AI Wrote the Code—Now Developers are Stuck Fixing It

It is a general belief that AI-driven tools help accelerate workflows for developers. While some developers may feel like Lightning McQueen, a fictional character in Pixar’s Cars franchise known for racing, others are encountering more bumps and roadblocks than anticipated.

In a blog post, Birgitta Boeckeler, global lead for AI-assisted software delivery at Thoughtworks, explained how she had to intervene, correct, and steer the flow of AI-driven development while using Cursor, Windsurf, and Cline. That slowed down the speed of the development process.

So, what are the takeaways from her findings? How do other developers feel about the same issue? More broadly, can AI-assisted coding sometimes hinder productivity rather than enhance it in certain contexts?

Where the Problem Occurs

Boeckeler began by expressing her appreciation for the recent progress in IDE integration. She noted how those integrations have made it easier and faster to get things done—from executing tests and fixing errors on the spot, to conducting web research, and more.

As she progressed through the AI-assisted coding sessions, Boeckeler found herself frequently needing to guide the tools and correct their suggestions, and sometimes, discarding the changes proposed by the AI altogether.

So much so that she mentioned, “We’re still far away from AI writing code autonomously for non-trivial tasks.”

Boeckeler explained the problem by categorising it into three areas of impact. 

Coding and Commit Time

To begin with, Boeckeler faced a common issue: receiving code that simply did not work. She then had to decide whether she would do it herself or start a new AI session to retry, which could potentially take more time.

In addition, it was observed that AI misdiagnoses a problem. She shared an example where the AI assumed that a Docker build issue was due to architecture settings. In reality, however, the issue originated from copying the wrong packages built for the wrong architecture.

Koushik Bhargava, a member of technical staff at Salesforce, told AIM, “I have experienced AI hallucinations a few times while trying to resolve some issues, wasting more time in the process. Often, I was able to resolve those issues with one simple Google search.”

Breaking the Team Flow

As per Boeckeler’s observation, AI focuses on the overall objective instead of focusing on building blocks of functionalities. She highlighted that this risks wasting large upfront work before realising the problem.

Moreover, AI has been noticed to take a brute-force approach to solving issues rather than analysing the root cause. This causes a delay in the solution coming into effect. 

An example included a scenario where the AI encountered a memory error during a Docker build. To resolve this, it increased the memory settings rather than investigating the problem of the error.

She added that the developer workflow can become more complicated than it should be at times, sometimes due to incomplete context in human prompts.

Long-term Maintainability 

AI-generated code does not focus on modularity, which makes its reusability difficult. This could be because AI does not always have the entire context of the project, potentially making it a problem for future code maintenance.

In some cases, AI generates overly complex and unnecessary code. In contrast, simpler code makes it easy to maintain.

So, a Speed Boost With a Catch?

In an exclusive chat with AIM, Mehul Gupta, data scientist at DBS Bank, said, “AI coding has definitely sped up my workflow. Boilerplate code? Done in seconds. It cuts down development time massively.”

“After years of programming, I’ve developed a sixth sense for where AI tends to slip up . So the AI handles the grunt work, and I fine-tune the details,” he added.

Gupta stressed that even with the benefits, there is a trade-off that he noticed—the time saved in writing the code often gets reallocated to reviewing. “AI isn’t great with edge cases, and in complex projects, it can introduce subtle, hard-to-spot errors. The risk jumps when I am working with an unfamiliar language or framework - AI-generated bugs can be sneaky and a nightmare to debug later.”

He further explained that AI coding tools are a significant productivity multiplier for experienced developers. Yet, beginners often find themselves in a frustrating loop of trial and error, lacking the foundational knowledge to leverage them effectively. He pointed out that while AI enhances existing programming skills, it doesn’t replace them.