Customers wanted results, not a system
Four paid pilots taught us to operate the service ourselves before deciding what the AI product should become.
The first change was another person.
I had known the developer who would become my cofounder for some time. He had shown little interest in the ideas I had brought up before. This time was different. He said he wanted to build Libera with me.
We did not incorporate a company right away. We decided to start with a small project and see how we worked together. We would split its revenue equally. If we later incorporated, we would start with a 50-50 equity split as well.
I was no longer testing my first hypothesis alone.
We narrowed the problem twice
At first, we wanted to solve companies' marketing problems.
But "companies" was too broad, and so was "marketing." With the customer and the problem both undefined, we could not decide what to build.
We first narrowed the customer to direct-to-consumer beauty brands selling products online.
I initially assumed beauty companies spent more on advertising and suffered more from marketing problems than companies in other industries. The evidence was not strong enough to make that claim. What I could verify was that beauty is highly dependent on marketing.
L'Oréal spent 32.2% of its 2024 revenue on advertising and promotion. According to NielsenIQ, online beauty sales in 2025 were growing nine times faster than in-store sales. Korean cosmetics exports reached $10.2 billion in 2024, up 20.3% from the previous year.
We cannot use the cost structure of one global company as a proxy for a small D2C brand. Still, beauty brands have to keep introducing products to new customers, and online sales are growing quickly. That was enough reason to test beauty as our first market.
Then we narrowed marketing down to performance advertising.
Performance advertising is a loop: form a hypothesis, run an ad, read the data, and decide on the next experiment. It rewards running many tests quickly and carrying each result into the next decision. We thought an AI employee might eventually do this better than a human.
That is not a fact. It is the hypothesis we are testing.
We chose Meta ads as our first channel. Not because Meta is the largest advertising channel overall. Instagram and Facebook make it easy to introduce visual products and compare how people respond to different creative and selling points. More importantly, it was the channel where we could execute and learn fastest at the time.
We sold four pilots
We started building a Meta advertising operating system for D2C beauty brands.
We sold the first two pilots at KRW 500,000 each. One customer came through outbound email. The other found us through a post I wrote on Threads.
Next, we went to two beauty industry events in person. We followed up with people whose business cards we collected and signed two more pilot customers. This time, we charged KRW 600,000 each.
The KRW 100,000 increase was deliberate. Price was another hypothesis. Rather than leave it unchanged, we wanted to learn what customers were willing to pay.
In total, we ran four paid pilots.
Customers did not want a system
I thought customers would keep using the service if we built a good enough system.
As the pilots approached their end, something else became clear. Customers did not really want a better system for using Meta ads. They did not want another tool to learn and operate themselves.
They wanted someone to run their marketing without being chased and turn that work into revenue.
I assumed that generating attributable sales would be enough to win the next contract. It was not. Customers compared the money spent on ads with the sales that appeared immediately. Money used to test a new campaign or creative often felt like a loss.
At first, I thought they simply did not understand advertising experiments. I am no longer sure that was fair.
Maybe we failed to explain what the experiment was meant to learn. Maybe we did not agree on a loss limit and a stop rule. Maybe we talked about ROAS without calculating allowable customer acquisition cost from product margins and repeat purchases. Or maybe the customer simply did not have the cash to absorb the learning period.
We also cannot assume that a ROAS below 100% will pay off over time. Repeat purchases matter only when the later contribution profit exceeds the loss on the first order. We need product cost, contribution margin, repeat-purchase rate, and payback period to make that call.
The lesson was not that "customers do not understand marketing."
Before running ads, we need to agree with the customer on how much downside they can absorb, what we are trying to learn, and which number tells us to continue or stop.
We started an agency service to build the product
Before the pilots ended, we ran a separate outbound campaign. It brought us four agency clients. Three are the D2C beauty brands we chose as our core customer. One is a hair salon, an offline beauty business.
These were not pilot customers converting into agency contracts. We recruited a separate group of clients.
We had not set out to build an agency. But we could not know what an AI employee should own until we had produced the outcome ourselves.
So we decided to use the agency work as a concierge MVP. Before the product is finished, we perform the service ourselves and observe the customer's actual workflow and recurring problems.
What information does the customer provide? What do they want to hand off? What do they insist on approving? When performance drops, what should be checked first? Which decisions repeat every day, and where do exceptions appear?
We will answer those questions by doing the work and keeping a record.
The agency itself is not the destination. We do not want a structure where four times as many customers requires four times as many hours from us.
Our job now is to find the work and decisions that repeat, then move them into the system one by one. We will build the AI employee while producing the results customers actually want.
The first hypothesis is now more specific.
We started with "a digital employee that handles repetitive work for companies." Now we are building "an AI performance marketer that runs Meta ads to grow revenue for D2C beauty brands."
We still do not know if that hypothesis is right.
Next, we will interview customers and use the results from the agency work to understand why customers hesitate to fund advertising experiments and what has to be true for them to keep paying us.