The First Five Minutes of Building a WhatsApp Agent

A founder messages me: “We want to build a WhatsApp agent for our business.”

I ask: “What should it do?”

They say: “Help customers.”

That is where most projects go wrong. Not in the technology. Not in the API. In the answer to that question.

“Help customers do what?” is the only question that matters in the first five minutes of building a WhatsApp agent. Everything else the model, the prompts, the welcome message, the button labels is secondary until that answer is specific.

The first five minutes of building a WhatsApp agent should not be spent writing prompts. They should be spent understanding the job. What is this agent supposed to remove from the business? A support burden? A sales bottleneck? A payment confirmation headache? A broken follow-up process?

Until that is clear, you are not building an agent. You are decorating a chatbot.

Minute one

Not what should the agent say. Not what personality it should have. Not which model it should use. What work should it do?

A useful WhatsApp agent must own a job. It helps customers place orders. It qualifies leads. It books ferry tickets. It follows up unpaid invoices. It confirms payments. It routes complaints. It sends receipts.

That is the level where serious product thinking starts. If the job is vague, the agent will be vague. If the job is “help customers”help them do what? Buy? Book? Pay? Complain? Track? Cancel?

A WhatsApp agent without a clear job becomes a polite confusion machine. It talks a lot. It solves little.

Minute two

Before you automate anything, watch the human process. Not the process written in the company document. The actual process. The messy one. The one happening inside WhatsApp chats, Excel sheets, phone calls, screenshots, and “boss nimefuatilia” updates.

A customer sends a message. Someone replies manually.

The customer asks for price. Someone checks a catalog.

The customer asks for availability. Someone calls another person.

The customer pays. Someone checks M-Pesa. A screenshot is forwarded.

A receipt is typed. A delivery person is called.

The customer asks again because nobody updated them.

That is the workflow. Do not laugh at it. Map it. Because that mess is where the product lives. A WhatsApp agent should not be built from imagination. It should be built from the real sequence of work already happening.

Minute three

Good automation starts where friction is repeated. Not where the founder is excited. Not where the technology is impressive. Where the pain is boring, frequent, and expensive.

Bad builder asks

“How do we make this agent smart?”

Good builder asks

“Where is the business leaking time, money, or customers?”

The best WhatsApp agents do not try to replace the whole business on day one. They remove one painful loop. Then another. Then another. That is how automation compounds.

Minute four

This is where most chatbot projects become useless. They answer questions, but they cannot act. They can explain payments, but they cannot check payment status. They can describe products, but they cannot check stock. They can apologize, but they cannot escalate to a human.

That is not an agent. That is a talking FAQ.

A real WhatsApp agent needs tools. It needs access to business data, APIs, rules, permissions, and escalation paths. It needs to know when to stop talking and hand over.

A chatbot responds.

An agent completes work.

Without action, the agent is just another layer between the customer and the answer. And customers hate unnecessary layers.

Minute five

Do not start by designing the full conversation. Start by designing the first successful outcome. What should happen in the first working version?

A customer asks about a product and gets the right item.

A passenger books a ticket and receives confirmation.

A lead answers three questions and lands in the sales pipeline.

A customer gets an accurate answer without waiting for a human.

That is the unit of value. Everything else is secondary. The greeting is secondary. The tone is secondary. The model choice is secondary.

Can the customer complete the thing they came to do? If yes, you have a product. If no, you have a demo.

Too many WhatsApp agents are demos pretending to be products.

People spend too much time polishing the welcome message. “Hi, welcome to ABC Company. How may I assist you today?” Nobody cares. The customer did not come to admire your greeting. They came to solve something.

The best WhatsApp agents get to value quickly. They do not trap users in long menus. They do not force people to read paragraphs. They do not behave like corporate IVR systems wearing a chatbot costume.

WhatsApp is not a website. People do not behave there the same way they behave on a dashboard. They send incomplete messages. They mix languages. They disappear and come back hours later. They expect memory and speed.

Someone will type:

“bei?”

“iko?”

“nataka mbili”

“nimeshalipa”

“naenda Zanzibar kesho”

“can I pay by card?”

“boss mbona haijaja?”

That is not noise. That is the interface. The agent has to understand messy intent and turn it into structured action. That is why WhatsApp agents are not just chatbots they are translators between human mess and business systems.

There is a dangerous lie in the market right now: that if you add an LLM to WhatsApp, you have built an agent. No. You have built a fluent responder. That is not the same thing.

An LLM can understand language. But the business still needs logic, boundaries, integrations, confirmation steps, audit trails, human escalation, and fallback behavior. A good agent is not just a model. It is a system. The model is one part. The workflow is the product.

This is where serious WhatsApp automation platforms matter. Not because businesses need a place to write prompts because they need a place to connect conversations, data, payments, APIs, and people into one working flow. That is what Sarufi is built to do. Not “build a chatbot.” Turn business processes into WhatsApp agents that can actually run.

One audience. One workflow. One primary outcome. One clear escalation path. One measurable metric. That is enough for version one. Once the agent proves value, expand. Do not build a giant brain. Build a reliable worker.

Many teams celebrate message volume. “10,000 conversations handled.” So what? Did sales increase? Did support load reduce? Did customers complete the process faster? Did the team save time?

Message volume is not value. In fact, too many messages can mean the agent is inefficient. The real metrics are business metrics.

Conversion rate.

Resolution time.

Payment and booking completion.

Manual hours saved.

Revenue recovered.

Customer drop-off rate.

That is how you know whether the agent is working.

If you start with prompts, you will build a talker.

If you start with menus, you will build a tree.

If you start with AI, you will build a demo.

If you start with the business process, you might build something useful.

A WhatsApp agent is not a chatbot with better language. It is a business process with a conversational interface. It should understand messy input, connect to real tools, complete real work, and produce measurable outcomes.

Anything less is automated noise.

The first five minutes are not about building the agent. They are about deciding whether the thing deserves to exist.