There are two camps arguing past each other right now, and both views are too extreme.
The first says AI will never really replace software engineers; it’s too complex, too human, too contextual. AI is just a better tool, like a smarter Integrated Development Environment (IDE). The second says the developer’s role disappears within five years, replaced entirely by autonomous agents. So, the real question is: where do you stand in this shift?
I’ve spent 18 years delivering software, for startups that fit in a single room and for corporations so large they’ve lost track of their own size. What I see is something more interesting than either tribe is willing to admit. The same shift is happening across the entire economy. Most people just haven’t noticed yet, because they’re looking in the wrong place.
The part everyone’s been missing
If you’ve read a single headline about AI in the last two years, you already know which jobs are supposedly under threat. Programmers being laid off. Junior developers are struggling to find work. Companies are slowing or freezing hiring, because one AI tool can now do the work of multiple entry-level engineers.
The discussion is loud, anxious, and some of that concern is justified. But a new MIT study suggests we’re focusing on the wrong problem.
Most headlines treat AI as something that replaces jobs. That’s not really how it works. More often, AI replaces the tasks inside them. A lawyer doesn’t disappear overnight, but the hours they spend reviewing routine contracts quietly reduce. A journalist keeps writing, but the time spent on background research compresses. And while that might sound like a softer version of the same story, it isn’t, because our entire economic system was built around jobs, not tasks. GDP, unemployment figures, wage data, they count jobs and people. They were never designed to look inside a job and ask which parts AI can already technically do.
So, the change isn’t happening where most people are looking for it.
By the time disruption shows up in official numbers, it’s already well underway.
Massachusetts Institute of Technology (MIT) decided to build something new to fix this. Not another prediction of which jobs disappear, but an actual map of where AI capabilities and human skills currently overlap, weighted by the economic value of that work. They called it the iceberg index.
When you measure the work AI can technically perform across the tech sector, it accounts for about 2.2% of total US labor market wage value, roughly $211 billion. That’s just the visible tip. But when you apply the same methodology to the whole economy, the number jumps to 11.7%, roughly $1.2 trillion. Five times larger. Most of this impact is hidden, and it’s happening in high-skilled, well-paid jobs that no one is really talking about.
The anxiety that’s been dominating headlines for two years has been aimed at roughly one-fifth of the actual problem. The other four-fifths have been sitting on a real risk that no government, no company, and no individual worker has been preparing for.
And the people sitting on it aren’t who you’d expect. According to a separate Anthropic study tracking actual AI usage in professional settings, the most exposed group earns 47% more on average than the least exposed, is nearly four times as likely to hold a graduate degree, and is 16 percentage points more likely to be female.
In plain terms: people whose working day is built around reading, writing, analyzing, and summarizing information. People who, by any reasonable measure, did everything society told them to do, and did it well.
Two earthquakes, not one
To understand where software specifically is going, you need to understand the two big changes that have already happened. Most people only remember the second one.
The first was 2020–2021. Covid forced every business with a physical presence to become a digital business overnight. Software stopped being “that IT project” and became the main sales channel, the main everything. Demand didn’t grow; it exploded.
The second was 2022–2023. OpenAI released GPT, and the world saw for the first time that language and code are the same problem to a machine. Most people made poems and silly images. A small group of engineers started paying very close attention.
By 2025–2026, the latest models crossed a quality threshold that actually matters, not impressive demos, but real commercial software delivered reliably. A generation of builders created agent layers that don’t just generate code, but manage workflows, make decisions, and loop back on failures. A job, not just a prompt.
The first earthquake changed who needed software. The second is changing who, or what, builds it. These are not the same disruptions.
The tribe that’s right for the wrong reasons
The “AI is just a tool” camp is correct that human judgment isn’t going anywhere. But they’re clinging to that truth to avoid a harder one: the nature of that judgment is changing completely.
For years, the best person in software was the one who could keep the most code in their head. That’s not really true now. What matters today is being able to say, clearly, what the software should do and why it matters.
A machine can write code. It can’t tell you what to build. The people who use software are human, and they know what they need better than any model.
So the skills worth learning now come before the code. The first is framing the problem, which means looking past what people ask for and finding what they really need. The second is thinking in systems, so you can see how one choice changes things later. The third is product sense, which you only get from building things and being wrong many times. And the last one, often overlooked, is taking a human need and writing it down clearly enough that someone can build it without asking you ten questions.
The “developers disappear” camp is equally wrong, but in the opposite direction. They see the automation of code generation and assume the whole profession collapses. What they’re missing is that the demand for software is about to grow faster than agents can build it. The appetite for new digital experiences doesn’t go away when the cost of building drops. It accelerates.
The price of software is about to fall. The demand for it is about to rise. These two things happening together is not a crisis. It’s a market.
Nobody notices the blue
There’s a criticism of AI-generated software that comes up constantly in professional circles. I want to address it head-on, because I think it reveals more about the people making it than about technology.
The argument goes like this: AI still generates inconsistent code. The design doesn’t match the mockup precisely. The blue in the generated UI isn’t quite the same blue as in the brand guidelines. The spacing is slightly off. A senior developer would never let that through. Therefore, AI-generated software isn’t production-ready.
Let me be direct: 99% of end-users of the software will never notice if the blue on the web page is a shade lighter than the blue in the logo. They will not measure the padding. What they will notice is whether the product loads fast, whether the checkout works, or whether the interface makes sense on their phone at 11 p.m. when they’re trying to get something done.
The pixel-perfection argument is, in most cases, a professional defense mechanism. It’s how craftspeople justify their rates when the machine starts doing the job at a fraction of the cost. The typesetter worried about kerning. The darkroom technician worried about color grading. They weren’t wrong that those things mattered; they were wrong about how much they mattered to the end customer.
None of this means quality stops mattering. There’s a segment of the market where the interface is the product, where customers pay for how the software looks and feels, where polish is what they’re buying. Human craft will keep earning a premium there, and that segment is not going away.
But it’s not where most software is built. Most software exists to solve problems. The customer has a job to do, the software helps them do it, and if it works, they’re satisfied. The bar for “good enough” in that world is lower than people inside the industry usually assume, and it happens to be exactly where agents are working today. Every quarter, they get better at clearing it.
The gap between capability and reality
Here’s something worth holding onto before the panic sets in the gap between what AI can technically do and what it’s doing in practice is still enormous.
According to Anthropic’s 2026 research paper “Labor market impacts of AI,” large language models are theoretically capable of handling 94% of tasks in computer and math occupations. But in observed professional use, they currently cover only 33%. A similar pattern shows legal work, architecture, and engineering. The technical capability is already there, but it’s being held back by regulation, integration challenges, and the simple fact that most organizations still require a human to check AI’s work.
These are all things that slow it down today but probably won’t forever. And the signs are already there. Entry-level employment in AI-exposed occupations has already dropped 14% compared to the pre-ChatGPT era. Entry-level job postings across the US have dropped 35% since January 2023. Job postings fall before employment does. This is the early signal, not the final one.
Software is entering an industrial phase, not through disruption overnight, but through rapid, continuous acceleration. Faster than most are prepared for.
The new human role
If that future is coming, and I believe it is, then the human role in software delivery reorganizes around three things.
Specification: Someone has to define what gets built, and to do that well, you have to understand the business problem first. That starts with research, talking to the people who will use the software, watching them work, reading the support tickets and the workarounds they built because the current tool didn’t fit. Then analysis: looking at the process end to end, spotting the steps that exist only because no one questioned them, separating what people say they want from what would actually solve the problem. Only then can you write a specification that is both complete and coherent. This skill is undervalued in engineering culture today, which prizes people who build over people who define. That is about to invert.
Supervision: Agents make mistakes in unpredictable ways, and someone needs to catch them. The human job is to understand where an agent drifts, to correct it, to train it on the specific context of a company’s codebase, its standards, its preferences. The agent becomes a digital employee. The engineer becomes the manager who makes that employee more capable over time.
Architecture: High-stakes decisions about system design, how data flows, where the boundaries are, what fails gracefully, and what can’t fail at all, require someone who carries accountability. That remains a human function, because accountability is a human concept. A model cannot be held responsible. A person can. Good architecture isn’t just about how a system works today, but how it evolves under pressure, at scale, over time, and across changing requirements.
The private banking model of software
There’s an analogy I keep returning to when I try to explain what human involvement in software delivery will look like in ten years. It comes from banking.
Think about what a digital bank is. The entire product is software. You open an account in four minutes on your phone. You move money, apply for credit, manage investments, all without speaking to a single person. And for the vast majority of customers, that’s not a compromise. It’s the preference.
But walk into a private bank. The moment you cross the threshold, a person knows your name. The relationship manager has read your file and has a view on your portfolio that accounts for your daughter starting university next year. That human presence isn’t an operational necessity; the software could handle the transaction. It’s a signal. It says: you are important enough that we assigned a person to you.
Software delivery is heading to exactly the same place. In five to ten years, the default mode of building software will be agent-led. Specify, the system builds, feedback loops are automated. That will be the standard tier, accessible, fast, and good enough for the overwhelming majority of products.
But some clients will want a person in the room. Not because the agents can’t do the job, they can. But because having a human architect who has seen a hundred similar systems, who can push back on your assumptions, who will take the call on a Friday evening when something breaks into production, that is worth something that can’t be automated. It becomes a premium service. A deliberate choice. A luxury: something you pay more for not because it’s necessary, but because it signals the level of care you expect.
This two-tier split will feel uncomfortable to many people currently working in software. But it’s already happening in every industry automation has touched. The physical bank branch is a luxury. A tailor-made suit is a luxury. A doctor who makes house calls is a luxury. Human time and judgment become more valuable, not less, as automation absorbs the routine. The market just prices it differently.
The workers AI can’t touch, and why that’s not great news either
There’s another group worth thinking about. About 30% of the workforce has essentially zero AI exposure: cooks, mechanics, nurses, plumbers, bartenders, childcare workers. People doing physical, relational, hands-on work that no language model can replicate.
Surely those workers are fine?
Not exactly.
There’s an economic pattern that’s been quietly working against these workers for decades, long before anyone had heard of a large language model. In 1965, a Princeton economist named William Baumol noticed something odd about the performing arts. A string quartet performing Beethoven in the 19th century required four musicians and about 25 minutes. A string quartet performing the same piece a century later required the exact same number of musicians and lasted about the same time. Nothing had gotten more efficient. And yet, the cost of putting on that concert had risen dramatically, dragged upward by wages rising everywhere else.
As manufacturing and industry got more productive, they could afford to pay their workers more, which pushed wages up across the whole economy. Musicians had to be paid competitively to attract talented people, even though the output, four people playing for 25 minutes, never changed. So, the cost per performance kept climbing.
Economists call this Baumol’s cost disease. You can see it clearly in prices over the last 50 years. The things that got more productive, electronics, computers, got cheaper. The things that couldn’t, childcare, education, haircuts, healthcare, kept getting more expensive.
AI is about to make this much worse. If cognitive and administrative work is about to get dramatically more productive, then the Baumol effect will accelerate. A financial analyst processing documents with AI assistance might shrink a day’s work into an hour. A software engineer with the right tools might do the work of three people. But the nurse will still need the same amount of time per patient. The plumber will still need to physically be there. Their output doesn’t scale with AI, so the relative cost of their work keeps rising, pulled upward by a productivity surge happening all around them.
And most of the work AI can’t touch is essential. People don’t pull their kids out of school because costs went up. They don’t skip the plumber when the pipes burst. Workers who are safe from AI disruption may find themselves in industries that governments will increasingly struggle to fund.
What I’d tell a software company right now
The businesses that will thrive in this transition are not the ones that hold the line on human-only development as a point of principle. Nor are they the ones that hand everything to agents and hope the output is coherent. They’re the ones that redesign their delivery process around the new division of labor, humans setting direction and maintaining accountability, agents executing at speed and scale.
The price of building software is dropping. If you’re in the business of selling software services, your margin is migrating upward, toward the decisions that require judgment, context, and trust. The companies that figure out how to sell those things, rather than the hours of implementation, will own the next decade.
The engineers who understand this, who start positioning themselves as the private banking tier of development, not as people competing with agents on implementation speed, are the ones who will still be doing very well in a decade.
Human presence in software development isn’t disappearing. It’s being repriced.
The factory isn’t finished. But the blueprint is on the table, and anyone paying attention can read it.
The author has 18 years of experience in software engineering and digital delivery, working with companies from early-stage startups to global corporations.
