Build a chatbot or answer the chats yourself
Judge whether your chat volume and question mix justify a bot, what it cannot answer, and how a missing handover path loses the enquiries worth money.
The same four questions all day
Look at a week of chats on your website and your WhatsApp number. In most small businesses the same handful of questions account for the bulk of the volume. What are your charges. Do you deliver to my pin code. Are you open on Sunday. Is this size in stock. Somebody answers these by hand dozens of times a day between doing their actual job.
That pattern is what makes a chatbot attractive. It is also why so many bots disappoint. The repetitive questions are easy to automate and they are rarely the ones that produce revenue. The enquiry worth money is usually the oddly worded one that no scripted flow anticipated.
What a bot cannot answer and what it costs
A bot cannot answer anything that is not already written down somewhere it can read. Live stock prices delivery dates order status and anything requiring judgement all sit outside its reach unless a system holds that data and the bot is connected to it. It also does not reduce headcount in a business where the questions do not repeat.
The costs are platform fees and design work. Chat platforms charge per conversation or per contact and WhatsApp carries its own messaging charges. GST at 18 per cent applies to the service fees. The design work is the real expense. Someone has to read the real conversations group them into intents write the answers in your brand's voice and define what happens when the bot fails. Journey mapping is published as a chatbot journey mapping engagement so that design effort can be priced against doing it in-house. Maintenance continues afterwards because every price change and new product changes the answers.
Who gains from a bot and who should answer manually
A bot earns its place where volume is high and questions repeat. Ecommerce stores clinics with appointment enquiries education providers with admission queries and any business receiving chats well outside working hours all qualify. It also helps where the same question arrives in three languages.
It does not suit a business where every enquiry is different. A custom fabricator an architect a corporate law practice or a specialist consultancy will find that almost nothing is repeatable and the bot becomes an obstacle between a serious buyer and a person. It also does not suit low volume. If you receive ten chats a day a person can answer all ten properly in less time than it takes to maintain a bot. Building one at that volume is a project that costs money and slows your response quality. A business whose pricing is always negotiated should be especially cautious because a bot that quotes anything creates expectations you then have to manage down.
How to build one that is worth having
Start from transcripts rather than from a flowchart. Export the last few hundred real conversations and sort them into groups by what the person actually wanted. The groups that appear repeatedly are your intents. Anything that appears twice is not an intent and should be left to a human.
Write the answers next and write them short. A chat answer that runs to a paragraph is not read. Where an answer depends on live data decide honestly whether the bot can reach that data. If it cannot then the bot should collect the question and pass it on rather than guess.
Then design the failure path before the happy path. Define how many attempts the bot makes before it offers a person. Define working hours and what it says outside them. Make the route to a human visible in every flow rather than hidden behind a menu. Launch with the three largest intents only and add more once the transcripts show what is being missed.
Judging whether it is working
The flattering measure is conversations handled by the bot. It counts every interaction including the ones where a frustrated visitor typed three times and left. Containment rate on its own flatters for the same reason because a conversation abandoned in disgust counts as contained.
The measures that matter are resolution without a human handover followed by no repeat contact from that person within a day and the rate at which visitors abandon mid-flow. Alongside those track enquiries that reached a human and converted. A bot that raises containment while total enquiries to sales fall is costing you money quietly.
The mistake that loses the good enquiries
The common mistake is launching without a working handover. The bot answers what it knows and loops on everything else. A visitor with an unusual requirement types their question twice gets the same menu back and leaves. That visitor was frequently the largest potential order of the week.
This is expensive in a way the reports hide. The dashboard shows a high containment rate and falling agent workload which looks like success. The lost enquiries never appear anywhere because they never became enquiries. Put the escape to a human in front of the visitor from the first message and accept a lower containment figure in exchange for keeping the conversations that matter.
This guide explains how something works. It is not legal, tax or financial advice, and rules change. Confirm anything regulatory with a qualified professional before you act on it.