Skip to content
Keystone Technologies
All insights
Agent as a Service

WhatsApp support with an AI agent: what works and what doesn't

Where an AI agent helps on WhatsApp, from order status to returns, where it struggles, and how to set up opt-ins, tone and hand-over to your team.

Keystone Technologies · 13 October 2026 · 4 min read

For many Indian brands, WhatsApp is where customers actually talk to you. They message to ask where an order is, to change an address, to start a return or simply to check that you are real. It is convenient for them and, as volumes grow, relentless for whoever is holding the phone.

An AI agent can take on a large share of those conversations. It can also make things worse if it is set up carelessly. This article covers what tends to work well on WhatsApp, what doesn't, and the decisions worth making before you switch anything on.

What an agent handles well on WhatsApp

The best fit is work that is frequent, follows clear rules and can be answered from your own systems. When the agent is connected to your store, courier tracking and helpdesk, it can give a real answer rather than a scripted one. Good starting points include:

  • Order status: looking up the order from the customer's number, checking the courier and replying with where the parcel is and when it should arrive.
  • Simple changes: updating a delivery address or contact detail before the order is shipped, within the rules you set.
  • Returns and exchanges: checking eligibility against your policy, collecting photos where needed and booking the pickup.
  • Product questions: sizes, materials, care instructions and availability, answered from your catalogue rather than from memory.
  • Payment and COD queries: confirming whether a payment went through or a cash-on-delivery order is still active.

In each case the agent is doing what a trained support person would do: look something up, apply a rule and reply clearly. The difference is that it does this at any hour, without a queue building up overnight.

Where it struggles, and what to do instead

Some conversations are a poor fit for an agent working alone. An angry customer whose order arrived damaged for the second time wants to feel heard by a person. A request that sits outside your policy needs judgement, not a rule. Long, rambling voice notes, mixed languages in a single message and questions about something that happened months ago can all trip an agent up.

The answer is not to avoid these conversations but to recognise them early and route them to your team. The agent can still do the groundwork: pull up the order, summarise what the customer has said and suggest a next step, so the person picking it up starts with context instead of scrolling back through the chat.

Opt-ins and messages you send first

Replying to a customer who messages you is one thing. Messaging a customer first, with shipping updates, feedback requests or offers, is another. WhatsApp's business platform has its own rules on consent and on the kinds of messages businesses may start, and those rules change from time to time. It is worth reading the current guidance before planning any outbound messages.

In practice, a few habits keep you on the right side of customers as well as the platform:

  • Collect a clear opt-in, for example at checkout, and record when and how it was given.
  • Keep order updates and promotional messages separate, so customers who want one don't get the other.
  • Make it easy to stop messages, and make sure the agent honours that request straight away.
  • Send fewer, more useful messages rather than a steady stream of reminders.

This is general information, not legal advice. If you are unsure how consent or data protection rules apply to your business, speak to a qualified adviser.

Getting the tone right

WhatsApp feels personal, so a reply that reads like a policy document stands out badly. Customers expect short, friendly messages that get to the point. Give the agent a simple tone guide: how formal to be, whether to use the customer's first name, how to apologise, which words to avoid and how to sign off.

Keep replies brief and break longer answers into steps. Avoid sending three messages where one will do. If your customers often write in Hindi or a regional language, decide whether the agent should reply in kind, and test that carefully before going live. Reviewing a sample of real replies each week in the early days is the quickest way to spot anything that sounds off.

A good rule of thumb: if you would be uncomfortable seeing a reply screenshotted and shared, the agent should not be sending it. Review early conversations with that test in mind.

When to hand over to a person

A clear hand-over is what makes customers trust the channel. Agree the triggers in advance. Common ones are a customer asking for a person, signs of frustration, a request above the agent's approval limits, a repeat complaint about the same order, or any question the agent cannot answer with confidence from your systems.

When the hand-over happens, tell the customer honestly that a team member will take it from here and roughly when to expect a reply. Pass the full context to your team in the helpdesk they already use. Then make sure someone actually picks it up: a hand-over that sits unanswered for a day does more damage than no agent at all.

Starting small

You don't need to hand the whole channel to an agent on day one. Many brands start with order status alone, check the replies closely, then add returns, then product questions. Each step shows you where the agent is reliable and where your team still needs to be involved.

A typical first agent goes from discovery to production in about three to five weeks, depending on system access. Pilots start from ₹40,000, and the final price depends on the systems and work involved.

If WhatsApp has become the busiest part of your support and you'd like an agent to take on the routine conversations, book a call with us. We'll look at what your customers ask most and suggest a sensible place to start.

Tell us what's slowing your team down. We'll show you what we'd build.

No pitch deck. A direct conversation about your systems, your workflows and what it would take to fix them.