Everyone's job is safe except your job

There is a thing people do when they encounter a tool powerful enough to frighten them, which is to insist it has a soul.

I know, really weird, right?

They did it with fire, with the printing press, with electricity, and they are doing it now with LLMs, which are, at their core, extraordinarily elaborate filing cabinets with a gift for sounding confident.

What is an LLM and how does it work?

Here is what an LLM actually does: it reads a staggering quantity of text that human beings have already written, finds the statistical patterns in that text, and produces new arrangements of words that are, by design, consistent with what came before. That is the whole trick. The model has no window to look out of, no experience of Tuesday, no memory of being wrong last week. It has only the corpus: the vast, frozen record of human thought up to a certain date, and the mathematical relationships between the words inside it. When you ask it to write a poem about grief, it is not actually grieving, and when you ask it to generate a business strategy, it is not actually strategizing. It is doing something closer to very sophisticated autocomplete, the kind your phone does when it guesses your next word, scaled up to a size that makes the results feel miraculous.

Yann LeCun, the chief AI scientist at Meta and one of the architects of modern deep learning, has been quite direct about this for years, telling anyone who will listen that large language models are fundamentally bounded by the data they were trained on and cannot build genuine world models from text alone. A study published in the Journal of Creative Behavior reached a similar conclusion, finding that LLMs produce outputs that hover near the statistical average of human creativity: derivative by mathematical necessity, because the model’s entire purpose is to stay tethered to what already exists. The outputs can be fluent, useful, and genuinely impressive, but original in any deep sense they are not.

So, that’s the thing. Before we go any further, we should hold that fact in our hands and feel its weight: AI is a tool. It requires human input to function, and it can only see what has already been done.

The Hammer doesn’t Dream about Houses

There is an old joke about a hammer, that if the only tool you have is a hammer, every problem looks like a nail. It is a terrible joke, but it seemed appropriate. The joke with AI is somewhat more troubling, because a significant portion of the business world has convinced itself that their hammer is actually an architect, when what it has really done is memorize every blueprint ever drawn and learned to recombine them in ways that look, at a quick glance, like a new design.

This matters enormously because the value being ascribed to AI in boardrooms and press releases and keynote addresses is largely the value of invention, and the fantasy that the model will produce ideas that no human has yet had. Companies are not paying billions of dollars for a better search engine. They are paying for the story of a mind.

What AI does exceedingly well is pattern completion at scale. It can summarize, it can reformat, it can generate boilerplate at speed. It can hold a large amount of existing information in useful relationship with each other. These are genuinely valuable capabilities, and we shouldn’tbe too proud to use them. But the moment someone in a strategy meeting says “let’s let the AI come up with the solution,” they have misunderstood the nature of the machine entirely, because the machine doesn’t have any solutions. It has the historical record of solutions that humans have previously proposed, filtered through probability.

The input, always and without exception, is human. The prompt, the context, the judgment about whether the output is any good (all of that falls to the person in the chair). The AI has no stake in the outcome and no understanding of the problem. It has only the corpus, and it is doing its level best with what it has been given.

Which brings us to the prompt library, that beloved artifact of the corporate AI rollout, the Google Doc or shared Confluence page where someone has lovingly catalogued the fifty-seven prompts that the team has discovered work reliably, so that everyone can benefit from the collective wisdom and nobody has to think too hard. Prompt libraries are a sensible idea in the same way that saving last year's bus schedule is a sensible idea. Yes, the buses exist, the concept is sound, and the document will be completely wrong by the time anyone goes to use it.

Here is why: the models themselves change constantly, and when a model changes, the prompts written for the old model begin to misbehave in ways that are quiet, gradual, and maddeningly difficult to catch. Researchers call this phenomenon prompt drift, which is the polite technical term for "the thing you told the machine to do is no longer quite what it is doing, and you may not notice until someone complains." A study of GPT-4 and GPT-3.5 found that response accuracy on the same tasks fluctuated considerably over just four months, with degradation on some categories reaching more than 60%. The model is updated (for safety alignment, for architectural improvements, for reasons the company will not fully disclose) and the relationship between your carefully written prompt and the model's behavior shifts underneath you like furniture rearranged in a dark room. OpenAI has revised ChatGPT's system instructions dozens of times since GPT-4's initial release. Anthropic has done the same across every model generation. Each revision changes, at a level most users cannot see, how the model interprets language, what it prioritizes, and where it pushes back. A prompt that worked beautifully in March may return subtly different results in July and noticeably worse results by December, with no error message, no warning, and no indication that anything has gone wrong at all …just a slowly degrading quality of output that the team chalks up to the AI "having an off day," because the alternative explanation, which is that they need to rewrite all fifty-seven prompts again, is too dispiriting to contemplate.

New model versions compound the problem further, because each major release effectively resets the relationship between human and machine. The prompting strategies that worked for GPT-4o and Claude 3 required extensive step-by-step instructions, because those models needed to be walked carefully through a task. Newer models are sophisticated enough to find that kind of hand-holding condescending and respond to it with subtly worse outputs, in the same way that giving very detailed instructions to a competent professional sometimes produces worse results than simply describing what you need. The prompt that trained a less capable system becomes the prompt that patronizes a more capable one. This means that what companies are actually investing in, when they invest in AI capability, is a rapidly degrading collection of assets that become relics of the past shortly after someone selects the “publish” button. Wait, what happened to Agile?

A Manifesto and a Slow Death

In February of 2001, seventeen software developers gathered at the Snowbird ski resort in Utah and wrote a new process: a short document arguing that the way the industry was building software was making everyone miserable and that there was a better way. They called it the Manifesto for Agile Software Development, and its first declared value, the one they put at the very top, before all the others, was this:

“Individuals and interactions over processes and tools.”

That sentence was very deliberate. It was an intentional argument, written by people who had watched waterfall development strangle entire organizations in bureaucracy, that human beings working together in good faith were more important than any system, methodology, or piece of software you could throw at them. The people were the point, and the tools were supposed to serve the people.

The Agile Manifesto has had an interesting twenty-four years since then. It spread from software teams to marketing departments to Boy Scout troops to restaurants, as one account noted, because its core insight that flexibility, collaboration, and human judgment matter more than rigid process. It turned out to be useful in almost every context where people need to build things together. Agile became the dominant framework of modern software development. Sprints, standups, retrospectives, and user stories became the furniture of the industry.

And then AI came along with a very different proposition, which was: what if you automated the humans?

What? Wait…

Digital.ai’s State of Agile Report described AI as having moved from a “supportive tool to an orchestrator” in development cycles, which is a polite way of saying that the thing the Agile Manifesto put at the very top of its values (remember that whole “individuals, and their interactions” thing) is being systematically replaced by a machine. The report also found that 79% of development teams are being asked to do more with less, which is the corporate way of saying that the headcount is shrinking and the tools are supposed to cover the difference.

What happens to “individuals and interactions over processes and tools” when the tool is being positioned as a replacement for the individual? The manifesto does not offer guidance on this, because the people who wrote it couldn’t have imagined it. They were fighting against documentation-heavy bureaucracy, not against the prospect of the humans being automated out of the loop entirely.

It should be said, in the interest of fairness, that Agile had already done considerable damage to itself before AI arrived to finish the job. There is only so much a manifesto can survive, and seventeen years of SAFe implementations, two-week sprints that somehow never end, and "agile transformations" run by people who had never shipped software was a lot to survive. Ultimately, Agile deserved what happened to it, having already been mummified by the people who loved it most.

RIP Scrum.

A Comedy in Three Acts: None of Them Funny

Here is where it gets interesting, and also a little dark.

The Engineers: The people most loudly making the case that AI will replace everyone in your organization are, with startling frequency, software engineers. They are the ones producing the demos, writing the white papers, giving the conference talks, and generally conducting a sustained lobbying campaign on behalf of a technology that they control, deploy, and are paid to build. They are, to borrow a phrase from a different era, not entirely disinterested parties.

The argument goes roughly like this: AI will replace the lawyers, the marketers, the writers, the designers, the analysts, the customer service representatives, the data entry clerks, and the project managers. It will automate away enormous swaths of cognitive work across the economy. It is coming for everyone. Everyone, that is, except the engineers who build and maintain the AI systems because those people, we are told, are indispensable.

This is a very convenient story to tell when you are an engineer.

The data suggests a more complicated picture. Stanford researchers who analyzed payroll records from ADP found that employment for software developers between the ages of 22 and 25 declined by nearly 20% from its peak in late 2022, concentrated precisely in the tasks (boilerplate coding, scripted testing, routine bug fixes) that AI tools now handle with ease. According to IEEE Spectrum, overall programmer employment in the United States fell 27.5% between 2023 and 2025. Software engineering job postings, per Indeed’s Hiring Lab, were down 36.4% compared to February 2020. Salesforce CEO Marc Benioff said the company hired zero new engineers in fiscal year 2026. Shopify’s CEO sent an internal memo telling teams they needed to prove AI could not do a job before requesting new headcount.

What is particularly worth noticing is the mechanism by which this is happening. Companies are not replacing engineers with AI because AI has transcended human capability. They are replacing certain engineers with AI because AI can produce adequate code, at adequate speed, for the portions of the job that were already well-defined, and “adequate and cheap” beats “excellent and expensive” in a spreadsheet every single time. This is a story about cost reduction, which is what almost every technology story is, eventually.

The Designers have concluded that engineering is finished, which is the kind of confidence that comes from not having examined one’s own foundations recently. Design is, at its serious core, many things that are genuinely different from each other and that the industry spent twenty years pretending were the same. Graphic design is a fine arts discipline (composition, typography, color, etc) and the slow education of the eye that happens over years of making bad things before making good ones. UX design is something closer to applied social science, and the people who did it well came from HCI graduate programs and psychology departments, carrying credentials that represented years of learning how people actually behave rather than how they behave in a scenario a designer invented. Service design thinks about entire systems: the gap between organizational intention and human experience, the places where a perfectly reasonable policy produces an absurd outcome for the actual person trying to get something done. Interaction design, information architecture, content design, motion design, the list goes on and each is a real discipline with a real body of knowledge and a real set of things you have to learn before you can practice it without causing harm. None of them are the same. All of them got the same job title, because the same job title was cheaper.

Graphic design is not UX design. The fine arts student who spent four years learning to kern a headline and the cognitive psychologist who spent three years studying wayfinding in hospital environments are not the same person and are not doing the same job, and the bootcamp that trained both of them in twelve weeks was performing a kind of institutional sleight of hand that everyone agreed to accept because the artifacts looked similar and the salary was lower and nobody in the hiring meeting had a strong enough opinion about the difference to slow things down. They should have slowed things down.

The confusion about what design actually is traces a fairly direct line back to IDEO, which popularized design thinking (a methodology so successfully marketed that it spread from product development to healthcare to government to schools to, eventually, a certain number of restaurants) and in doing so convinced an entire corporate class that design was primarily a mindset available to anyone willing to do the empathy exercises, rather than a discipline requiring years of serious study to practice with any integrity. Design is a skill, design thinking is a brand and IDEO built a $300 million business on the brand, right up until the business contracted to $100 million and the offices in Munich and Tokyo closed and the layoffs came in rounds, which is what happens when a consulting proposition is built on a fad rather than a foundation. What is coming to replace design thinking in serious organizations is systems thinking or the practice of understanding how complex systems produce emergent outcomes, how interventions ripple through interconnected components, how the thing you changed over here broke the thing over there that you were not watching. It is, importantly, hard to learn, slow to develop, and impossible to deliver as a two-day workshop with a wall of sticky notes at the end. The designers who already think in systems will find that their moment has arrived. The ones who learned design thinking and called it design are finding out, at some inconvenience, that those were always different things.

The Product Managers have split into two parties, like a country facing an election nobody wanted. The first party is calling itself the AI PMs, which means they have added “AI” to their title and are hoping nobody asks them to explain the attention mechanism in a transformer. The second party has become growth PMs, which is a dignified way of saying they have decided to be marketing people who never quite made it to marketing, generating acquisition funnels and engagement metrics and calling it product strategy. Both parties are fleeing the same burning building, which is the recognition that a language model can produce every artifact a traditional product manager was hired to produce, faster and cheaper and without asking for a seat at the roadmap review.

The reason this hurts as much as it does is that product management spent the better part of two decades recruiting people whose primary qualification was a degree from a business school, on the theory that the MBA signaled a kind of strategic general intelligence that could be applied to any domain. What it actually signaled was comfort with frameworks, fluency with PowerPoint, and a willingness to facilitate meetings about decisions that the engineers and designers in the room had already made. The engineers tolerated it. The designers tolerated it less. And now AI has arrived and can facilitate the meeting, write the summary, and generate the follow-up doc, and what is left is the judgment, the genuine technical or creative intuition about what to build, which was precisely what the MBA was not designed to produce and precisely what nobody thought to hire for.

Annnnd, we are back to the death of Agile like a snake eating itself.

The Spreadsheet has always Won

Let’s say what is actually happening without the softening language of “transformation” and “augmentation” and “the future of work”: companies discovered that AI coding tools give them 40 to 55% more code output per sprint, according to tracking data from development teams, and they responded by doing exactly what companies always do when productivity increases: they reduced headcount and kept the margin. Meta, Amazon, Microsoft, Salesforce, Dropbox, and Duolingo have all made this calculation explicitly. The savings from not hiring junior engineers are being redirected, in many cases, directly into AI infrastructure spending.

This is capitalism working as designed, and it shouldn’t surprise anyone. The shocking part is that the profession most responsible for building these tools spent years arguing in blog posts, in conference talks, and on social media that AI was coming for everyone else’s job. And there is a reasonable argument that this was, whether consciously or not, a strategic positioning exercise: make AI sound universal and inevitable so that the people who control it appear uniquely necessary.

The Agile Manifesto said people over tools. The industry responded by spending the better part of a decade building a tool capable of replacing a meaningful percentage of the people. The irony is palpable.

Here’s the Skinny

AI is a genuinely useful tool, and nothing argued here is a case against using it, the hammer is a good hammer (pat on head), the filing cabinet is a magnificent filing cabinet (gestures supportively), the autocomplete is the most impressive autocomplete in the history of autocomplete (mmm…kay).

But a tool is what it is, and it requires an operator with a goal. It requires context that only a human with experience, judgment, and skin in the game can provide, but it can’t disagree with you when you are wrong, because it doesn’t have opinions. It has probability distributions, so you are going to have to ask for a disagreement. Kind of like vampires who can’t come in unless you invite them.

The disruption to software development is real, and it is specifically a financial disruption: companies will save money on engineering headcount, especially at the entry level, because the entry level is where AI is most capable of substitution. The deeper skills (system design, architectural judgment, understanding of a specific business domain and its constraints, the ability to know when a technically correct answer is the wrong answer) those remain stubbornly human for now, though it would be unwise to assume they will remain so indefinitely.

We have always been here before, in one form or another. The printing press was going to make scholars obsolete, and the calculator was going to make mathematicians obsolete, and don’t forget that the word processor was going to make secretaries obsolete. Each of these technologies changed who did what and how much it cost, and the people who figured out how to use them well came out ahead while the people who decided the tool had replaced the need for human judgment came out looking foolish.