When building three candidate designs costs roughly what building one used to, the obvious reading is that more gets built. The consequential change is elsewhere: the organisation now has to choose between working options, and that is a different problem from producing them.
What the evidence supports
An experiment comparing parallel and serial prototyping among 33 completers, on banner-ad design, holding prototype count and allotted time constant, reported better design outcomes in the parallel condition.
The scope is narrow and worth stating. Different domain, largely student participants, no AI involved, and no test of enterprise cost. The study also changed the timing and grouping of feedback, so it does not isolate parallelism as such, and it establishes no optimal number of alternatives.
What it supports is the direction: comparing alternatives against each other, with feedback structured accordingly, produced better outcomes than developing them in sequence.
Variants are not alternatives
The failure that cheap generation invites is counting output instead of ideas.
Alternatives should differ on a meaningful mechanism or user proposition, including an existing or non-build option where one applies. Several variants sharing the same model, the same prompt assumptions, and the same source data can share a common failure, and generating them establishes nothing about that failure.
A team comparing three interface approaches that all rest on the same unverified eligibility assumption has one untested assumption and three demonstrations of it. Investigating the assumption is worth more than any comparison between the three, and the comparison actively conceals it by making the shared part invisible.
So the count that matters is distinct tested ideas, and the discipline is to state the uncertainty first and derive the alternatives from it.
The budget moved to evaluation
Each additional candidate carries generation, setup, evaluation, integration, opportunity, and disposal cost. Generation is the term that collapsed. The others did not.
An additional candidate is worth considering when the expected improvement in the decision justifies those incremental costs — a decision heuristic rather than a calibrated formula.
This is why the constraint is review capacity. Comparable feedback conditions, selection criteria recorded before a favourite emerges, and retained rejected alternatives with their reasons are all work performed by people, and the people are the same ones who were the bottleneck before.
The organisational consequence is that a team newly able to produce five candidates and still able to evaluate one has not gained five times the options. It has gained a queue and a selection problem it is not equipped to solve, which will be resolved by whoever demonstrates most persuasively.
Cheap parity is not parity
The adjacent claim is that cheap generation makes replacement attractive, because reaching parity with an existing system is now affordable.
Reconstructing existing behaviour is difficult, and some of that behaviour is no longer wanted. Both halves matter: the reconstruction is hard, and part of what would be reconstructed should be discarded.
A generator recreating visible screens quickly while missing a month-end reconciliation rule has produced something that looks equivalent. Parity is unestablished, and the replacement comparison has to be revised rather than accelerated.
The comparison that decides it uses forward costs: required behaviour, intentionally retired behaviour, still-uncertain behaviour, and the future verification, migration, support, and operating effort — separately from code production. Coexistence costs and the consequences of changing data or external interfaces belong in it.
Four options rather than two: keep with bounded remediation, retain only a component, replace incrementally, or rebuild and cut over. An artifact can be disposable while its data, user promises, and integration obligations persist, and a well-understood small component can be worth keeping even where recreating it is cheap.
What the sequencing argument becomes
The discipline of establishing that something works before committing to scale it survives, with one term repriced.
Cheap generation changes the cost of running the experiment. It establishes nothing about demand, about a workable operating model, or about justified scale. A team that generates a complete mock workflow cheaply, with no evidence that anyone needs it, holds an artifact rather than a finding.
The new difficulty is that a cheap complete-looking artifact is more persuasive than a cheap obviously-partial one. The pressure to skip the test rose at the same moment the test got cheaper.
The rule
What stays fixed is that a decision improves when the alternatives differ in ways that matter and someone competent compares them. What changes is the cost of producing the alternatives, and the comparison capacity now sets the ceiling.
Not to be confused with
A measured saving. No cost advantage for parallel work in a software setting is established, and the supporting experiment is from a different domain with different participants.
More options being better. Beyond review capacity, additional candidates degrade the decision by consuming the attention that would have distinguished the ones already produced.