A table for two
A man came with an idea for a coupon app and left with a table-for-two booking. Today he would turn that idea into an app within the hour, and skip exactly the part of the work that cannot be sped up.
Every restaurant has a table nobody sells. The table for two. It brings in the best turnover in the house and is almost always free at the last minute, and no app deals with it. I know this because a man once came to me wanting to build a coupon app.
Coupons were the thing to build at the time, and anyone with some experience could have put a stack together for him. I asked him almost nothing about coupons. I asked what he had done until then and who he knew. He knew people at dating apps, and whoever runs a dating app has users who need to meet somewhere tonight, the two of them. Nothing was left of the coupons. What was left was a product cut to fit him.
The cheap part
If that man came to me today, he would already have his coupons. He would type the idea into a tool and get an app that works. Building has become the cheap part of the job, and that is good news, I use it every day.
What is interesting is the conclusion drawn by the people who make those tools. The approach growing fastest this year is called spec-driven development: a person first writes down what the product is and which rules apply, and the code follows from that. GitHub's tool for working this way has collected more than 130 thousand stars and reached version 1.0 in August. InfoQ says the same in one sentence: writing code has become easier, and the harder and more important skill is saying what is wanted and within which limits.
So the weight has not gone away. It has moved in front of the code.
Quiet decisions
Whoever skips that part is not left without decisions, they get them ready-made. Researchers at the University of Southern Denmark described this year how agents choose the framework, the database and the hosting before a person sees anything, and how the same task, worded differently, produces a differently built system.
I see it myself. Almost every new project stands on Supabase, because Lovable sets it up by default. It is a good database and I use it, but I chose it, and to most people it happened. And the database is the smallest of those decisions. In the same way, without a question being asked, an answer also appears to what the product is and who it is for.
Before the start
Those decisions do not follow from the idea, they follow from the founder. What shape the idea should take, who the co-founder is, what team it needs, who the first clients are: all of it depends on what that person knows, who they know and who they can sell to. Two people with the same coupons need two different products. A tool will give them the same one, because it sees only the idea.
One feature
That work does not stop once the project is under way. On one of ours we had a superior product with a pile of features, and that was exactly what held us back. The more the product could do, the more people at the customer had to take part in choosing it, so decisions dragged on. For the last three months we have been offering one single feature. A decision now takes a few days, the sale fits into two meetings, and the result is up to eight times better.
The market is already paying for this work. The number of job postings for engineers who work directly at the customer rose about eightfold in a year. In one survey, half of the companies building such teams say the same thing: AI products reach the client and stall before production.
The wrong speed
That is why speeding up everything is not cheap, even though it looks like an hour and a monthly subscription. Whoever hands the decisions to AI as well will not get a bad product. They will get a product that works and that they have no influence over: they did not decide what it is or who it is for, so when the customer reacts differently than expected, they do not know what to turn. The phases that were skipped do not disappear. Their turn comes later, when the money has already been spent.
None of this is a complaint about AI. When we know exactly what we want, there is no better tool: what used to take months now takes days. But it speeds up whatever it is given. Give it an unfinished decision, and we will arrive at the wrong place faster.
The table for two was not in the idea. It was in the man who brought it, and that part could not be made any faster. Everything after it can.