For years, software companies treated engineering capacity as the scarce resource.
There were always more ideas than people who could build them. Roadmaps were negotiations over developer time. A prototype was expensive enough to feel like a decision.
AI is changing that constraint quickly. An individual can now move from an idea to a working application in days or even hours. Teams can create more experiments, integrations, internal tools, and customer-specific variants than their old delivery systems were designed to absorb.
That sounds like a golden age for software.
It may also become a graveyard of perfectly functional apps that nobody needs, finds, trusts, or returns to.
More Code Is Not the Same as More Value
A recent NBER working paper gives this tension unusually concrete form. The researchers studied more than 100,000 GitHub developers across successive generations of AI coding tools. They found large increases in coding activity, but those gains became smaller at every step toward a finished product. The cumulative effect associated with autonomous agents reached 180% for commits, then fell to 50% for projects and 30% for releases.
The marketplace data is even more revealing. Across iOS, Android, Chrome, and SourceForge, the researchers observed more new applications but no corresponding increase in aggregate usage. On iOS, monthly releases accelerated sharply while the share of new apps failing to reach even ten ratings in their first three months rose.
The paper is careful about what this means. It cannot fully separate lower product quality from congestion in discovery channels, and its data only runs through early 2026. Adoption may take time. Ratings and downloads are also imperfect measures of value.
But the strategic signal is already strong:
the ability to produce software is expanding faster than the market's capacity to notice and absorb it.
Shipping Was Never the Finish Line
The industry often talks as if an app moves through a simple funnel:
idea, code, release, success.
The real sequence is longer:
problem selection, product design, code, integration, reliability, positioning, discovery, onboarding, trust, habit, support, and renewal.
AI has attacked one expensive part of that chain. It is beginning to help with several others. But automating the build does not automatically solve the whole system.
A customer does not experience the elegance of your code-generation workflow. They experience whether they immediately understand the product, whether it works in their context, whether their data feels safe, whether the first result is worth the effort, and whether the product keeps its promise on the fifth visit.
That is why “we can build it” is becoming a weaker argument for building anything.
Customer Experience Is the Product Boundary
When competitors can reproduce features quickly, the durable differentiation moves outward from the feature itself.
It lives in the complete experience:
- how clearly the product identifies the customer's problem,
- how little effort is required to reach a useful outcome,
- how gracefully it handles uncertainty and failure,
- how well it fits existing habits and systems,
- how responsive the company is after launch,
- and how consistently every interaction reinforces trust.
This is not decorative polish applied after engineering. It is the architecture of adoption.
An AI-generated application can be technically impressive and still transfer too much work to the user. It can offer twenty capabilities while leaving the first useful action unclear. It can generate plausible output while making verification exhausting. It can save ten minutes in a workflow and introduce a new permission, billing, or reliability concern that costs more than it saves.
The winning experience is rarely the one that demonstrates the most capability. It is the one that removes the most uncertainty.
Distribution Is No Longer a Launch Function
Distribution is often mistaken for promotion: buy attention, publish content, announce on social media, and push people toward a landing page.
That is only the visible edge.
Real distribution is the system that repeatedly connects a product with the right customer at the right moment. It includes channel access, partnerships, reputation, community, search visibility, integrations, sales motion, word of mouth, and the product's ability to invite its own next use.
It also includes context.
A new standalone app asks the customer to discover a brand, evaluate a promise, create an account, understand an interface, grant permissions, move data, and form a new habit. Each step is a tax. A product embedded in a workflow the customer already uses can remove half of those taxes before the first interaction.
As app creation becomes cheaper, these routes to market do not become cheaper at the same rate. Customer attention remains finite. Trust accumulates slowly. Organizational access is negotiated. Habits resist change.
That makes distribution more valuable precisely because production is becoming abundant.
The New Competitive Unit Is the Whole System
This shift is uncomfortable for builders because it weakens the relationship between technical effort and market reward.
A small team may produce an extraordinary product and still lose to an adequate product with an established audience, a familiar workflow, a trusted brand, and a simpler path to purchase. This is not always fair, but it is how markets work.
The answer is not to give up and assume incumbents will win everything. Lower production costs also let focused teams compete in places that were previously uneconomic. A small company can serve a narrow profession, integrate deeply with a specialist workflow, and learn faster from a concentrated group of customers.
But it has to choose its distribution advantage deliberately.
That advantage might be a community the founders already belong to. It might be proprietary access to a workflow, a trusted relationship with a profession, a channel partnership, an unusually strong service layer, or a product loop that naturally creates collaboration and referrals.
“We will launch and see” is not a distribution strategy.
Build the Path to Demand Before the Product
Before committing to another application, I would ask five questions:
- Which specific customer already feels this problem strongly enough to change behavior?
- Where do those customers already gather, buy, and work?
- What trust must exist before they will connect their data or depend on the result?
- How quickly can they reach a valuable first outcome?
- What makes continued use—or sharing the product with someone else—the natural next step?
If the answers are vague, a faster build will only produce a faster lesson in indifference.
The age of AI-generated software will create far more products. That is almost certain. It will not create proportionally more attention, trust, or customer time.
Those remain scarce.
And in markets, the scarce resource eventually captures the value.
Source
The research discussed here is “Writing Code vs. Shipping Code: Productivity Effects Across Generations of AI Coding Tools,” NBER Working Paper 35275 by Mert Demirer, Leon Musolff, and Liyuan Yang. It is a May 2026 working paper circulated for discussion and has not yet been peer-reviewed.