Who says no now?
On what replaces scarcity as the constraint on software
Almost every SaaS platform and a good share of consumer apps shipped an AI feature in the last two years. Most are not very good. Some have little AI beyond the interface. Others work as intended but solve no problem anybody needed solved.
I often wonder how these features got there.
Bad product ideas existed before AI. What changed is the friction between an idea and the product.
Software has always had more proposed work than engineering capacity. Prioritisation was the response, and it quietly did another job too: preserving product strategy, keeping technical debt survivable and protecting the customer experience. When you can ship ten things a quarter and forty are queued, you have to say no thirty times. Even a mediocre team, forced to say no thirty times, can do only so much damage.
Scarcity was a subsidy for taste. By taste I mean something narrower than visual refinement: the ability to tell what makes a product more itself from what merely makes it larger. Scarcity forced organisations that lacked discipline to behave, occasionally, as though they possessed it.
That subsidy is disappearing.
AI is removing friction across software production. Designs, code, tests and migrations all come faster now.
Kyle Daigle, GitHub’s COO, posted in April:
Yup, platform activity is surging. There were 1 billion commits in 2025. Now, it’s 275 million per week, on pace for 14 billion this year if growth remains linear (spoiler: it won’t.)
I would hold that loosely. Daigle disclaims his own extrapolation, the two figures are not measured the same way, and commits are not features. But the direction matters more than the number.
For a team with a coherent thesis, this is good news. It can spend the capacity on reliability, latency, polish and deletions. A team without one will spend it on features. The decision stays exactly where it was. The throughput multiplies.
Scarcity never prevented bloat. Word and Salesforce became bloated while engineering was still expensive. Scarcity merely capped the speed at which bloat accumulated, and AI is removing the cap while leaving the incentives untouched.
Companies face enormous pressure to demonstrate AI adoption in the product. Boards ask what the AI strategy is. Competitors launch copilots, so product roadmaps acquire copilots. That pressure manufactures features whose primary purpose is to signal movement.
These incentives extend well beyond AI:
Enterprise software needs features to survive RFPs and expand contracts.
Consumer products need new surfaces to increase engagement.
Internally, a team with no roadmap risks losing headcount, while a product leader who removes three features has less to display than one who launches three.
Accommodation always looks reasonable in the moment. But software that bends in every direction transfers the cost of every decision to the user. Compressed development cycles supply the capacity, organisational incentives supply the demand, and the combination breeds bloatware.
The visible alternative is opinionated software: software that stays coherent about the job it does. That is a different axis from how large it is.
Obsidian is extensible through plugins and immovable about its substrate: local files, plain-text Markdown and user ownership of data. Permissive where variation is cheap, stubborn where it is expensive. Things has declined collaboration for close to two decades, remaining a task manager for one person.
Both are independent companies, which makes refusal easier. But founder taste is not a scalable substitute for structure. Linear is venture-backed and still stubborn, because for Linear the opinion is the positioning and its customers are buying the refusal. Companies have to make saying no institutionally safe, so that refusing a feature does not cost somebody revenue, relevance or resources.
There is another reason restraint matters. Steph Ango, who runs Obsidian, asks a question I think about often:
What are the chances this thing will be abandonware in a month? You can make something in a few days but will you commit to maintaining and improving it for the next few years?
AI has collapsed the cost of building, but building was only the visible half of the cost. The expensive half was always the promise: a commitment to everybody who depends on it, a constraint on every future redesign, another behaviour to test and document, a permanent row in the support queue. A feature is not finished when it ships. It begins accumulating obligations.
Agents will make parts of maintenance cheaper. The evidence so far says not the expensive parts. A large study of AI-generated code found that more than 15% of commits from every assistant introduced at least one issue, and that 22.7% of the AI-introduced issues it tracked were still alive in the repository’s latest version. Google’s 2025 DORA report found the same pattern one level up: AI amplifies whatever system is already there. Strong testing converted the speed into outcomes; weak systems converted it into disorder. What still lands on people is judgment, coordination, accountability and the willingness to stay responsible after the person who proposed it has moved on.
So who says no now?
The backlog used to, automatically. A feature had to be valuable enough to displace something else. The process was slow and often dysfunctional, and good ideas died inside it, but it imposed a cost on addition.
Now that building is cheap, that cost has to be recreated deliberately. The replacement needs three things: a thesis about the job the product does, a company arranged so that refusing features does not cost revenue or headcount, and the willingness to own what gets shipped for as long as it exists.
None of them pays off within a quarter, which is why they get skipped. But restraint compounds. A product that has stayed coherent for five years is one people trust with their work, because they can predict what it will still be. Trust is one of the few moats in software that gets harder to cross with time, and every competitor spending those years adding surface area is building the case against itself.