance brand: 3.5x+ ROAS on Meta for 6+ months, with 35% lower CPL. That result came from ads and lead work, not from an AI agent. It taught me what this article is about. The ad brings the buyer. What happens in the chat after the click decides whether she stays.

Up next So how do you start without risking the brand you built? One job, seven days.