Why the stakes are higher here

A chatbot on your website is a widget somebody can close.

A chatbot on WhatsApp is in the same app as their family and friends. It feels personal, and a bad experience there registers differently.

A customer stuck in a loop on WhatsApp is more annoyed than one stuck on a website. That is the practical reason to scope these tightly.

What a WhatsApp bot can do well

Four things.

Answer a defined set of factual questions. Opening hours, order status, returns policy, delivery areas. Definite answers, definite questions.

Route the conversation. Ask one or two questions and pass it to the right team, with what has been said already attached.

Collect information before a person joins. Order number, account reference, a description of the problem. The agent starts with context instead of asking for it.

Cover out of hours. Acknowledge, capture, and set an expectation — provided it is honest about being a bot and about when somebody will reply.

What it does badly

Anything outside its list. The moment a question is unusual, a bot produces a confident irrelevant answer — which is worse than no answer.

A customer who is already annoyed. Asking somebody to rephrase when they are frustrated makes it worse, quickly.

Multi-part problems. “My order arrived damaged and I was charged twice” needs a person.

Anything with money attached. Refunds, credits, disputes.

The rules that decide whether it helps

Four. Breaking any one costs you more than the bot saves.

Escape after one failure

If the bot cannot answer, it says so and offers a person. Immediately.

Not after three attempts. Not after “could you rephrase that”. Once.

A customer being asked to reword their question is forming an opinion about your company, and it is not a good one.

Say it is a bot

Do not give it a human name and photograph. People work it out, and being caught pretending costs more trust than the automation saves.

A simple line at the start is enough.

Keep the scope small

Write down exactly what it handles. Everything else goes to a person.

A bot with a small confident scope is useful. A bot claiming to handle everything gives confident wrong answers on the questions that mattered most.

Never invent a fact

No prices it made up. No delivery dates it estimated. No promises about what you can do.

If the answer is not a fact from your system, it should not be given as one.

The handover

The moment the bot passes to a person is where most of the damage happens.

During working hours: the conversation history goes with it. The customer never repeats themselves. If nobody is free, say so honestly — “everyone is with a customer, we will reply within the hour” beats silence.

Outside hours: be clear it is out of hours, capture their details and question, and give a realistic time. Then meet it.

Never leave somebody waiting on a bot with nobody behind it. A conversation with no reply is worse than no bot at all.

Menus and free text

Two approaches, and they suit different situations.

Menu-driven. Numbered options — “Reply 1 for order status, 2 for returns”. Predictable, and it makes the scope obvious. It also constrains the customer.

Free text. The customer types what they want, and the bot works out what they mean. Feels natural, and it fails less gracefully.

A menu is usually the safer choice on WhatsApp, because it sets expectations. The customer sees what the bot can do and does not waste a message on something it cannot.

A useful pattern: a menu, plus “or type your question and we will pass you to someone”.

Measuring it

Five numbers. One matters more than the rest.

Conversations started. Volume.

Resolved by the bot. How much it actually handled.

Escalations. How often a person was needed. Not a failure metric — a well-scoped bot escalates often, and that is correct.

Abandonment. How often somebody stops replying mid-conversation.

Blocks and reports. Directly threatens your quality rating.

Watch abandonment most closely. It tells you whether the bot is helping or driving people away, and it rarely appears in vendor demonstrations.

When not to have one

Three cases where the answer is no.

Low message volume. If you get twenty conversations a day, a person handles them. A bot adds a step and no capacity.

Complex or high-value products. If every enquiry needs a real conversation, a bot delays the conversation people want.

No capacity to respond. If messages already sit unanswered, generating more of them is not an improvement.

A sensible rollout

Weeks 1–4. No bot. Handle conversations manually and record what people actually ask.

Week 5. Build a bot for the five most common questions only.

Week 6. Add routing and information collection.

Week 8. Review abandonment and escalation. Adjust the scope.

Then expand slowly, based on what people actually ask rather than what you assumed.

Point one matters most. Building a bot before you know what people ask means guessing, and guessing produces a bot with the wrong scope.

The short version

A WhatsApp bot works when it is narrow, honest about being a bot, and one message from a person.

Escape after one failure. Never invent a fact. Watch abandonment.

And run manually for a month first. What customers actually ask is rarely what you would have guessed.