A product from scratch, with no access to users

The problem

When I joined there was no product. There was a bet: that AI could attack the operating cost of anyone producing advertising at scale. I was there from ideation, when there was no screen, no user, no installed base and not a single line of code to look at. At that point the business problem is not a bad screen, it is an open question: is there enough pain here for someone to pay, and where exactly does it hurt.

The bet had outside backing. The project got into the AI startup programmes at Google, Nvidia and Amazon, which settled technical feasibility and infrastructure cost. It did not settle the product question. Cheap infrastructure lets you build fast, and building the wrong thing fast is the most expensive way to find out the bet was crooked.

The research

The first constraint arrived before the first question: the C-level vetoed any and all interaction with users. No interviews, no testing, no validation with the people who would use it. Designing that way, in a product that does not exist yet, is the easiest scenario to turn into guesswork, so the question I had to answer stopped being just what to build and became where to find evidence when I cannot talk to anyone outside.

I built discovery on three sources. I studied how a marketing agency works, where its operating costs sit and what it had already tried to cut them. I ran an extensive competitive benchmark, looking at both the AI tools that were being born and the established advertising suites, and assessed the feasibility of integrating the product with Google Ads. And I used the company's own account manager, trained in advertising, as a domain reference, because he knew from the inside the routine we were trying to replace.

The agency bottleneck was not in creating the campaign, it was in multiplying it. The same concept had to become dozens of formats, and that is where the cost per asset blew up.

The recommendation

I recommended the product not compete on creating the asset, but on multiplying it: a campaign concept goes in, and the platform hands back the whole set of formats ready to publish, in batch. That is the direct reading of where money burned in the agency routine, and it is the only place where an AI tool buys back real human time instead of just speeding up what was already fast.

I also recommended the design system, and it came out in two stages for a practical reason. The first version shipped in plain HTML, CSS and JavaScript, so the product would not be locked to a framework while the stack was still being decided. The second went to Svelte and reached more than 40 components. With no PM, eight developers and two designers, the system was less a consistency indulgence and more the only way eight people could build a new product in parallel without rebuilding the same screen three times. What stayed refused to the end was contact with users.

The result

The platform got to generating more than 200 ad formats in one hour from a single concept, which repositions what the agency does: human time comes out of repetitive execution and goes back to the idea. Cost per asset fell by around 40% against the manual process the market research had mapped, and the production time for a full campaign went from a scale of months to a scale of weeks.

On the team side, the design system reached 80% adoption by engineering, measured on production screens against the Svelte package. For a product born from nothing, with no PM and no process, that number is what explains the other two: delivery speed did not come from rushing, it came from not having to re-decide the button, the form and the loading state on every new screen.

What I took away

I take away a stubbornness. If I started over, I would fight for user research and plan usability testing from day one instead of accepting the benchmark as a substitute. A benchmark shows you what competitors have already decided, and a competitor's decision is not evidence: it is somebody else's guess, with more mileage behind it. A new product needs to know what is missing, and that only shows up in the mouth of the person living the problem.

I also learned that this fight is not won on the argument, it is won on the format and the size of the ask. Expensive, long research is easy to veto. What I would bring to that same table today is the small test, five people and one week, cheap enough to be hard to refuse. And I take with me what worked despite the veto: looking inside the building for someone who has lived the user's problem, like the adman on the team, beats having no evidence at all.