A good customer service chatbot answers the questions your team receives most often, quickly and correctly, and passes everything else to a person without making the customer repeat themselves. Building one comes down to a handful of decisions: which questions it handles, which channel it lives on, what kind of bot it is, which systems it reads from, how it hands over to humans and how you measure it.
Large language models have made chatbots far better at understanding informal, messy messages. They have also made it easier to build a bot that sounds confident while giving wrong answers. The steps below aim to keep the first benefit without the second risk.
Start with your chat logs, not the technology
Export two to four weeks of conversations from WhatsApp, Instagram, email and your website chat, and sort them by topic. For an online store, the list often includes "where is my order," "is this item in stock," "how much is shipping to my city," "how do I pay" and "can I return this." For a clinic it might be schedules, prices and BPJS coverage; for an insurer, claim status and required documents.
Then place each topic in one of three groups:
| Type of question | Examples | What the bot needs |
|---|---|---|
| Static answers | Opening hours, return policy, payment methods | Approved answer text |
| Data lookups | Order status, stock, appointment slots | Integration with your systems |
| Judgment calls | Complaints, refunds, damaged goods, special requests | A fast handover to a person |
A chatbot earns its keep on the first two groups. If most of your volume sits in the third, improve your agents' tools and reply templates first. A bot would only add a step between the customer and the person who can help.
Choose the kind of chatbot
There are three common designs, and most practical bots combine them.
| Approach | How it works | Strengths | Weaknesses |
|---|---|---|---|
| Menu or button bot | Customer taps options | Predictable and cheap to run | Frustrating for anything off the menu |
| Intent-based bot | Classifies each message into trained topics | Answers stay controlled | Needs many example phrases; struggles with slang |
| LLM with retrieval | A language model writes answers from your approved content | Handles free text and informal Indonesian well | Needs guardrails, testing and per-message model costs |
A sensible hybrid uses buttons for well-defined paths such as order tracking, a language model for free-text questions answered from approved content, and a clear route to a human. For other places this technology helps, see our list of practical generative AI use cases.
You also face a buy-or-build decision. Omnichannel inbox platforms, including several Indonesian providers, include bot builders and are quick to start. A custom build makes more sense when the bot must read from your own systems, follow specific business rules or keep data under your control. The trade-offs mirror those in our guide to custom software vs off-the-shelf.
Building the chatbot step by step
Step 1: Pick the first channel
WhatsApp is where many Indonesian customers already talk to businesses, so it is often the right first channel. Automated bots need the WhatsApp Business Platform, accessed directly through Meta or through an official Business Solution Provider; the free WhatsApp Business app does not support them. Replies are free-form within 24 hours of the customer's last message, while messages you start outside that window must use pre-approved templates.
Step 2: Write the answers once, in one place
Collect policies, product information, shipping rules and payment methods (QRIS, virtual accounts, e-wallets, cash on delivery) in a single knowledge base, written in the tone your brand uses. Give it an owner who updates it when anything changes. A language model can only be as current as this content.
Step 3: Connect the data lookups
Link the bot to the systems that hold live answers: order status, courier tracking through the courier's API or a shipping aggregator, stock levels or appointment slots. Verify the customer before showing details, for example by matching the order number with the phone number. Start with read-only access, and add actions such as cancellations later, with explicit confirmation.
Step 4: Design the handover to a person
Hand over when the customer asks for a person, when the bot is unsure, when the topic is a complaint or refund, and after two failed attempts to help. Pass the full transcript and order details to the agent so the customer never repeats themselves. Outside working hours, say honestly when a person will reply.
Step 5: Set guardrails
- Answer only from approved content and live data; otherwise say so and offer a person
- Never promise refunds, discounts or compensation. Route those decisions to staff
- Never guess prices, stock or delivery dates
- Ask only for the data you need, and mask sensitive fields in logs
- Log every answer so it can be reviewed
This matters legally as well as commercially. In February 2024, a Canadian tribunal held Air Canada responsible for a refund policy that its website chatbot had described incorrectly. Customers treat what your bot says as your company's word.
Step 6: Test with real questions
Build a test set of 100 to 200 real past messages across all topics, including typos, abbreviations and mixed Indonesian and English. Have customer service staff grade each answer as correct, partly correct, wrong or "should have handed over." Fix the content and rules, then retest until wrong answers are rare and caught by the handover.
Step 7: Launch in stages
Start with one channel, a few topics or selected hours. Review transcripts every week, and expand only when the results hold. A staged launch follows the same logic as our MVP guide: learn from real users before you scale.
Indonesian language and data protection
Indonesian customer messages are informal and full of shorthand: "kak," "min," "brp," "ongkir," "cekout," often mixed with English and, in some regions, local languages. Modern language models handle this far better than older intent-based bots, but your test set should prove it. Decide the form of address too: "Kak" suits a fashion brand, while "Bapak/Ibu" may suit a bank or hospital.
Chat transcripts contain personal data such as names, phone numbers and addresses, and sometimes health or financial details. Indonesia's Personal Data Protection Law (Law No. 27 of 2022, UU PDP) applies, and its two-year transition period ends in October 2024. Practical steps:
- Tell customers they are talking to an automated assistant
- Define how long transcripts are kept and who can read them
- Treat health and financial data as specific personal data with stricter handling
- Check where your AI provider processes data, because UU PDP sets conditions for transfers outside Indonesia
What to measure and what drives the cost
Track these numbers from the first week:
- Resolved without handover, counting only conversations where the answer was actually right
- Handover rate and reasons, which show where content or integrations are missing
- Wrong answers found in review, the number that protects your reputation
- Response and resolution time compared with your human-only baseline
- Customer satisfaction after the chat, and repeat contacts about the same issue
Running costs come from four places: WhatsApp charges, which Meta bills per 24-hour conversation at rates that vary by conversation category and country; model usage, billed by the amount of text processed; integration and hosting; and staff time to maintain content and review transcripts. Teams often underestimate the last item, yet it decides whether the bot stays accurate after launch.
FAQ
Will a chatbot replace our customer service team?
No. A well-designed bot takes repetitive questions off your team's plate, but complaints, refunds and judgment calls still need people. Someone also has to maintain the knowledge base and review transcripts. Plan for your team's work to shift toward escalations, content quality and the conversations where a human makes the difference, rather than planning for the team to disappear.
Can we use the regular WhatsApp Business app for a chatbot?
Not for a real bot. The free WhatsApp Business app offers greeting messages, away messages and quick replies, which help a small team respond faster. An automated assistant that understands questions and reads from your order system needs the WhatsApp Business Platform, accessed directly through Meta or through an official Business Solution Provider.
How long does it take to build a customer service chatbot?
As a rough planning range, assuming one channel, five to ten topics and one system integration such as order status, a small team can often have a tested first version live in 6 to 10 weeks. Content preparation, integration work and testing drive the timeline far more than the AI model itself.
How Weaveward can build this for you
We would start by analyzing a sample of your recent chat logs to separate the questions worth automating from those that should always reach a person. Then we would connect the bot to your first channel, for example WhatsApp through the Business Platform, link it to your order or booking system for live lookups, and build the handover into the inbox your agents already use. The first release covers your top questions, and we validate it against a test set of your real past messages and a staged launch before adding topics.
Learn more about our AI solutions for customer service, or talk to our team about the questions your customers ask most.