Planning a WhatsApp AI assistant for customer support

For many businesses in India and other WhatsApp-first markets, customers already send their questions on WhatsApp. An AI assistant can answer the repetitive ones instantly and route the rest to your team — but only if a few decisions are made before any code is written.

This is the checklist we work through when planning a WhatsApp assistant.

Understand the platform rules first

Business messaging on WhatsApp runs through the WhatsApp Business Platform, either directly on Meta's Cloud API or through a provider. A few rules shape the whole design:

• The customer service window. When a customer messages you, a 24-hour window opens in which you can reply freely. Outside that window, you can only start a conversation with a pre-approved message template.
• Templates are reviewed by Meta. Order updates, reminders and follow-ups must be written as templates and approved before use, so plan them early.
• Opt-in is required before you send business-initiated messages to a customer.
• Pricing depends on message type. Meta charges differently for categories such as marketing, utility and authentication templates. The model has changed over time, so check Meta's current pricing before estimating costs.

These rules mean an assistant that "messages customers whenever it likes" is not possible — and designing around the service window up front avoids rework.

Decide what to automate first

Resist automating everything. Start with questions that are frequent, answerable from existing information and low-risk if the answer is imperfect:

1. Frequently asked questions — timings, locations, fees, policies, eligibility.
2. Status lookups — order, application or appointment status through your existing systems.
3. Lead capture — collecting name, requirement and contact preference before a salesperson calls back.
4. Booking — appointments or demos, confirmed with a utility template.

Leave complaints, refunds, payment disputes and anything with legal or financial consequences to people, at least at first.

Prepare the knowledge the assistant will use

An assistant is only as good as the content behind it. Before launch:

• Collect the real sources — website pages, brochures, policy documents, price lists and past support replies.
• Remove contradictions. If two documents give different answers, the assistant will too.
• Write short answers for the top questions in the tone you want customers to hear.
• Decide what the assistant must never say, such as unconfirmed prices, medical advice or promises about delivery dates.

Retrieval-augmented generation lets the assistant answer from these sources and cite them, and to say it does not know when the answer is not there. That is far safer than letting a model improvise.

Design the handoff to your team

The handoff is where most assistants succeed or fail. Define:

• When to hand off — the customer asks for a person, the question is out of scope, the assistant has low confidence, or the conversation shows frustration.
• Who receives it — a shared inbox, with the full conversation and the customer's details, so nobody asks the customer to repeat themselves.
• What the customer is told — honest expectations such as "A team member will reply within business hours."
• Coverage outside working hours — the assistant collects details and confirms that someone will follow up.

Connect it to the systems that hold the answers

Status questions need live data. Plan read-only integrations with your order system, CRM, booking tool or student records first. Actions that change data — cancelling an order, changing an appointment — should start with a confirmation step and be added only after the read-only flows are reliable.

Measure whether it is working

Agree on a few measures before launch and review them weekly:

• Resolution rate — conversations answered without a person stepping in.
• Handoff rate and reasons — why the assistant passed conversations on.
• Unanswered questions — the most useful list for improving content.
• Lead quality — how many captured leads turned into real opportunities.
• Customer feedback — a simple thumbs up or down at the end of a conversation.

The unanswered-questions list is the engine of improvement: each week, add or fix the content behind the most common gaps.

A realistic rollout

1. Launch on the website widget and one WhatsApp number, with a small set of topics.
2. Review conversations daily for the first two weeks and fix content gaps.
3. Add status lookups and booking once answers are reliable.
4. Introduce templates for reminders and follow-ups, with clear opt-in.
5. Expand to further channels, such as Instagram or voice, once the WhatsApp flows are stable.

A narrow, reliable assistant that hands off gracefully earns more trust — from customers and from your own team — than an ambitious one that guesses.