AI Should Serve a Business With Orders, Cash and Customers

One bug in our systems stays with me far longer than any outright outage. A data-collection job lost its login session. It did not complain. It returned a zero, and that zero was written over the real figure of 1,481. The job finished, there was no red line anywhere, and everything looked normal. The only problem was that the number the whole company was looking at had become wrong.

In a demo, nobody notices an error like that, because nobody reconciles anything. In a business with orders, waybills, bank statements and people waiting for payroll, it is the most expensive kind of error: silent, and plausible. This essay is about how I think about AI from that starting point. AI should serve a real business with customers, orders and cash flow. It is not there to put on a show, and it is not there to teach other people how to use AI.

A silent wrong number costs more than a system that goes down

After that incident we set a rule and wrote it into code: failure must be loud. If a collection job returns zero rows, it stops, writes nothing, keeps the previous figure and raises a warning flag. Every new data source must be connected to a shared watchdog before anyone is allowed to trust it. The rule sounds simple, but it changed how I choose work for AI. The first question is no longer "what can AI do?" but "if it is wrong, who finds out, and when?"

In the same spirit, AI at HOPE has no authority over money. Ad autopilot is off by default; the machine only measures and recommends. Every budget increase, even a manual one, is capped at VND 1 million per day (roughly US$38) in one single place in the code. The control for changing spend exists only in my private chat. AI cannot make payments. I do not see this as a limitation of AI. It is the condition under which an owner is willing to hand work over to it.

Where a demo and an operation differ

A demo only has to be right while someone is watching. An operating system has to be right when nobody is: Monday morning, when orders pour in and the person who reconciles has not yet started the shift. Three differences stand out to me.

• Results must be reconcilable. Our revenue is extracted automatically from the order group chat, then matched against shipping waybills and bank statements. AI extracts orders very quickly, but the value lies in that ledger agreeing with real money.
• Quality needs a yardstick. Each important AI agent has an automated test set built from real cases, run every Sunday: operations advisor 22 cases, lead routing 16, customer replies 15, sales-conversation scoring 13. The latest scores, on 10 October 2026, were 20/22, 16/16, 15/15 and 12/13. I state the two imperfect results on purpose, because a test set where everything scores full marks is usually a test set that is too easy. An agent is given more responsibility, and we switch to a newer AI model, only when scores match or improve.
• Real customers are the final test. In parallel, the system measures real response times: a median of one minute, and 80% of messages answered within fifteen minutes (30 days to 10 October 2026). The machine drafts replies; staff approve them before they go out.

None of this is glamorous, and none of it looks good in a video. But it is the part that separates a tool from an operating capability.

Learning AI by solving the company's own numbers

I did not learn AI by looking for courses. When I interpreted the six-piece model of an organization, I wrote: "Turn the organization into a school, and the CEO must study relentlessly", meaning study exactly what solves the problem in front of the numbers, not whatever is interesting. That applies directly to AI. HOPE's problems are concrete: which orders are real, how much was collected this month, which customers are waiting for a reply, who has bought a second time, which post is allowed to go out. Each question forced me to learn one more piece of AI, because if I did not, the number in the ledger would be wrong.

This way of learning has two consequences that few people mention. First, my knowledge always has something to hold on to: I understand a technique because I have seen it get a number right or wrong, and that number was mine to answer for. Second, I know when AI cannot be trusted, because I have paid for the times it could not. It is slower than attending a talk, but what you learn does not fall away when real work arrives. I wrote about the six-piece model separately in The Six Pieces of an Organization.

Build it internally first, then use it as the frame

In August, in a working instruction, I wrote: "Build it internally first, then use that as the frame to carry over quickly." The order matters. A frame built for a customer who does not yet exist is only a good-looking frame that nobody has put load on. A frame that has already run inside our own company, through bad data, messages arriving at midnight and tightened advertising rules, already has the scratches it needs.

As a result HOPE has two identities at once, and both are real. As a D2C dietary-supplement business we sell to individual customers and live with advertising law, inventory and cash flow. At the same time, that way of operating has been written down as processes, watchdogs, scoring suites and permission layers, clearly enough that a partner could use it as a frame for their own business. I say "could", not "has": this is a capability we have, and I describe it only at the level of principle.

Who a Chief AI Officer is

I believe the Chief AI Officer role will become more and more necessary, but I want to be clear about how I understand it, because the most common misreading is "the person best at AI in the company". To me it is someone who understands the business, then coordinates tools to produce real results. Tools change by the quarter: a new model arrives, an older one steps back. What remains is the ability to look at a process and say which step AI can do and a person should approve, which step should not be handed over at all, and what yardstick shows it is working.

So the person in that seat must be able to read a reconciliation, understand why content has to pass a banned-claims filter, and know where a budget cap belongs. Engineering is half of it. The other half is answering for the final number. For someone like me, who started as a software engineer and moved into international sales, those two halves are one job. This is my own judgement: teams that only demonstrate AI will struggle to hold their position once people start asking about results, while teams that use AI to serve a real business will be less affected, because they already measure in orders and money.

What investors and partners should understand about HOPE

• Costs scale with software more than with headcount. Paid advertising in 2026 is about 1% of revenue, roughly VND 26,000 (about US$1) per paying customer. Customers come mainly from an owned community of about 240,000 followers.
• Reconciled numbers. Net revenue for the first nine months of 2026 was about VND 3.57 billion (about US$137,000), a run-rate of about VND 4.4 billion a year (about US$170,000). Reconciled figures begin in 2026, so I do not publish earlier years.
• AI has boundaries. A person approves content before it goes out; customer data is masked before it reaches AI; every AI call is logged, counting calls without storing content.
• Weaknesses remain. We are a small company, reconciled data exists only from this year, and a couple of our test sets still contain cases we fail. I would rather say so than have someone else find out.

I do not want HOPE to be remembered as a company that uses a lot of AI. I want it to be the company where, when someone opens the ledger, the numbers match the money in the account, and when AI gets something wrong, someone knows immediately. Everything else is just tools.