A flow chart cannot answer "eh ada ready stock tak?"
Most things sold as a WhatsApp chatbot are keyword rules and menu trees. Some are AI agents that read what was written and answer from your business knowledge. They fail in completely different ways, and the difference is not obvious from a pricing page.
Auto-reply / flow builder
Keywords, menus and branches you draw in advance. Predictable, and blind to anything unplanned.
AI agent
Reads the message and composes an answer from the knowledge you gave it.
How they differ where it matters
| How they differ where it matters | Auto-reply / flow builder | AI agent |
|---|---|---|
| An unexpected question | Falls through to a default reply, or repeats the menu. | Answered, if the knowledge covers it. Escalated if not. |
| Malay, English, or both in one sentence | Each phrasing is a keyword you have to have thought of. | Handled as written, including a mix of both. |
| Typos and shorthand | Missed. "bkn", "brape", "ready stok" are each separate rules. | Understood. |
| Setup effort | You draw every path. The work grows with every case you think of. | You write down what your business does, once. |
| Maintenance | A new product or price means editing the flow. | Update the knowledge or the product list; the answers follow. |
| Predictability | Total. It can only say what you wrote. | Depends on the guardrails — a good one refuses rather than guesses. |
| Cost shape | Usually flat. Sending is cheap because nothing is thinking. | Usually metered, because each reply is real computation. |
| Worst failure | A customer stuck in a menu that does not contain their question. | A confidently wrong answer, if it was not given limits. |
Which one is right for you
Pick an auto-reply tool if
Your inbound is genuinely repetitive and narrow — an order-status lookup, a fixed menu of three services, a form. Rules are cheaper, and predictability is worth more than flexibility when the question set really is closed.
Pick an AI agent if
Your customers write in their own words, ask about prices and stock, mix Malay and English, and go off-script constantly. Which is to say: if you sell things to the Malaysian public.
Where aiBalas lands
aiBalas is an AI agent, and the guardrails are the part we spend the most care on, because the worst failure above is the one that matters. Each of the four agents is written to refuse rather than improvise: prices come from your product list, and anything outside what it was told is handed to a human instead of being invented. You choose what it must never answer — a clinic agent, for example, does not touch symptoms or diagnosis. The bar we hold it to is not "does it always answer", it is "when it does not know, does it say so".
Common questions
How do I tell which one I am being sold?
Ask what happens to a question the vendor did not anticipate. If the answer involves adding a keyword, a branch or a fallback message, it is a rule engine. Ask to see it answer something off-script during the demo — with a typo in it.
Will an AI agent make up prices?
A badly built one will. A well built one quotes from a product list you control and says it will check with you when something is not there. Ask specifically where prices come from, and test it on a product that does not exist.
Do I need to write a lot to set an AI agent up?
Less than you would expect, and much less than drawing a flow. You describe what your business does, your hours, your prices and your policies — closer to briefing a new staff member than to building software.
Can an AI agent hand over to a person?
A good one must. Handover is the release valve that makes the whole thing safe: the agent stops replying, a human is alerted, and the customer is not left arguing with software.
Try it before you decide
Free to start, no credit card. WhatsApp, Telegram and website chat on one agent.