About AI in Software Dev Teams

AI Breakthrough

I still remember my first contact with generative AI. It was two years ago, during the summer of 2022, when a then-novel tool—DALL·E mini—became viral on Twitter. At first, it was just a fun experiment: you could create amusing images such as “Mickey Mouse in a Shrek movie” or “Ronald McDonald fighting Colonel Sanders.”

A few weeks later, I discovered ChatGPT, at this moment in its GPT-3 version. It wasn’t very impressive at the time—it could code simple tasks and help debug errors if you already understood the application and tech stack. However, a couple of months later, when GPT-4 was released, the real breakthrough began. This new model was capable of building standalone functions and writing the code that no one really wants to write. Although it occasionally made mistakes or misspelled variable names, its overall performance was remarkable.

About a year and a half later, a new model (O1) emerged as a true game changer. It could take a few instructions and generate complete files, optimize the performance of large functions, and—thanks to progressively larger context windows—even design and write full application features. The current SOA models (03-mini-high and R1) build upon this foundation, consistently producing plug-and-play code, generating complex queries from natural language descriptions, and refactoring large files with minimal errors.

How AI Has Impacted Workflows

In just three years, we have transitioned from a silly novelty application to having a reliable, versatile software engineer that never gets tired and remembers virtually every syntax detail across languages and frameworks. This evolution has transformed teams in several key areas. In my team, I have identified six main use cases in which AI has greatly improved productivity:

Coding

The most obvious benefit is in coding. Instead of spending 5–20 minutes writing a function, you can simply describe its input, desired behavior, and output—and have the code generated in a minute. Also when debugging with today’s SOA models, you can even include a stack trace along with your code, and more than half the time receive a fix or useful guidance on the error. Additionally, AI can handle the tedious parts of coding, such as exposing multiple functions through an API.

Querying

Beyond the creative queries that manipulate data in interesting ways, much of SQL work is repetitive. Who wants to manually define six columns or write a query with three joins? Large language models excel at these routine tasks.

Getting Context

As a team lead, I don’t have every detail of our applications memorized. I often need to quickly understand sections of code or make modifications when the responsible team member is unavailable. With tools like O1, you can paste files or components and ask, “How does this work?” or “Why does this happen? How can I add this?” to gain the necessary context.

Learning New Technologies

Not long ago, getting started with a new framework or language required reading quickstart guides, poring over documentation, or watching lengthy videos before you could even build a “Hello World” application. Today, you can simply ask, “How do I create a simple application using [target technology]?” and iteratively refine your questions until you achieve your goal. While traditional methods still have their place, this approach is significantly faster.

MVPs (Minimal Viable Products)

MVPs are all about speed. You don’t need flawless code, perfect scalability, or meticulously optimized performance—just a quick release to test if a solution is viable. In this respect, AI can take you from whiteboard to production in less than an hour.

Technical Writing

Just try giving ChatGPT a file and asking it to “Write a README.md file explaining the application’s purpose, how it works (a high-level overview), and include detailed descriptions for every function along with useful annotations.” I tested this with GPT-4.5 for its high-level language capabilities and with O1 for its coding expertise. O1 outperformed GPT-4.5 by delivering a comprehensive document—what would have taken 90–120 minutes to write was completed in just 5 minutes.

Current Limitations of AI

Currently, to make these models truly useful, you need to be very clear about what you want to achieve. Instead of simply saying, “I need a report on my best locations,” you must specify: “Give me a [desired technology] script that queries this [table structure] to identify my top 5 locations based on [criteria] and outputs the result in [file format].” While this level of detail might seem trivial, the information within the brackets is not common knowledge for someone without a technical background.

Additionally, AI is not very adept at making design or architectural decisions. Even for small-scale applications, it struggles to devise an application from first principles. In my experience, it tends to follow the direction you suggest rather than offering alternative solutions. You might be playing in Cursor for hours, building an application, yet it will rarely indicate that you might need essential components like authentication, a database, or a navigation menu—unless you explicitly ask for them. Thus, gaining hands-on experience remains essential before relying entirely on an AI copilot.

Impact on Future Developers

I don’t believe that AI will create a generation of less capable computer science graduates. In the past, some students copied classmates’ assignments or did just enough to pass without truly learning. Those same individuals might now simply copy and paste LLM-generated responses. Similarly, students who once spent an all-nighter building a not needed graphical interface for his C++ console application can now produce a React application where the assignment where a basic HTML CRUD assignment. Ultimately, curious and motivated students will use AI to achieve even more, while those who are less engaged may continue to get by.

Regarding the “AI will take your job” narrative, every new technology was once expected to eliminate the need for developers—but that never happened. SQL was designed so that your secretary could manipulate data; instead, it led to database courses and specialized database analyst roles. COBOL, once touted as a language anyone could use to code in plain English, gave rise to specialized COBOL developer roles. Content management systems (CMS) were expected to replace webmasters, but they instead created dedicated web development agencies and roles for WordPress/Shopify developers. In the same way, LLMs will create new roles and empower current and future developers to build better and faster.

Today, you can build a React static application in just a couple of minutes—a task that used to take hours a few years ago, days a decade ago, and even weeks at the start of the millennium.

Final Thoughts

As an example, this site was built with Next.js. Do I know Next.js? Not really. The last time I built a React application from scratch, Create React App was still in use. However, with just a couple of prompts, I was able to build this blog. And dont get me wrong, the shitty minimalistic look its intentionall, aligning with my preference for backend and systems design. Also, while I wrote this entry entirely, as a non-native English speaker I used AI to "correct any orthographical, grammatical, and stylistic errors, as well as to make any changes required to improve clarity or better structure the information." By sacrificing some of my personal writing style, I was able to publish a more coherent and polished text.

Note: All of these reflections are based on current commercial models. APIs represent an entirely separate domain that would warrant its own discussion.