For the past two years, the loudest debates inside our company have used the same vocabulary: scale, productivity, AI. Every quarter delivered a fresh version of the same conversation. How do we scale sales? How do we make customer success more productive? Which AI tool takes which workload off whose desk? The vocabulary came from outside, from the discourse every founder swims in, and it always sounded like the right conversation to have.

It was not. Looking back, I think we spent our energy one rung too high. Our actual problem was never that we could not do things ten times in parallel. It was that we could not dependably do them twice.

Scalability is a prestigious problem. It implies you have something worth multiplying. Admitting that you cannot repeat your own wins is a much less flattering diagnosis, so nobody volunteers it. When the industry conversation is about scaling and AI-powered productivity, those become the frames you reach for, whether or not they fit your situation. Paul Graham’s advice to do things that don’t scale gets quoted in every startup circle, including ours. We read it the convenient way: as permission for a temporary phase, something to exit as fast as possible. We skipped the part where the unscalable phase is supposed to teach you what the repeatable core of your business actually is.

What we called a scaling problem in sales was usually a knowledge problem. What we called a productivity problem in operations was usually a documentation problem. Therefore we should have focused more on reproducibility than on brainstorming how to make a barely running product more productive and scale it. Like for example how to design training for CS Reps or Sales Reps and how to create products which solve the problem of one segment consistently.

What does it mean to reproduce a problem solution?

Reproducing a solution means the same result can be achieved again absent the original people. It has two directions: a capable new hire can learn to solve the problem quickly, and a solution that functions for one customer can be transferred to other segments, companies, and countries.

The first direction is about people. In sales, customer success, and operations, our best answers live in individual heads. One person can price a complex deal. One person can resolve a specific type of customer escalation. They are good at it. The mistake we made was that as soon as we saw early signs of success, we wanted to scale it already. Only while trying this we noticed that other people weren’t able to close deals or if another person took over backoffice tasks we had a certain lack of quality.

The second direction is about transfer of a working solution. A solution we build for one customer should carry over to the next segment or at least the next company which was similar. Ours mostly did not. Every customer became a special case, handmade every time. That produces a lot of charm and very little repetition. Entering a neighboring market felt like starting from zero because, in terms of documented, transferable solutions, we were.

Enter new markets with consulting first

The instinct of many startup founders is to enter with already scaleable products. I now think the opposite is right. A new market deserves an agency phase. Work the first customers the way a consultancy would: handmade, senior people on the account, priced for learning rather than for margin. The output that matters in that phase is not revenue per head. It is documented solutions: what the problem actually looks like there, what we built, and why it worked.

That is the reading of Graham’s advice we skipped the first time. The unscalable phase is not an embarrassment to exit as fast as possible. It is the instrument that produces the material worth scaling.

The phase does need an exit criterion, otherwise it becomes a lifestyle. Ours, tentatively: once we have delivered the same solution to a third customer, and the third delivery worked without the person who built the first one, it is ready to be productized. Before that point, scaling talk is premature. After it, scaling stops being a debate and becomes execution.