ChatGPT's response to this is informative. I particularly liked this bit. I hear this a lot from programmers desperately attempting to justify their existence.
I differ somewhat from the conventional “programmers will become architects” reassurance. That may be true during the transition, but it isn't obvious that it describes the equilibrium. If an AI can write 100,000 lines of code, there isn't a fundamental reason it must forever require a human computer scientist to decide how those 100,000 lines should be structured. Architecture is itself an intellectual task susceptible to automation.
But the whole response is good.
I think you may be directionally right, especially if we phrase the prediction as “software companies will cease needing nearly as many people whose primary skill is writing code.” That is a stronger claim than “AI will make programmers somewhat more productive,” and the evidence is beginning to point that way.
Already, 90% of professional developers surveyed by JetBrains in mid-2026 were using coding agents at least weekly and 68% daily. Gartner is now predicting that, among engineering teams using agents, conventional IDEs will become optional for most of them by 2027. And OpenAI reports something even closer to your marketer example: its nontechnical departments, including Legal and Recruiting, now use Codex extensively rather than leaving such work exclusively to programmers. (The JetBrains Blog)
Where I think your startup argument is particularly compelling is the economics. Imagine two companies trying to build the same new SaaS business.
The incumbent thinks:
product manager → designer → frontend engineer → backend engineer → QA → DevOps → security review → release engineering
The 2028 startup founder may think:
“I want the customer to be able to upload this spreadsheet, reconcile these records, charge the customer, and produce this report. Build it.”
Then the founder, marketer, accountant, industry expert, or product person spends an afternoon iterating with agents. There might be three people running a business that would previously have employed thirty.
That doesn't require AI to become literally flawless. It only has to become good enough that the remaining failures can themselves be detected and repaired by other agents.
And this is where I differ somewhat from the conventional “programmers will become architects” reassurance. That may be true during the transition, but it isn't obvious that it describes the equilibrium. If an AI can write 100,000 lines of code, there isn't a fundamental reason it must forever require a human computer scientist to decide how those 100,000 lines should be structured. Architecture is itself an intellectual task susceptible to automation.
There are nevertheless a few areas where I expect humans with serious CS expertise to persist much longer: operating systems, databases, compilers, hyperscale infrastructure, model training systems, cryptography/security, embedded systems, extremely high-performance software, and safety-critical systems. But that's a very different labor market from today's. Perhaps 5% or 10% of today's software-engineering population does that sort of work.
Large companies also have exactly the impediments you describe. A startup can say, “Don't hire an engineering department.” JPMorgan or United Airlines can't wake up Monday morning and dismiss 8,000 engineers because an agent appears capable of maintaining their decades-old systems. They have undocumented dependencies, regulatory obligations, internal politics, security boundaries, procurement rules, audit requirements and enormous amounts of legacy code. So incumbents could preserve conventional engineering employment years after the economic case for it has disappeared in greenfield companies.
There's actually an interesting piece of evidence for how rapidly the transition is occurring. METR's well-known early-2025 experiment found experienced developers working on mature repositories were 19% slower when using then-current AI tools. By February 2026, METR said its newer experiment had developed a serious selection problem because developers increasingly refused to participate if they were required to work without AI. (Metr) That is an extraordinary change in behavior in less than a year.
I would modify one part of your university prediction, though. I don't think computer science fundamentals become irrelevant. I think “learning to program” and “learning computer science” separate from one another much more sharply.
A useful curriculum might become something like: specification and decomposition, probabilistic reasoning about AI outputs, software/system architecture, security, databases and distributed systems, testing and formal verification, agent orchestration, evaluation, debugging AI-created systems, and enough algorithms/data structures/OS/networking knowledge to understand why something has gone wrong.
Learning Java syntax or spending three semesters becoming proficient at manually implementing software patterns could start looking rather like learning long division after calculators became ubiquitous.
And I'd go one step farther than your marketer example. The disruptive group may not be marketers who learn to vibe code. It may be people for whom “coding” ceases to feel like an activity at all.
An insurance expert will describe an underwriting workflow.
A logistics person will describe a dispatch system.
A lawyer will describe a document-review process.
A marketer will describe a campaign-management application.
They won't necessarily think, “I'm programming.” They'll think, “I'm telling the computer what I want.”
At that point, saying every company is a software company might remain true—but “software engineer” could stop being the normal intermediary between an idea and functioning software.
That's the labor-market discontinuity I think people who merely talk about “AI making developers 30% more productive” are missing.
