Taste as the scarce asset when building is free
Taste is the new scarce asset. For the first time in my career, that is actually true.
Building got cheap this year. Not free, but close enough that the old constraint stopped doing its job.
For a long time engineering scarcity did our selection for us. Mediocre ideas died in the backlog because there was no capacity to build them, and we got to call that prioritization. When I was building Payit, before Rappi acquired us, the roadmap was decided by arithmetic: four engineers, two real things a quarter. The other eighteen ideas were not rejected on merit. They were rejected by physics.
That protection is gone. What replaces it is taste, and I think taste is going to decide who wins the next few years.
I mean that concretely, not as a compliment you pay to designers. Taste in UX, so the fifth screen gets deleted instead of defended. Taste in product sense, so you can tell a real problem from a well-argued one. Taste in what people actually want, which is almost never what they ask for in the call.
Taste is the only input in product work that did not get cheaper this year.
The simple part
The cost of a wrong pick did not go down. It moved somewhere harder to see.
When building was expensive, a bad bet cost you a quarter of engineering time and you felt it the same week. Now it costs two days of agent time, and the bill arrives later, in usage. Everything extra you ship lands in the product someone has to work through every day: one more option in the menu, one more toggle in settings, one more page in the help center, one more thing support has to explain to a customer who never asked for it.
At Yuno a merchant integrates once and then lives inside that surface for years. Whatever we shipped that should not exist is still there when they open the docs to add their tenth market, and it is our name on it.
So the ratio moved. Cheap to build, same cost to use. Multiply the build side by ten and you get zombie products at industrial scale: alive in the codebase, dead in the roadmap, and standing between the customer and the thing they actually came for.
Which means the question changed. Instead of "can we build this", the question is "of the forty things we could ship this week, which one earns its place in front of the user". That question has no answer in the backlog, no answer in the capacity plan, and no answer in the model. It gets answered by whoever in the room has the best taste, and everyone else finds out whether they were right sixty days later.
Capacity used to be the thing we were measured against. Taste is the thing we will be measured against now.
The complicated part
Here is what taste looked like on a product I ran for a long time. At Revolut I led RevPoints, the loyalty programme. A consumer product, millions of users across many countries, and a mechanism that looks trivial from the outside.
Take the mechanism: a point. Every point you give a user is a promise that you will hand something back later. So a loyalty programme is a small economy with four levers. The earn rate, how many points a user gets per unit of behaviour. The redemption value, what a point is worth when they spend it. The catalogue, what they can spend it on. And breakage, the share of points that never get redeemed, which is the difference between a programme that funds itself and one that quietly eats margin. Defining those four took a lot of smart people a long time, to make sure the programme was going to be successful.
Now an agent can build all of that before lunch. Ledger, balances, expiry job, a redemption screen. It compiles, it balances, it ships. The taste is in the settings.
Whether points expire, and when. Expire them fast and breakage looks great for two quarters, then users learn the points are fake and stop noticing them. Whether a premium tier earns more, which is the moment a loyalty programme stops being a cost centre and starts selling the subscription. Whether a point is worth the same on every redemption or worth more on the ones you want to push. Whether you let other brands put your points in front of their own customers, which turns the programme into distribution and also means a decision you make about redemption value now moves someone else's numbers.
Then the reader. A user does not want to understand any of this. They want three things in the first second of the screen: how many points they have, what that gets them, and how that makes them feel. Every earn rule we put above those three is a rule they scroll past.
Now the part that matters for the argument. An agent will hand you forty versions of that programme before lunch. All forty compile. All forty pay out correctly on the happy path. Several would pass review. The difference between them is whether users who hold a balance come back and use more of the product, or treat the balance as noise. There was only one way to find out: run the experiment, market by market, and watch retention and usage move or not move. Which earn rate reads as generous in one country and as noise in another is written down nowhere. You learn it by watching a lot of cohorts.
The bank version of this is the same story with a different reader. At Rappi we launched banks in five countries (Mexico, Colombia, Brazil, Peru and Chile), and the arguments that mattered were never about whether a screen could be built. They were about which field to remove from onboarding, what to assume about the user instead of asking, and how much of a digital flow survives a market like Mexico, where 80% of transactions are still cash. Every one of those calls had a plausible case on both sides. The good ones came from people who had watched a lot of users abandon a lot of flows.
So here is my working definition. Taste is compressed exposure to a specific system. It is narrow, it is local, and it decays the moment you stop looking at the data. It does not transfer cleanly from one market to another, which is exactly why it stays scarce while everything around it gets cheap.
The useful thing about compressed exposure is that it produces predictions, and predictions can be checked.
What I'm not sure about
What I am not so sure about is whether taste is enough on its own.
Taste with nothing underneath it is confident preference. It needs a real problem, one you can state in a sentence without the customer in the room, and a hypothesis you can actually prove, with a number and a date in it. Without those two, you cannot tell the difference between good judgment and a good story, and neither can anyone on your team.
The other direction fails just as often. A solution nobody likes, attached to a problem that is unmistakably real, is also not a product. You can have the problem right, the mechanism right, the users polite about it in research, and adoption that never moves.
So maybe taste is not the thing that decides the next few years. Maybe it is the tiebreaker, and what decides is whether you picked a problem worth solving and wrote down what you expected before you built anything. Real problem, provable hypothesis, then taste to choose between the forty ways to build it. In that order.
There is also the model question. If taste is compressed exposure, models compress better than I do, and on a system with years of data the honest expectation is that they end up holding more of it than the people who lived through it. What may stay with us is narrower than we would like: choosing which problem to point all that compression at, and carrying the consequences when the answer is wrong. I still think that is the scarce asset. I am less sure it stays scarce for as long as we are telling ourselves.
So think about the last ten product calls you made on taste alone. For how many of them could you name the real problem in one sentence, and the prediction you would have been wrong about?