How AI will change software engineering

Estimated reading time: 27 minutes

The consensus about the future of the software engineering industry is that there is no consensus.

That’s thanks to the rapid emergence of generative AI, a type of artificial intelligence that creates new content through large language models (LLMs), which seek to replicate, and scale, the learning processes of the human brain.

Generative AI, which had its breakout moment in late 2022 with the emergence of OpenAI’s ChatGPT, has made an immediate impact on the craft of software engineering, setting off a tidal wave of bold predictions about where the industry is headed.

Meta CEO Mark Zuckerberg has predicted that 2025 will be the year when it becomes possible to build an AI agent with coding and problem-solving abilities equivalent to a mid-level engineer, while Microsoft’s chief technology officer Kevin Scott believes that 95% of code will be AI-generated in the next five years.

Yet beyond the C-suite, technology leaders who are observing the rollout of AI within the industry are approaching the topic with less certainty.

“Anybody who claims to have any confidence about where AI is going to be more than a year or two out, the more confident they claim to be, the less you should believe them,” said Charity Majors, cofounder of the observability platform Honeycomb. “There’s so much hype,” she said. “Honestly, if you zoom out a little bit, so much of ‘AI’ you could just replace with ‘automation’.”

Joel Greensite, vice president of technology at digital consulting firm Publicis Sapient, has stopped making predictions altogether because developments, and existing predictions, are happening at a much faster rate than he anticipated.

“It’s just a highly uncertain time,” said Matt Beane, assistant professor at the University of California, Santa Barbara, and author of The Skill Code. ”Anybody who says that [they have] a very firm confident prediction about what skills are safe, what the software job of the future is going to look like, how you might get there, what junior people should do, all of these things – they’re selling something. I mean no one has the data to be that confident.”

Lines and lines of code… 

Generative AI’s biggest impact so far has been in the domains of code generation and content creation, forcing white collar workers and creatives to feel the most threatened.

Code generation has become so powerful that even those without technical skills are “vibe coding,” developing entire applications through AI prompts. Several startup founders have been telling Denny Gabriel, a venture investor at Runa Capital, that they are able to output code with up to 65% less headcount. This echoes recent sentiments shared by Y Combinator’s CEO Garry Tan, who said a quarter of its latest cohort of startups were launching with 95% AI-written code.

While code generation can be helpful to startups who want to move fast and break things, the applications often stumble when facing real-world demand and scale, said Alex Patow, an analytics engineer for venture firm Inflection. This is what Google engineer Addy Osmani calls the 70% problem. 

Holly Cummins, senior principal software engineer at open source software company Red Hat, recognized this when building a Quarkus demo application using AI prompts. “I looked at the code and I realized 80% of the code was not necessary,” she said. “We tend to spend a lot more time reading a codebase than writing a codebase. If I produce loads of flab in my AI-generated code, then I’ve reduced the cost of generating it, but I’ve increased the cost to the rest of the team of reading this code.”

Organizations are now “on the receiving end of unprecedented influx of software of unknown origin” as a result of engineers tinkering with code generation, Majors said.

Google’s 2024 DORA report demonstrated this displacement of productivity, noting that “AI does not appear to be a panacea.” While developers are being more productive with code generation, it’s starting to have a negative impact on software delivery throughput and stability with larger and less manageable changes.

In DORA’s statistical evaluation of responses from over 39,000 software engineering professionals, there’s an estimated 1.5% reduction in throughput and 7.2% reduction in delivery stability for every 25% increase in AI adoption – a marked charge from prior reports which showed improvements in software delivery.

GenAI, the new junior developer? 

As generative AI tools move up the stack, senior engineers will get more of the benefits than juniors, said Martin Reynolds, field chief technology officer at AI-powered software delivery platform Harness. By removing the mundane parts of the job, more experienced engineers can focus on solving complex problems.

As a result, many pathways for junior engineers to enter the workforce and learn on the job are being closed off. Some firms are looking to hire only senior developers, researcher Beane said.  

For example, Salesforce has said it may not hire any engineers this year due to AI productivity gains. Hiring activity for software developers fell at a slightly greater rate relative to other tech positions during the past two years from January 2023 through to December 2024 and has been slower to rebound at the start of this year, according to data analysis from trade association CompTIA. 

The data also points to a slight shift towards companies recruiting more experienced software engineers, the firm said, while also noting that it’s still early and these trends tend to play out over much longer time periods.

“It’s almost a false economy because you have to look beyond the one year into the three and the five years and say how does this become sustainable?” Reynolds said. “If everybody takes this route, how expensive are those top-line resources going to get? And will you be able to maintain your business and will it be successful?”

Senior developers are also becoming more reluctant to oversee juniors when working on complicated tasks because of the availability of AI tools, said Beane, based on conversations with engineers through his research.

A field study at three Fortune 100 companies found a 26% increase in completed tasks amongst developers using AI tools, noting that less experienced developers had higher adoption rates and greater productivity gains. 

However, Frank Licea, chief technology officer and cofounder of Howdy.com, highlights how this can create an imbalance where code is produced faster than it can be reviewed and integrated as senior developers aren’t experiencing the same level of productivity gains.

This is an apprenticeship industry, Majors said. The goal is for the team to be productive, not the individual. “You have to see your team as a system,” she said. “You have to pay attention to that system. I think that it’s very short-sighted to just be like, ‘we’re going to get the most senior people we can get for the least money.’”

Junior developers should take control of their learning and ask tools like Gemini and ChatGPT to explain the reasoning behind the prompt responses and not accept it at face value, Reynolds said.

Roles reimagined

Developers at all levels will need to adjust their skill sets to some extent to stay relevant, as  roles will be redefined rather than replaced. The ability to think critically and communicate effectively will become even more crucial when prompting a fleet of AI assistants, said Sabrina Farmer, chief technology officer at GitLab, in an email.

Peter Morales, CEO of edge computing platform Code Metal, expects there to be more of a shift toward systems engineering and high-level architecture design, while Farmer sees data engineering expertise as the next big boom.

“The ability to manage, process, protect, and deliver high-quality data is becoming ever more crucial because it’s the backbone of AI development,” she said. “Protection of data is so much more critical in an AI world.”

Developers will need to extend their logical mindsets toward sophisticated and targeted prompt engineering, said Reynolds, highlighting that there can be hundreds of lines of prompts to create a simple AI service such as a Siri request to play a song from a specific playlist. Plus AI doesn’t always give you the right answer, so persistently challenging and refining prompts often yields significantly better results, Farmer said. 

“Knowing [programming] languages and deep understanding of complex systems will never not be an asset,” Majors said. “But in order to make the most of that asset, you also have to skate where the puck is going.”

Move fast and break things

The challenge is that the puck could be moving too fast.

The industry is in the midst of real discontinuity, Beane said. The change from 2022 to late-2024 is going to seem like a warm up act compared to what will be experienced from now through to the next year and a half, he said.

An example of this is the SWE benchmark, Joe Kim, CEO of the analytics company Sumo Logic, said. The benchmark tests whether large language models can solve real engineering problems on GitHub. A year ago, the accuracy of these models were in the single digits. Now the accuracy comfortably sits in double digits for leading models, which Kim notes as demonstrating the rate of change.

Different LLMs can talk, argue and reason with one another, with increasing accuracy, Kim said. His company showed off this capability during the recent AWS Re:Invent conference, where an engineer could kick off a conversation with multiple site reliability engineering AI agents who work together. Despite the levels of accuracy, some human interaction will still be needed because their responses are non-deterministic, but it’s likely that jobs will have to adjust as this technology becomes more sophisticated, he said.

“Roles, even understanding what the work is [and] how it’s organized in firms, is up for far more disruption in the next 365 days than it was in the last 365,” Beane said. “It’s important for individuals and firms to have a plan that is useful if that turns out to be true.”

The problem is that most people don’t have a plan beyond platitudes, such as keep experimenting or playing with tools. Beane expects that CEOs will have to get to grips with the fact that humans won’t be able to “flex fast enough” and they will need to manage the fact that their core production function becomes a “live wire” in the company.

How to stand out in this landscape

Both Beane and Kim agree that creativity and initiative will help developers, especially juniors, stand out in this landscape, but they also acknowledge that it’s on companies to allow that creativity to thrive.

“I’d almost look at it as, especially as a software engineer, like it’s an amazing time to be alive right now,” Kim said. “There’s so many opportunities that this vehicle can bring.”

The rate of change is part of the reason Publicis Sapient’s Greensite is reluctant to make any predictions. In the short-term, he’s confident there will be better tooling for software development professionals, but in the medium-to-long term, he questions whether developers will even touch code. 

Natural language may become the new programming language, he said, where developers simply specify what they need and don’t need to touch programming languages like Python, not dissimilar to how the industry has very few people writing assembly or configuring cloud infrastructure these days. 

In this new world, developers will have more time to be creative and to experiment to solve client or business problems, he said. For example, products always have to make compromises and now they won’t because the technology unlocks the ability to cater applications to individuals through increased resources and capabilities.

“I don’t subscribe to the idea that we’re never going to write code again, but I do think the shape of what code is will change,” Greensite said. “I can’t tell you exactly what that’ll be, but only that the outcome will probably be that we’ll work on much more interesting things.”

Get articles like this in your inbox

Choose your LeadDev newsletters to subscribe to.