Deciding What to Build When Building Is Cheap

Almost every conversation about AI in software is a conversation about speed: how to build faster, ship more, generate at scale. It is the wrong conversation to be having near the top of an organisation. When building gets cheap, the constraint does not disappear — it moves upstream, to the decision. The scarce thing is no longer the capacity to build; it is the judgement to know what is worth building, and the nerve to say no to the rest. Discovery is becoming the new development.

The asymmetry nobody priced in

Here is the part the speed narrative quietly skips: building got cheaper, but being wrong did not. The cost of constructing the wrong feature has fallen; the cost of having built it — the support, the surface area, the maintenance, the cognitive load it adds to everything around it, the strategic attention it absorbs — has not. Cheap construction lowers one line on the bill and leaves the larger, slower line untouched.

And most things we build are wrong, in the only sense that matters: they do not move the outcome. In well-optimised products, controlled experiments at Bing and Google find only about 10 to 20 percent of changes positively move their target metric; across mature companies the working rule is roughly a third positive, a third flat, a third negative. These are figures for instrumented, A/B-tested changes in mature products, not a law of all software — but the direction is unambiguous and it is decades old. Organisations are routinely, structurally wrong about what their customers want. Experimentation exists precisely because we cannot tell the good ideas from the bad ones in advance.

Abundance industrialises being wrong

Now make building nearly free and remove the discipline that decides what is worth shipping. You do not get a leaner product; you get a bloated one. As the experimentation literature puts it bluntly: without experimentation, most of these negligible-impact changes would simply be deployed, creating a product that is constantly changing without adding real customer value. Cheap building, absent judgement, does not reduce waste. It manufactures it at scale.

The 2026 evidence makes this concrete, and it is more honest than the marketing. AI is an amplifier, not an accelerator: Google’s DORA research finds its primary effect is to magnify an organisation’s existing strengths and weaknesses, so a disciplined org gets more disciplined output and an undisciplined one gets more mess, faster. The speed itself is not even reliable — a controlled trial by METR found experienced developers were 19 percent slower with early-2025 AI tools on code they knew well, while believing they were 20 percent faster (a result bounded to those models and familiar codebases, and likely narrowing since, but a warning against trusting the feeling of speed). And the cost does not vanish; it migrates. Early studies of agent-authored pull requests find markedly more redundant code accepted with fewer objections from reviewers — debt that accrues quietly downstream in maintenance, not a bill that anyone sees this quarter.

Put those together and the contrarian conclusion is the honest one. Cheap building does not lower the cost of being wrong. It raises it — because abundance multiplies low-value output, and the consequences land later and out of sight.

So judgement becomes the scarce factor

The practitioner framing for this is Marty Cagan’s: as the cost of delivery drops, the bottleneck and the competitive advantage move to discovery — to deciding what to build. We agree with the direction, and we will go one step past it. The leverage is not in discovery as an activity you add; it is in the specific decision discovery exists to make: the decision not to build. When construction was expensive, the cost of building gatekept your portfolio for you — you simply could not afford to build everything, so something had to be cut. Take that constraint away and nothing gets cut by default. The no-build decision, which used to be made for you by scarcity, now has to be made on purpose.

The Conviction Gap

We call the space this opens the Conviction Gap: the distance between what you can now build and what you have evidence to justify scaling. Cheap building widens it relentlessly — capability shoots up while conviction stays exactly as hard-won as it ever was. You can build almost anything; you still have no more reason than before to believe most of it will work. The gap is not a tooling problem. It is a judgement problem, and it is where the work now is.

Closing it does not mean building less for its own sake. It means separating two thresholds that cheap building tempts you to collapse into one. The first is the threshold to learn: deliberately low. Building to learn — a prototype, a fake door, a throwaway — is exactly what cheap construction is good for, and you should do more of it, not less. The second is the threshold to earn: deliberately high. Building to earn — committing real production surface area, real maintenance, real customer dependence — should require evidence, and the bar for it should not fall just because the keystrokes got cheaper. The failure mode of the AI era is collapsing the two: treating everything you can cheaply build as something you should ship and support.

• Separate the two thresholds explicitly. Make “build to learn” cheap, fast and disposable — and make “build to earn” require evidence, every time, regardless of how easy it was to produce.
• Fund discovery, not just delivery. The bottleneck moved upstream; the budget and the senior attention have to move with it, or you will simply build the wrong things more efficiently.
• Make the no-build decision legitimate and visible. Killing or declining an idea should be a celebrated outcome of good judgement, not a quiet admission of failure — otherwise everything that can be built, will be.
• Measure what you decided not to build. Track a no-build rate the way you track delivery throughput. A team that never kills anything is not disciplined; it is a feature factory with cheaper machinery.

As building got cheaper, organisations did not become more selective — they became less. The friction that used to force prioritisation quietly disappeared, and the question stopped being “can we afford to build this?” and became “why wouldn’t we?” The teams pulling ahead use cheap building to learn faster while getting more disciplined about what earns a permanent place: the advantage is not building more things, but accumulating fewer wrong ones.

But isn’t cheap building an argument to build more?

It is — for learning, and we should be honest about that tension rather than wave it away. The lean instinct is right that cheap experimentation lets you test more ideas against reality instead of arguing about them in a room, and that disposable prototypes often beat up-front analysis. Nothing here contradicts that. The error is applying the build-more logic to the wrong threshold: using cheap construction as a reason to ship and commit to more, rather than to learn from more. “Do the wrong things faster” is only a virtue while the things are disposable. The reconciliation is the two thresholds: lower the bar to build-to-learn as far as you like; hold the bar to build-to-earn exactly where the evidence puts it.

The lesson most will get wrong

We suspect many organisations are about to learn the wrong lesson from AI. They will see that building got cheaper and conclude that they should build more. The teams that create lasting advantage will reach the opposite conclusion: they will use cheap building to increase learning, not commitment — making experimentation abundant and production deliberate. And they will discover that the rarest capability in an age of infinite construction is still the oldest one: judgement.