I Am No Longer the One Pressing the Buttons: The CEO as Supervisor

There is one question I ask of every automated task in our company: if it hits a decision point halfway through, and I am not sitting next to it, what will it do?

The honest answer in the early months was: it guesses. A machine does not hesitate the way a person does. Faced with ambiguity, it picks a direction, carries on, and reports back in a very confident voice. That question taught me that moving from doer to supervisor is not about watching machines run. It is about designing a loop, and writing the rules into the exact place where the supervisor will be absent.

From the person who presses the button to the person who supervises

In mid-July I told our engineering team something simple: "I will stand in the supervisor's role." The process I wanted was this: the system surveys the situation on its own, proposes a strategy, I approve, then it executes and reports the result. Not me touching everything. It is a loop, not a one-off handover: propose, approve, execute, report, and when something goes wrong, back to propose.

At first I thought the hard part was making the machines good enough. In practice the harder part was making the loop dense enough. I told the team plainly that too little intensity was not acceptable, and that running once a week was not enough. The work had to run by itself every day, some of it one to three times a day. A supervisor, however sharp, cannot rescue a sparse loop.

The consequence is that the CEO role changes completely. I no longer measure myself by how many tasks I did by hand. I measure by the quality of the approval points, and by whether the loop keeps turning when I am not looking at it.

Three small habits that keep the loop turning

Assigned means started. When I hand over a task with clear criteria, the system starts work. It does not insert an extra confirmation step just to be safe. Every unnecessary confirmation gate is a place where the loop stops and waits for me. What I need to provide is clear criteria; I ask for two or three options, choose one, and then extend trust.

Ask the CEO for exactly one action. Some steps only I can do. When the system is blocked there, it must hand back to me one short task with everything else already prepared, rather than throwing the whole problem onto my desk. Do the maximum yourself, and ask only for the part no one else can do. A supervisor's time is limited; if each question becomes a small meeting, I am back to being the doer.

Every report carries a recommendation. A number standing alone forces me to work out the next step myself, which means I am doing the work again. A number accompanied by a recommendation lets me nod or shake my head. Then I ask the familiar follow-up: how far along is it? Reports come daily, after completion, not whenever someone remembers.

When nobody answers halfway through

A new employee can turn to a colleague and ask. A task running while everyone is asleep has no one to ask. So I drew a principle: the rules must already be inside the work order, travelling with the task to the very end, rather than sitting in a shared handbook that a long job may forget. Every order we give a machine carries the following:

• Do not delete anything outside your own scope: no data, no files belonging to others.
• Do not message real customers and do not publish anything publicly without approval.
• Do not let secrets or customers' personal information leak into logs, reports or change history.
• Do not push anything live until the checks are satisfied.

And one rule above all the others: where you are unsure, stop at that item, write it into the report, and continue with the remaining items. This turns ambiguity into information for the reviewer, instead of a silent decision made by the machine. Alongside it sits a rule about truth: if something cannot be done, say so and give the reason; never write that it is done.

I also do not trust the phrase "it should be fine." Every change must be verified by actually running it. The lesson we paid for (in the same vein as the legal guardrail is part of the product) is that a test suite showing all green can still let most real advertising lines slip through once they are tried. Green in the lab is not the same as correct in the field.

Beyond rules, there are ceilings. Each working session for a machine has a limit on steps and on the amount of computation. Sensitive work passes through two separate modes: plan only, and execute only after approval. For a machine operating directly on another system's interface, the permitted actions form an allowlist; touching anything outside it stops the entire run, rather than skipping it and carrying on.

The most sensitive area is money. Our automated advertising mode is off by default: the machine only measures and recommends. Every budget increase, even one made by hand, is capped at VND 1 million per day at a single point, and the button that changes budgets exists only in my private conversation. The machine has no authority to move money. This is another way of stating the principle in AI reads widely, writes narrowly.

Supervision is not blind trust

Real supervision needs measurement. Each important AI agent of ours has an automated test set built from real cases, run every Sunday; there are currently four. An agent is given more work, or switched to a newer AI model, only when its score matches or beats the previous one. I do not have to read every output to know whether a machine is still standing; I read the scores, I read reports that carry recommendations, and I approve at the points that carry consequences for customers or money.

I should also state the limit plainly. This loop reduces how often I have to touch things, but it does not make me redundant. Anything involving money, words spoken to customers, or legal exposure still needs a human to look before it goes out. That is a design decision, not a gap. The aim is to concentrate my scarce attention exactly there.

What investors and partners should take from this

First, control at HOPE rests on structure, not on the machine's good intentions: rules inside the work order, narrow write permissions, a person approving before anything goes outside, and ceilings on money and on each run.

Second, as a result, operating cost scales with software rather than with headcount. A three-layer AI system handles orders, customer care, content, reporting and payroll, automating more than 60% of the work of a 12-person team, and one supervisor can keep them all in view because the loop has a steady rhythm and the reports carry recommendations.

Third, this is an organisational capability, not a tooling trick. I tend to believe a strong leader is one who uses leverage and does not do one hundred percent of the work personally. With AI, leverage is only safe when the person holding the lever has written the rules for the moments when they are absent.

Closing

I began by doing everything by hand and arrived at a different role: the person who asks the right question, writes the rules into the work order, and approves exactly where approval is needed. If you are considering a partnership or an investment, the question I would like you to ask us is the one I opened with: when nobody answers halfway through, what does your system do? For us, the answer is written down before the work begins.