Legal and ethical guardrails are part of the product

This year we reviewed and took down a large batch of our own older articles. They had been published earlier, written the way the whole industry was used to writing, and when advertising rules for dietary supplements tightened, they became risk sitting in public. Removing posts is not hard technically. What is hard is looking back and admitting that part of the content asset you built turned out to be a liability.

After that I stopped asking "how do we comply and move on". I asked a different question: how do we make this kind of mistake impossible to repeat, even when I am not watching?

Guardrails are not paperwork

In dietary supplements and beauty, the familiar way to compete is the medical claim: promise a little more power than the competitor, a little more speed. It wins in the short term, because a stronger promise draws more clicks. But it carries two costs. Regulators keep tightening, so today's edge becomes tomorrow's violation. And customers lose trust: when everyone promises the same thing, the promise stops meaning anything, and the whole category pays in credibility.

We chose the opposite direction. HOPE and IKI sell dietary supplements, and we say exactly that. No comparisons with medicine, no promises of results within a number of days, no customers' before-and-after photos as proof, no overstating what a product does. In my family this attitude has an old name, "phúc đức": doing business decently, with clean products. I do not regard it as decoration. For a company that wants to last, it is a property of the product, like the cleanliness of an ingredient. Customers do not see it every day, but without it everything else loses its footing.

Promotions follow the same logic. We only give gifts; we do not discount by percentage. Percentage discounts pull the whole brand into a price race, while a gift leaves the product's price intact.

A rule written in words does not stop code

The foundational lesson for me came from something unrelated to advertising. For a period our revenue was read from accounting software that people filled in by hand: it ran a few days late, and one month it was missing 125 orders. We already had a rule, written in the handbook, not to take sales figures from that source. Yet 53 places in the codebase were still querying exactly that source, and the company was nearly blind to its own revenue for nine days.

Since then I have kept one sentence for the team: a rule written in words does not stop code. The rule has to live in the code. Applied to legal and ethical matters, it means each guardrail below is a mechanism that actually runs, not a line in a handbook.

Three guardrails for content

• One banned-claims filter, and only one. Everything the machines generate, from blog posts to replies to customers, passes through the same gate. It blocks medical-claim language, time-bound result promises and percentage discounts. A single gate means no shortcuts, and a rule changed in one place changes across the whole system. Our health blog has more than 300 articles, with English and Japanese editions, and every one passed through this gate.
• Evals that deliberately try to make the AI break the rules. Every Sunday, four automated test sets run against our key AI agents, built from real cases. Within them are cases written solely to see whether the AI will say something illegal when pushed. An agent is given more responsibility, and we switch to a new AI model, only when scores match or improve. The latest scores, on 10 October 2026, were 20/22, 16/16, 15/15 and 12/13. I mention the imperfect ones on purpose: a system that measures itself has scores that are not round, and those numbers tell us exactly what to fix.
• A person approves before anything is published. Our principle is that AI reads broadly and writes narrowly. Posts and customer replies are drafted by machines and go out only after a person approves. AI has no permission to move money, delete data or change infrastructure.

Ad money and customer data

Money is where one mistake costs for real, so the guardrails here are the tightest. Ad autopilot is off by default: the system only measures and recommends. Every budget increase, even a manual one, is clamped at VND 1 million per day (about US$38) in a single place in the code. I do not patch the cap at each caller, because with many patches one of them will eventually be forgotten. The control for changing budgets exists only in my private chat, never in a shared group that includes staff. The result is that even if a machine is wrong or a person taps the wrong button, the maximum damage is known in advance. Our total paid advertising in 2026 is about VND 34 million, roughly 1% of revenue, so the cap is not what limits our growth; it limits the size of a mistake.

Customer data is handled the same way. Phone numbers and emails are masked before they reach AI or appear in reports. The AI assistant answers only in allow-listed chats. AI call logs record counts only, never content. And right before code is saved there is an automatic gate: any commit containing phone lists, secret keys or secret configuration files is blocked. I put the gate there because that is the moment when a human slip can still be stopped at the lowest cost.

What investors should take from this about HOPE

The first item on my own risk list is that advertising rules for dietary supplements keep tightening. I regard it as both a risk and a moat, and a moat has to be watched constantly. What I want you to see is that HOPE's legal and brand risk has been engineered: there is one filter, a weekly scored test set, a money cap in one place, a person approving before anything goes out, and a gate before code is saved.

This matters to investors for two reasons. First, IKI is positioned on "measured, not promised", which needs no claims, so the tighter the rules become, the less we have to change. Second, compliance cost scales with software rather than headcount: when content multiplies, the filter is still one gate.

I am not saying the risk is zero. Eval scores are not perfect, rules will keep changing, and people can still err. What we can do is move risk from something we hope will not happen to something with a mechanism that blocks it, a number that measures it and a person who answers for it.

Closing

Doing business decently sounds like an ethical choice. I see it as an engineering choice too: build the guardrails into the product from the start rather than bolt them on after an incident. For us, "phúc đức" is something written as code, scored every week and held by a named person.