The explosion in AI chatbots over the past few years is having a massive impact on software development. At the moment hot takes are ranging from “developers will all be replaced by AI” to “this is an over-hyped grift that will crash & burn”.

As the owner of a 13 year old software development agency I wanted to write about how AI impacts our development processes in the here & now.

My main motivations in writing this are to communicate with our clients about how things are changing, & lock in a snapshot so I can nostalgically read this post five years from now.

I’m going to use two analogies to explain the current state of play:

AI as an Author

Imagine feeding an AI LLM the contents of a library containing every fiction book written in the past 100 years & then asking it to write a novel. As we know, the AI will have no problem generating pages upon pages of grammatically correct sentences. However, it will fall over in the following ways:

• Creativity: AI’s aren’t creative - they can refactor human creativity, but that’s not the same. The topic of the novel will probably be an average of the topics of all of the novels it has read, with common tropes, characters & themes.
  
  
• Quality vs Quantity: When you read “every fiction book written in the past 100 years” it’s easy to think about famous books like “The Hobbit” or “1984”. But the reality is, out of the millions of books that have been written in this time period only a fraction of them are well written. An AI trained on all of these books is going to be reading a lot of garbage.
  
  
• Lived Human Experience: Creating a believable fictional world takes a massive amount of understanding of the world. Most novels are about people & our understanding of humanity is built from thousands of years of evolution & decades of lived experience. AI’s don’t have the breadth of understanding or the intuition to comprehend human interactions.

So while an AI is capable of writing a novel about a young woman caught in a love triangle between a vampire & werewolf in under an hour for less than $5, it would be the literary equivalent of wallpaper paste.

AI as a Builder

Just as AI can write a novel in less than an hour, it’s very very fast at writing code. But there’s more to software development than writing code.

When talking to clients I often use house building analogies when talking about software development. At a simplified level there are architects, engineers & builders - who all work together to build a house. Currently there are AI coding agents that act as very enthusiastic, very energetic, very fast builders. Think of a house frame going up in under 5 minutes & you get the idea.

But, like our AI author, these AI coders have been taught on a library of trillions of lines of code. Like our hypothetical library a lot of that code is absolute garbage - outdated, buggy & littered with security issues. However - there are libraries of reference material on various programming languages & API which give more context on what good code looks like.

In the real world, our family recently had a block wall built from plans created by a structural engineer. Workers excavated dirt, built framing, re-routed electrical wiring, poured concrete & laid blocks. Imagine having a team of AI robots who were trained on the building code & libraries of building reference material building the same block wall.

At face value this sounds great - a team of block laying robots could out-pace humans & all they’d need to do is follow the engineering plan & their own training - right? Which brings me to:

• Plans Exposed to Reality: We had an excellent structural engineer, but during the prep work the builders identified an issue around how the planned wall would impact an existing structure. Had they built exactly to plan it would have generated a domino effect of disruptive & expensive additional work - which they stopped by talking amongst themselves & calling the engineer to work out a solution.
  
  
• The World is Chaos: Ground is never flat, buildings are never square & there is no existing diagram of the electrical wiring that needed rerouting. Humans are expert improvisers, even down to how we subconsciously adjust our gait when walking over uneven ground, or change our breathing based on exertion, air composition & temperature.

Most successful uses of machine automation exist in tightly controlled environments where there are no surprises & the plan can be applied exactly how it was written.

Those two analogies bring us to:

AI in Software Development

You can probably see where I’m heading with this. While parts of software development take place within heavily codified logic based frameworks the important parts interact with deeply complex humans in a chaotic world.

We have enthusiastic, energetic, mind-bendingly fast builders - but without guidance they can & will generate a lot of code that’s fit to plan, but not fit to use. They will also make mistakes, get distracted, go down rabbit holes & use bad code from their training data - all at breakneck pace.

So AI coders need firm, detailed documentation to write quality code. But the last coding revolution - the move to “agile” development - found success in moving from building to rigid blueprints to a flexible building process where the project evolves as new functionality, requirements & limitations surface. In 2020 I wrote about the dangers of developing to a rigid software scope. AI agents building to spec risks losing the flexibility that has defined the past decade of development.

UpShift’s current development process looks roughly like this:

1. Architecture & Software Engineering: As a team of humans with decades of software development experience & even more decades of human experience we work out an architectural plan of the software we’re building & an engineering plan for how it will be implemented.
   
   
2. Prompt Engineering: You may have heard this term thrown around. It’s the science & art of giving an AI agent the information it needs to build what you want it to build.
   
   
3. AI Coder Goes Vroom: Even a couple of years into the AI coding revolution I’m still astounded watching Claude Code write code several times faster than a human. Methods created in seconds, hours of refactoring in minutes, but…
   
   
4. Code Review & Quality Control: Real humans now need to check the AI generated code. My news feeds are overflowing with cautionary tales of people & companies that let AI write their code, didn’t check it & discovered catastrophic issues when it had been published into production.

Our bottleneck used to be writing the code - but now our bottlenecks are the engineering & the code review.

Yes, we are building software a lot faster than we were two years ago. But we’re heading towards a point where the only way to increase development speed is to let unsupervised builders spew out unreviewed code - which raises the risk of bugs, security holes & structural issues in our applications.

Looking to the Future

I could easily write another thousands words on the future - but two things jump out at me:

• Can AI do the Architecture, Engineering & Review: The AI industry is telling us that their products are capable of managing the entire development process. But this has yet to reach a point where any professional software developer would trust leaving AI to do everything. How quickly this promised future arrives, if ever comes down to AI evolution.
  
  
• The Rising Cost of AI Development: At the moment AI companies are bleeding money while building their customer base. Following the standard enshittifaction product cycle there will be a point where the AI companies will start to raise their prices. Based on the software industry's reliance on AI agentic coding services we’ll probably be the first to experience this rise in cost (spoiler alert: it’s already happening). My suspicion is that monthly subscriptions will go from hundreds to thousands of dollars.

Part of the reason I’ve built a career around software development is I find the ever changing pace of technology exhilarating & challenging. The past three years have seen more changes than the previous three decades.

Having a friendly robot that writes code at an eye watering speed is both awesome & riddled with risk. Humans need to collaborate with machines in a process which makes the best use of human/ machine strengths & weaknesses. How the next three years track is going to be a wild & interesting ride.

Footnote: "AI as an Author" is already happening in genres like romance - there's a good article here & a podcast interview here about how AI is impacting writing novels.

Banner Photo by Artturi Jalli