AI is probably not going to steal your job – or at least, not yet.

As good news goes, that’s about as hopeful as you can get in this brave new world of agentic algorithms. Humans still have the edge (though perhaps only just) in those essential skills needed for work – communication, critical thinking, creativity, and adaptability, to name a few.

But even if we are keeping ahead of the bots in soft skills for now, our opposable thumbs are no match for the sheer technical wizardry of today’s LLMs. AI can code in minutes or seconds projects that would take a human days, maybe weeks, to complete, and evidence shows it routinely outperforms in data-heavy, repetitive tasks.

The technology is only going to get faster, with everything pointed to AI swiftly replacing humans. The World Economic Forum has predicted that AI will transform or displace 92 million jobs by 2030, with research suggesting 3 million UK jobs are on the line.   

Being on the front lines of this shift, IT workers are rightly worried about their future, but rather than see AI as an existential threat, the technology should be embraced as an opportunity to elevate their roles.

That was the core message from Jonny Brooks-Bartlett, senior machine learning engineer at Spotify, who addressed a standing-room-only crowd at this year’s DIGIT Expo with a session that challenged the audience to look beyond the hype and fearmongering to find real opportunity in AI productivity.

The process starts by taking a long, hard look in the mirror.

Find the Workflows AI Can’t Crack 

“If you’re worried that AI is going to take your job right now, it’s probably because you’re not doing enough,” said Brooks-Bartlett.

Harsh, but fair. So far, AI has only fully displaced the simplest tasks, like scheduling, customer service, and back‑office admin, with recent figures suggesting that traditional data entry jobs will fall by 25% in less than ten years, while up to 80% of routine customer service issues could be fully automated by 2030.

However, while AI might excel at repetitive, rules‑based work, modern jobs are a complex mosaic of workflows, fragments that technology cannot overtake wholesale.

Take, for example, the position of a machine learning engineer at Spotify.

“What do I do? Well, most people say, ‘oh, you code’, but actually, coding is a relatively small amount of my time,” explained Brooks-Bartlett.

“I also propose experiment designs, I’ll do planning, I write Request for Comment documents, I implement new features, I communicate results, I present, and I’ll write emails and blog posts. There’s a lot of stuff I do in my job that an AI coding agent does not do.”

Those workflows break down into countless tasks, each calling for some element of human skill, with Brooks-Bartlett arguing it’s crucial firms are clear‑eyed about where AI adds value, and where it can’t replace human insight.

For instance, an AI coding agent might have the same technical skills as a room full of developers, and churn out reams of fairly clean code in double time, but it’s important to take a holistic view of a human’s role.

An AI can code, but it can’t work on a problem as part of a team, it can’t see past the data to find creative solutions, and it can’t yet provide trusted, unbiased feedback.

“If I, as a company, say I’m going to bring an AI agent to replace all of my engineers, I’m going to be in a world of pain, because it’s only going to do part of the job,” he said.

“You need to think about the different workflows that each role does. Split those up into the tasks that make up those workflows, and then ask what tasks does this tool actually improve? 

“The better question to ask is not will AI take my job, it’s can one or more of these tools improve how I tackle different tasks in my workplace?”

Make AI Work Smarter – And Harder

At Spotify, one area AI is increasingly shaping is product development. According to Brooks‑Bartlett, incoming co‑CEO Gustav Söderström has pressed product managers to rely more on AI evaluation when building and refining new features, a step-change from the product requirements documents of the past.

The shift has forced a fundamental re‑examination of process, with the company attempting to define what, exactly, a good idea looks like, then encoding that in a series of evaluations and tests to score AI output.

Training an LLM to do this is a meticulous task. Engineers must first sift through examples of input and output, gathering all the data needed to design an initial prompt, then devise ways for the AI to measure its own progress. 

Once the eval loop begins, the AI and engineer together grind through round after round of fine‑tuning, honing the prompt until the output looks like a product that could actually ship, a process that ends with asking the AI what original input prompt would have led to this answer quicker. 

It takes time, and it can be frustrating to wrangle AI into alignment, but according to Brooks-Bartlett, these cycles of trial and error are unlocking real gains for the company. 

“As a product manager, I can actually prototype the feature that I need before I have to give it to an engineer,” he explained.

“I’m no longer handing them a document that tries to define, in some natural language, wishy-washy, hand-wavy way, this is a target audience, this is what I want the feature to do. 

“I have a prompt that shows the input that I want and the quality of the output that I want, and I can pass on something clear to engineers when creating something new.”

While working so closely and collaboratively with AI would have been seen as fairly radical even a year ago, Brooks-Bartlett isn’t too far ahead of the curve in so tightly weaving the tech into his daily practice.

According to Google’s latest DORA Report, AI adoption among software development professionals surged to 90% this year, a 14% jump from 2024, with the majority (65%) heavily relying on AI for software development and typically dedicating around two hours a day to working side by side with LLMs.

Recommended reading

• 96% of Leaders Say AI Fails to Deliver ROI, Atlassian Report Claims
• Gartner: AI Will Touch All IT Work by 2030
• Developers Expect AI to Redefine Their Roles in 2026

That shouldn’t be surprising. As Brooks‑Bartlett points out, every line of human‑written code is a liability to be maintained, and most software pros already see that AI is outpacing them in this and countless other core tasks.

Instead of plugging away at a terminal for hours to write a thousand lines of code, the job of a software developer or engineer has been reimagined as that of a creative, where value lies in the feature or the product built, not the code it takes to get there.  

“I try not to code now,” said Brooks-Bartlett. “I’m an engineer, but I am in this position where I do not want to code anymore. I don’t have to. What I want to do is go to an LLM and say I need to do this thing. Can you do it for me?

“That is what I want my job to be.”   

Verifiability is the Limit

AI may be central to his workflow, but Brooks‑Bartlett is clear that he uses it under tight conditions.

“I don’t trust any of these tools enough to just let them run wild,” he said.

“There are parts of my workflow that this tool can automate and improve, but there are parts that it still can’t. And so there are parts of my job that still need me to do, as a human.”

One of the key responsibilities for humans that he sees continuing well into the future is double‑checking that an AI is really delivering what it’s asked to do.

“In the past, when we wrote the code ourselves, and we wrote the tests, it was our own suite that actually tested whether the feature worked,” said Brooks-Bartlett. “That is not going to be the case anymore.”

Now that AI handles most of the coding and much of the drudgery, the Spotify engineer sees his role as verifying outputs, ensuring the system doesn’t hallucinate, lose the context of the prompt, or introduce security flaws.

That oversight is critical. Recent studies show that 25% of developers estimate that 1 in 5 AI-generated suggestions contain factual errors or misleading code, and 60% who use it for testing, writing, or reviewing, complain that LLMs miss relevant context.

“I would be very, very wary of making AI automate something if I cannot independently verify the outputs,” said Brooks-Bartlett. “If you are not in a position to be able to verify the output, proceed with caution.”