WhatsApp AI: buy the platform, or build the agent
Most businesses asking us for a custom WhatsApp agent would be better off paying for a subscription platform. Five questions tell you which one you are, and you can answer all five today.
Say your business takes sixty WhatsApp messages a day. Most of them are the same five questions: are you open, do you deliver to my area, how much is it, do you have it, where are you. Someone has quoted you for a custom AI agent, and you have also seen the self-serve platforms you can configure yourself in an afternoon. You want to know which one is right, and nobody selling either one will tell you straight.
We build the custom side. We also tell people to go buy a platform more often than you would expect, because a custom agent that answers five fixed questions is a bad purchase. The line between the two categories has nothing to do with how modern your business is. It is about how much of the answer lives outside the conversation.
The two categories, plainly
An off-the-shelf WhatsApp platform is a subscription product. You connect your business number, type in your questions and answers or upload a document, draw a few flows, and it replies. The good ones put a language model over the text you gave it, so the replies read like a person instead of a phone tree. What they all share is this: the knowledge is whatever you typed in, and it goes stale the moment your business changes and you do not retype it.
A custom agent is software someone writes and runs for you. It also replies on WhatsApp, but it can reach into your systems mid-conversation: check the real stock count, quote the price that applies to that specific client, hold a slot and write the booking into the calendar your staff already opens every morning. It can escalate to a human on rules you define. It costs a build, plus hosting and model usage every month, plus somebody accountable when your backend changes. That last line is the one buyers forget, and it is permanent.
Five questions to answer about your own business
Answer these honestly, about the business you have this month, not the one you are planning.
- How many messages a day. Under twenty or so, a human with saved replies is still cheaper and better than either option. Sixty to a few hundred is where automation starts to earn its keep. Thousands, and the question stops being whether and becomes which.
- How many of the questions actually repeat. Open WhatsApp and read the last hundred conversations. If eight in ten are the same handful of questions, any automation will pay. If every conversation is genuinely different, you are buying a very expensive way to say "let me check with a colleague".
- Does the answer change during the day. If prices, stock or availability move hourly, a platform will confidently repeat yesterday's answer to today's customer, and it will do it politely and at scale.
- Does it have to write into another system. Reading is the easy half. Creating an order, moving inventory, booking a slot that must never double-book: that is where you need real integration and a person responsible for it when it breaks.
- What does a wrong answer cost. Wrong opening hours costs one walk-in. A wrong quote, a wrong dosage, a wrong legal deadline costs a great deal more. The bigger the downside, the tighter the control you need over what the thing is allowed to say, and that control is custom work.
Rough scoring. If your answers are high volume, mostly repeated, information that is stable for weeks, nothing to write back, and a cheap worst case, buy the platform and stop reading. If two or more of live data, write-back and expensive mistakes are a yes, the platform will disappoint you around month two and you will pay for the migration on top of the subscription you already bought.
When the honest answer is no
We turn down WhatsApp work in three situations. When the volume is low and seasonal, because the build never amortises and you would be maintaining software for eleven quiet months. When the customer wants a custom agent for prestige rather than for a bottleneck they can describe. And when the questions are not repeatable at all, which usually means the business is really a consultancy and the conversation is the product.
The fourth answer is not no but not yet. If nobody in the company can sit down and write the ten most common answers into a document, the problem is not the tooling. Do that document first. It takes an afternoon, it improves your human replies immediately, and it is the exact input both categories need. Plenty of businesses discover at that point that the document alone solved most of it.
An agent cannot know a policy that exists only in the owner's head.
What changes the answer as you grow
Three things move the line. Volume, because at some point the per-conversation cost plus the time your staff spends on handoffs passes the cost of a build. Systems, because the day you have a real inventory or booking backend with an endpoint to call, live answers become possible and stale answers become embarrassing. Rules, because as your pricing turns conditional on client type, volume or zone, a flat list of questions and answers simply cannot express it.
The sensible sequence for most people is to start on a platform, keep every transcript, and re-read them six months later. The transcripts show you exactly which questions needed live data, which ones ended in a handoff, and which ones the customer asked three times because the answer was wrong. That is the specification for a custom agent, and it is a far better specification than anything you or we could write from scratch today.