If the workflow worked,
he wouldn’t be selling it.
A Threads post this week asked for someone to build an AI that answers WhatsApp and email inquiries, writes quotations, follows up, runs the courses, manages the trainers, collects payment, issues certificates and does the marketing on three channels. “Freelance pon boleh.” Seven hours later someone replied that he is certified in agentic AI architecture and sets up exactly this kind of end-to-end multi-agent system routinely — DM for a prototype. Twenty-eight replies, most of them similar. I read it twice, then asked myself the only question that matters.
A screenshot of a Threads post came across my feed. A training business asking, in plain Malay, for someone to build it an AI. Then the list. Then, a few hours later, the replies. I want to walk through it slowly, because everything that is currently wrong with how AI gets bought and sold in Malaysia is sitting inside that one thread.
What was actually asked for
Strip the word AI out and read the list again. A CRM. A quoting engine. A follow-up sequence. A learning-management system. A staff roster. A payment gateway with reconciliation. A certificate issuer. A marketing operation across social, WhatsApp and email. That is seven products. A software company would put a team on it for a year and still ship it in phases.
It was posted as one job, for one person, part-time. And within a working day, several people said yes.
I am not writing this to mock either side. The buyer asked an honest question. The sellers answered the way the market has taught them to. I am writing it because I build these systems for a living, and the gap between what was promised in that thread and what I know it takes is the whole story of AI in Malaysian SMEs right now.
The question nobody in the thread asked
If you can build a system that answers every inquiry, quotes, follows up, takes the money and markets itself across three channels — you own the most valuable thing in small business: a company that sells without staff.
So why are you selling it as a freelance service, in a Threads reply, at a training centre’s budget?
There is only one honest answer. The machine does not run the way the reply says it does. If it did, it would be running for him. A workflow that works gets kept. A workflow that gets sold is usually the one that did not.
I wrote the same thing about AI courses: the guru sells the method because the method does not pay as well as selling it (operator or guru — run what you teach). Trading signals, dropshipping playbooks, now agentic workflows. The product changes. The tell does not.
Why people believe it anyway
Because the demo is real. An agent answering a WhatsApp message in your brand voice, in two minutes, on a screen share — it works, and it is impressive. I have built those demos. AI is extraordinary at producing something that works today.
What the demo cannot show is month three. The trainer roster changes and nobody updated the prompt. The payment gateway returns a case the flow never saw. The model provider ships an update and the quotation that was perfect in July now rounds differently in September. A participant receives someone else’s certificate. And the person who built it has a day job, because “part-time” was in the offer. The word was the disclosure.
I have said before that AI builds MVPs, not release products, and that we have not shipped a fully stable AI product ourselves. That was not modesty. Ninety-nine per cent reliable is a demo. A product is what happens to the other one per cent at nine at night, and who answers.
The buyer in that thread used the right word without knowing it. He asked for a sustainable solution. Sustainable is not a feature of the software. It is a person under an obligation to keep it running. A freelancer in a reply has no such obligation, and you cannot buy one from someone who does not carry it. When the wrong certificate goes out, the seller is not on the complaint. You are.
Half of it is not AI at all
Look at items four to six again — schedules, trainers, payments, certificates. None of that is an AI problem. It is ordinary software that has existed for a decade and costs less per month than one trainer’s lunch allowance. Putting a language model in charge of payment reconciliation is not innovation; it is choosing a probabilistic tool for the one job where you need the answer to be the same every time. I learned that the expensive way and went back to deterministic automation for anything repeated.
AI belongs in the parts that involve language and judgment at volume: the first reply, the follow-up, the draft quotation a human approves, the marketing copy. That is roughly half the list. The other half should be bought off the shelf and connected. Anyone proposing to “multi-agent” your payments is telling you they have not run one.
To be fair to the lone rangers
Some of them are good. Every agency I respect started as one person with a laptop. Being able to build a thing is not the same as wanting to run a training business, and a builder is entitled to sell his hours instead of his own product — a good plumber does not have to own the building.
But a plumber sells a bounded craft with a visible track record. The claim in that thread was “end-to-end, seven functions, routinely.” That is not a craft claim. It is a claim to have solved something the industry has not, made in a reply, for free.
And the buyer set the price of the answers he got. Asking for seven products at freelance rates guarantees that the only people who say yes are the people who do not yet know what it costs. The mismatch is on both sides of that thread.
What I would actually do with that list
Buy one function, not seven. Start with the inquiry reply — cheapest to build, most visible, and the cheapest mistake, because a human can correct a bad reply and cannot un-issue a certificate. Rank the rest by the cost of being wrong, and keep the language model out of the bottom of the list entirely.
Then ask any seller three questions before a ringgit moves. Show me one of these running for six months at a paying customer I can call. When it breaks at night, who answers, and under what agreement. And what do you carry if it sends the wrong thing to the wrong person — because advice is worth exactly the consequences the adviser carries.
Then the fourth question, the one I started with. Show me your own funnel. If the man selling you a business that runs itself still answers his own DMs by hand, you already have your answer.
The bottom line
A system that could really sell without staff would be selling for him, not being sold by him.
Buy the half that is language. Buy the other half off the shelf. And before you buy anything, ask to see the seller’s own machine running — not the prototype, the one that pays his rent.
Published by IMA AI — September 2026. Written by someone who builds these systems for a living and has not yet shipped one that runs itself.