What a local business will pay an AI agent to do

The job, in one sentence

A small business gets the same twenty questions on WhatsApp every day: the timings, the rates, the address, "is a room free tomorrow?", "do you have a Sunday batch?". Many arrive at 11 at night or during the lunch rush, when nobody can answer. A reply agent answers the ones it can from the business's own information, collects the details for the ones it can't, and hands those to a person.

That is the whole job, and it is narrow on purpose. The agent is not a general assistant, does not run the business, and never promises what only a person can promise. This course teaches you to build that agent in plain PHP, put it live for a real business, and sell it as a service you can support.

Who has this problem

Walk down any market road and you will find it:

• A guest house: room rates, check-in time, parking, "kal ke liye room milega?"
• A clinic: OPD timings, the consultation fee, how to book. Never symptoms or medicines: those go to the doctor.
• A coaching class: batch timings, fees, the demo class, what the course covers.
• A salon: services and prices, Sunday hours, appointment requests.
• A shop: "do you have this in stock?", the delivery area, the return policy.

They share three things. The questions repeat. The answers live in a short list of facts the owner already knows. And a message left unanswered for an hour is often a customer who has already gone to the next shop.

Why "a reply agent" sells and "an AI chatbot" doesn't

The owner of a guest house is not shopping for AI. They want two things: no enquiry left unanswered at night, and their staff not typing the same rates forty times a day. Sell that outcome, in their words.

Be just as clear about the limits, because they are what makes the agent safe to switch on. It answers only from what the owner wrote down. It never confirms a booking, takes a payment or quotes a price that isn't on the list. It tells every customer it is an AI and that they can reply HUMAN at any time. An owner who hears those limits up front trusts you more, not less.

What Meta allows on WhatsApp

Since 15 January 2026, Meta's WhatsApp Business Solution Terms have barred general-purpose AI assistants from the WhatsApp Business Platform. Bots that serve a business's own customers, such as support, enquiries, bookings and orders, are still allowed. That is exactly the agent in this course: it answers questions about one business and politely declines everything else (lesson 7 builds that in). Meta changes its terms from time to time, so read the current version before you take on each new client. This summary is as of 19 October 2026.

Is this really an agent?

Yes, but a carefully fenced one. The model decides how to word each answer and whether a person is needed, and it extracts the details of a booking request. Your PHP decides everything else: which messages the model sees, what happens before it is asked (opt-outs, people who want a human, spending limits), and what is checked after it answers (prices, length, empty replies). For an agent that talks to real customers of a real business, that split is the point. The cost of a wrong answer is a wrong price quoted to a customer, so the code keeps the final say.

If you want the general theory of agents, tools and loops, the Agentic AI Course on this site covers it. This course is about one agent that earns its keep, done properly.

What you will build

The course has four parts:

1. Find the job (lessons 1–3): choose a niche, find a first client, and write down exactly what the agent will and won't do.
2. Build it (lessons 4–10): a complete reply agent in plain PHP and MySQL, with no Composer and no background workers, so it runs on ordinary shared hosting. Every line of its code is printed in the lessons, and each part is explained where it appears.
3. Go live (lessons 11–13): connect a client's number to the WhatsApp Cloud API, work out the running costs in rupees, and handle their customers' data properly.
4. Sell and support (lessons 14–16): price the work, write the proposal and agreement, hand it over, and look after it every month.

If you would rather start from a finished, installable version, with a setup wizard, a web inbox and a leads page, the AI Reply Agent is sold separately and together with this course. The lessons point out where each idea lives in its code, so you can read one against the other.

How the code in this course was tested

Every code example in this course was tested against simulated versions of the Claude API and the WhatsApp Cloud API that run on your own computer, not against the live services. Lesson 5 prints both stand-ins in full, so you can practise without an API key and without spending anything. The stand-ins check your requests the way the real services do, but their answers are scripted. Real answers from Claude will be worded differently. That is why lesson 10 shows you how to rehearse with your own key, and have the owner approve the answers, before any customer sees the agent.

One more honest note. This course does not promise you clients or income. What you earn depends on the businesses you find and the work you put in. What the course gives you is the method, working code and real cost figures, so that what you sell is something you can stand behind.

Try it

Before the next lesson, list three businesses near you that you could visit this week. For each one, write down five questions you think their customers send on WhatsApp. Lesson 2 turns that list into a first client.