I have spent years in software, working across product, the commercial side and building companies. I understand enough about technology to discuss APIs, infrastructure and how a product should work. I am not a conventional software engineer.

That combination used to leave a fairly clear boundary around what I could build myself. I could develop an idea, research it and define the product, but turning it into working software meant finding people with the necessary technical skills. Often it meant finding the money to pay for them first.

AI is changing that sequence. I can now investigate an idea, build a prototype and test parts of it much further before I need a conventional development team. That is the change I find most exciting.

The bottleneck used to be translation

A software idea needs several kinds of knowledge. Product, design, front end, back end, infrastructure and engineering leadership all contribute something different. A founder has to bring enough of those capabilities together to make progress.

Each hand-off also takes time. Someone explains the idea, someone else interprets it, and questions come back before the next part can start. Even a small experiment can consume scarce development capacity.

With AI, I can describe what I am trying to achieve, inspect the result and work through the next decision. I still need to understand the problem. What has changed is how much technical work I can direct without first assembling a team.

Capability is moving up the stack

I think about the old barriers as knowledge, capability and capital. They were connected. If I did not have a particular capability, I needed to hire it, and hiring it required money.

AI coding tools give me access to a much wider range of technical work. I can use them across different parts of an application and, with the right integrations, connect development to infrastructure, previews and testing.

The useful skill is increasingly knowing how to describe the outcome, question the implementation and recognise when I need deeper expertise. My product and commercial experience help me do that. They do not make me an expert in every part of the system.

Building is not the same as engineering

I would not want this argument reduced to the idea that engineering expertise no longer matters. Producing code and engineering a dependable system are different responsibilities.

  • Architecture determines how the parts fit together.
  • Security and data integrity affect what can safely be released.
  • Testing needs to cover more than the path that happened to work once.
  • Maintenance matters after the first version is built.

A prototype can be useful before all of those questions are resolved, provided I am clear about what it is. The danger is mistaking something that looks finished for something that has been properly tested.

AI lets me cross more of the distance towards a working product. It also makes it important to recognise where my own knowledge runs out.

Entrepreneurship gets cheaper

Research spikes are one of the clearest changes in my own work. I remember investigations that could take two or more weeks and involve several developers. I can now do some early feasibility work with AI in around 20 minutes.

Those are personal comparisons, not like-for-like benchmarks. The depth of an investigation can differ substantially. Twenty minutes of early exploration does not replace every form of technical diligence.

Even with that qualification, the effect on my behaviour is significant. If I can establish a technical baseline and test a few assumptions quickly, ideas become worth examining that I might previously have left alone.

This changes when I need capital. I may still need investment to build a reliable business, hire specialists or scale a product. But I do not necessarily need it simply to discover whether the first version of an idea has any merit.

Taste becomes more valuable

Cheaper execution does not make every idea good. It makes the decision about what to build more important.

I can generate several approaches quickly, but somebody still needs to judge them:

  • Does this solve a problem people care about?
  • Is the workflow better, or merely different?
  • What should be removed rather than added?
  • What would make the result trustworthy enough to use?

My experience in product and sales becomes useful here. Understanding a customer problem and being able to explain its value are still necessary. AI gives those skills more room to operate; it does not supply a business case automatically.

Teams may become strangely shaped

I do not expect the disciplines involved in software development to disappear. I do expect more people to work across the boundaries between them.

A product person can test an interaction. A designer can explore an implementation. A founder can develop an internal tool before deciding whether it deserves a larger investment.

That may allow a smaller team to attempt more, but it will depend on the work and the people involved. Broad output should not be confused with universal expertise. Someone still needs to take responsibility for the parts that carry real consequences.

The new literacy is supervision

The ability to ask for code is only a starting point. I need to be able to challenge what has been created and decide whether the next step is safe.

  • Break the problem into pieces that can be evaluated.
  • Test assumptions rather than accepting a plausible explanation.
  • Ask what changed and why.
  • Check the important failure cases.
  • Bring in specialist help where the stakes justify it.

I can also use AI to research a problem, produce an execution plan and test parts of that plan with connected tools. That makes investigation more tangible, but I still need to distinguish research, prototype validation and production engineering.

Ideas have fewer excuses

The biggest change for me is that fewer ideas have to stop at the point where I would previously have needed someone else to build the first version.

I can find out more before asking for money or committing a team. Some ideas will fail that early examination, which is useful. Others will become concrete enough to deserve more attention.

I still have to choose well. What I can increasingly do is make that choice with something in front of me that works, rather than a document describing what I hope might work.

Source: OpenAI — GPT-5.6