- Published on
The Decisions We Made When Software Was Expensive
- Authors
- Name
- Iván González Sáiz
- @dreamingechoes
Table of contents6 sections
What we can afford to build shapes which ideas get a chance. There can be a useful idea, a clear need, and still no room to explore it alongside everything the team has already committed to.
I enjoy how much easier coding agents make it to explore some ideas. Having spent time both building software and leading engineering teams, I wonder which of the decisions around that work might deserve another look. If part of the implementation takes less effort, some reasons for setting an idea aside may no longer hold.
In the first article in this series, I explored how we might know whether AI was making a team better at delivery. Here the starting point is that it has: the team delivers more software, maintains quality, and has capacity it did not have before. That gives us more possibilities to choose from.
Some things were worth building all along, but we could not afford to find out. Others may have looked worthwhile mainly because they were the best use of the limited time we had. The extra capacity could go straight into the next item on the roadmap, or give us a reason to revisit the choices behind it.
The ideas we could not justify
When I look at this from the leadership side, the difficult part is that saying no to an idea does not necessarily mean doubting its value. You can understand the need, believe the proposal would help, and still be unable to justify taking the team away from something else for a month. The need stays real after the decision. It simply does not make it onto the plan.
The benefit of faster engineering may go well beyond getting through the work we already selected. There are ideas we set aside not because they lacked value, but because we couldn't justify the effort they required.
If the cost comes down, some of those decisions deserve another look. The original answer may have been entirely reasonable. It might even still be right, because implementation was not the only cost. But continuing to say no without reconsidering why would miss something I find exciting about working with agents: more ideas can become worth exploring without first becoming large commitments.
In Cui, Demirer and colleagues' experiments, companies randomly assigned access to Copilot. The researchers estimated that using it increased completed development tasks, measured through pull requests. Whether that work helped customers was outside the experiment. That is the part I want to follow.
Sometimes the answer is straightforward. A known problem gets fixed sooner and people benefit sooner. We do not need to turn every overdue fix into a discovery exercise. Things get more complicated when we know exactly how to build something, but we are much less certain why we should. Making it cheaper gives us a reason to revisit the idea, not a reason to stop asking.
The constraint beyond implementation
The uncomfortable possibility is that a team might reach the end of what the organization understands well enough to build before it runs out of capacity to build it. That will not be the limiting factor for every team. A team with years of well-understood customer needs waiting behind an implementation backlog may benefit from faster engineering for a long time.
Suppose a team can now implement several alternatives before its next research session. If the unanswered question is which version people can use, earlier prototypes are helpful. If it is whether the problem matters enough for someone to change their habits, the team may still need access to customers, a workable distribution channel, and time to observe what they do. More versions do not automatically supply any of those things. A calendar full of research sessions can become the limiting factor; so can a product with too few active users to evaluate many experiments at once.
A convincing prototype can make that distinction surprisingly easy to overlook. You can see that the interaction works, that the data is there, that the team could finish it. Whether someone will care enough to use it remains much harder to see. A faster implementation can help you get closer to that answer, provided someone still has time to listen to it.
The 2025 DORA report offers some support for that concern. Its survey model showed a more favorable relationship between AI adoption and reported team performance at higher levels of user focus. That does not establish causation, and performance here is respondents' assessment of how effectively their team works, not company profit.
None of this makes Product or Design the department now failing to keep up. Engineering can help investigate a need, challenge an assumption or make an alternative cheap enough to try. AI can help with research synthesis and prototyping too. I would like that to bring us into more of the same conversations, rather than leave one group producing specifications faster so another can implement them faster. We may have more room to work out the problem together.
The cheaper decision to be wrong
I keep coming back to the difference between building something to find out and building something because we have already decided.
A lower implementation cost makes the first option more attractive. An idea that could not justify a month might justify a small test. Requiring the same certainty for both would throw away part of the gain. Sometimes the sensible response to cheaper software is to try an idea with less evidence, provided the attempt is small enough to abandon and capable of teaching you something.
But the cost of writing the first version is only part of the commitment. Imagine adding a new reporting option because an agent makes it easy. Once customers depend on it, the option needs documentation, support, compatible behaviour and decisions about its future. You can finish the implementation in an afternoon and still acquire years of obligations. Even with good code and stable releases, a product can become harder to understand through a series of individually inexpensive additions.
That makes the boundary between a test and a commitment more consequential. A prototype shown to a few people can reveal that the underlying idea is wrong. A feature rolled out widely, included in sales material and built into customer workflows is harder to put down. Cheap implementation gives you more room to be wrong early; it does not make every later consequence equally cheap.
Before starting, there is still the question of what happens if we learn that the idea is not useful. Can we put it down? Or will the fact that it already exists become a reason to keep adding to it? It is possible to call something an experiment while treating its continuation as the only acceptable result.
A small attempt may be worth trying without the confidence we used to need before committing a month. But making room for more experiments also means making room to stop them, change the question, or discover that the useful answer requires no new feature at all. Otherwise we can execute an uncertain decision much faster and end up more committed to it before we have learned whether it was a good one.
The capacity that already has a destination
A backlog makes extra capacity look easy to allocate. There is always something next. The danger is less that nobody knows what to do with the time than that the answer arrives before anyone considers what has changed.
You could use it to explore more product ideas, or finally address the needs that never looked large enough to justify their engineering cost. You could spend it reducing operational risk, simplifying a difficult part of the system, or investigating a strategic question without first finding a spare evening. More planned features are one option among several.
And this becomes more complicated when we start translating the saved time into money. An engineer's estimate of hours saved is not yet capacity the team can reliably assign elsewhere. Even when the team can use the time elsewhere, the salary bill stays where it was.
Imagine agents make it possible to complete a long-postponed change to a costly background job. A lower infrastructure bill is a saving we can point to. A product experiment may leave the bill unchanged and still change the team's next decision. If the time goes into a feature nobody needs and the team learns nothing that affects what follows, the released capacity has been used without producing additional value. The engineering improvement can be real in all three cases.
This is why I am reluctant to call capacity a saving before knowing where it went. Or rather, knowing where it went is only the start: a team can describe its new work clearly and still be uncertain whether it was worthwhile.
The startup experiment by Kim, Kim and Koning is an encouraging counterpoint. All 515 firms received AI tools and training; a randomly selected group also studied how to reorganize work around AI. By the end of the three-month accelerator, that group was more likely to have paying customers. The study relied on founders' reports and tested a broader intervention than faster coding.
What I take from it is that the way a firm finds uses for AI can make a difference to the result. There is room to act here. A team does not have to accept that any benefit beyond implementation is too distant to influence, or that the next roadmap item is automatically the best destination for the time.
The choice affects the people doing the work too. Space for reliability or investigation makes for a different day from filling every available hour with another commitment. The previous article explored what people want to keep in their work; that still matters. The question here is whether we make a choice about the new capacity before the existing plan makes it for us.
The business result that takes longer
An implementation finishes on a particular day. Its consequences rarely arrive together.
A customer may only need the new capability at month-end. A change that makes a service safer might matter most during an incident that has not happened. Maintenance work could make future changes easier without moving this quarter's revenue. Less time spent recovering from failures, fewer manual workarounds and a product more people can use are valuable outcomes, even when their financial effects are indirect.
That does not make an absence of results reassuring. If a feature remains unused after the intended customers have had a fair opportunity to find and try it, the original belief deserves another look. Returning to work after it ships matters more when the team can already be several ideas further on. Waiting for a benefit that has a reason to arrive later is different from leaving an unsuccessful initiative permanently under evaluation.
McKinsey's August 2026 State of AI survey captures some of that distance. Eighty percent of respondents reported better individual productivity; 37% attributed a positive contribution to their organization's earnings before interest and taxes to AI. These are self-reports across industries, answering different questions, not a conversion rate from saved hours to profit. Respondents also reported benefits such as customer satisfaction. A lot can happen between finishing the work and seeing a benefit in company accounts.
For a team shipping ordinary software with AI assistance, attribution is harder. If a useful feature ships earlier and adoption grows, pricing, distribution or a change in the market may also have contributed. Even where the feature clearly helps, the narrower AI benefit may be the extra time customers had access to it, or the opportunity to use engineering capacity elsewhere. Crediting AI with the feature's entire lifetime value would make too much of the tool; ignoring the earlier access would make too little.
Those uncertainties deserve different responses. A benefit that needs more time is different from one already visible in operations, even if neither appears in revenue. Some work will turn out not to help, and some useful work will still have cost more than its benefit can justify. We can acknowledge the better engineering without treating it as evidence that every decision made with that capacity was worthwhile.
This post is part of the series Software Engineering in the AI Era — exploring how AI is reshaping software engineering beyond the editor: delivery, leadership, business outcomes, and the people behind the work.
Final thoughts
I keep returning to the ideas a team could not justify before. Some may still be wrong for the product. Some may have been waiting for a less expensive way to find out. Being able to revisit them is part of the appeal of working with agents, and I would not want to lose that possibility inside a larger commitment to finish the same roadmap sooner.
Having more capacity does not make the choice easy. It does make me less comfortable treating an old decision as settled when one of the reasons for it has changed. We can be pleased that the team is faster and still ask whether the plan reflects what has become possible.
Perhaps some of that additional capacity is worth spending on conversations we couldn't make room for before. Going back to ideas we set aside, understanding why we made those decisions, and asking whether the reasons still hold.
Enjoyed this article?
Subscribe via RSS
Follow along in your favourite feed reader. Every new post lands there as soon as it's published — no account needed.