the new builders

On June 18, 2026, Accenture lost roughly a fifth of its market value, the steepest one-day fall on record for the stock. Revenue had actually grown, six percent in dollars. But new bookings shrank by two percent, guidance came down, and investors, as Bloomberg reported, read a bigger story into it: the fear that AI is repricing the consulting market. Capgemini was dragged down the same day, more than eight percent to a 52-week low. A year earlier, Accenture had cut over eleven thousand jobs in a restructuring of some 865 million dollars, with CEO Julie Sweet explaining on the earnings call that for some of those leaving, reskilling toward the AI skills the firm now needs was no longer a viable path.

In that same market, the opposite story was unfolding. Fortune described how Maor Shlomo built the Base44 platform in four months, working essentially alone: a tool that lets non-technical users create software by describing what they want. Its first month after launch brought in nearly one and a half million dollars in subscriptions. Six months after founding, Wix bought it for eighty million dollars, as TechCrunch reported. The same magazine profiled Dana Snyder around the same time, a nonprofit consultant with no technical background, who built her own software platform in half a year and runs it as its only full-time employee.

Two extremes of one movement. What the giants are losing and what the solo builders are gaining flow from the same source: building software has changed in composition.

the craft did not get faster, it got different

The hardest evidence available right now comes from the Anthropic Economic Index, an analysis of some four hundred thousand working sessions from roughly 235,000 people building with agents, measured from October 2025 to April 2026. Two caveats up front. Anthropic measures its own users, so this is telemetry from the front of the pack, not a snapshot of the whole market. And success in this research is a classification of the conversation, not proof that the result ever reached production.

What the telemetry shows is not acceleration but a shift. The share of sessions spent repairing broken code dropped from 33 to 19 percent in half a year. The share spent operating and running software rose from 14 to 21 percent, and the share spent on writing and data analysis roughly doubled, from about ten to twenty percent. Less time on what is broken, more time on what needs to exist. And there is no straight line to full automation in these numbers: between two consecutive samples, the share of automated interactions in API traffic actually fell sharply.

Anyone who has worked in a traditional team knows what that shift means. The classic software project spends most of its calendar not on thinking but on friction: handovers, alignment, waiting, repair. When an agent takes over the typing and the repair work, the friction does not disappear, it relocates. The scarce skill is no longer being able to write code. It is knowing precisely what should be built, and being able to judge what was built.

domain knowledge beats the diploma

The most striking number in that research is not about speed. Anthropic compared how often people from different occupations complete their building task successfully, and found that every major occupational group, from managers to lawyers to social scientists, succeeds at coding tasks at nearly the same rate as professional software engineers. The entire top ten sits within about seven percentage points. And the more domain knowledge someone brings, the more work the agent performs per instruction.

That is the real disruption, and it is quieter than the demos on LinkedIn. Not that everyone can program now, but that the predictor of success has moved. For decades, the first question was: who can build this? The question has become: who understands this process well enough to direct a system that builds it?

I see it daily. My value in a project sits less and less in the lines of code I type myself, and more and more in the fourteen CRM projects I have carried from start to finish. They taught me how a quote actually moves through a company, where a service process jams, and which edge case will hurt on day ninety. The customer portal I built for Elite Klimaat went live after a month of building, for under five thousand euros. Five years ago that would have taken a team and a multiple of the budget. The difference is not working harder. It is a different way of building, steered by process knowledge.

the old arithmetic breaks

The traditional delivery model is arithmetic: hourly rate times team size times months. Every factor in that equation is under pressure. When one builder with agents delivers what used to take a team, and lead times drop from months to weeks, little of the equation survives. No hard, verified number for that gap exists yet; anyone quoting one is making it up. But the direction is in the books: new bookings at Accenture shrank last quarter, and the stock market is already pricing in the doubt.

The counterargument deserves its stage: Accenture’s CEO disputes the AI explanation. She points to weak demand, US government spending cuts and geopolitical unrest, and puts a number of her own against it: bookings for deals above a hundred million dollars grew thirteen percent. All of that can be true at once, and a share price measures fear, not proven cause. But a sector that restructures itself around AI skills, lets thousands go, partly because reskilling no longer works, and gets repriced by the market on AI fear is not behaving like a sector that believes nothing is happening.

the downside the pioneers prefer to skip

A warning belongs here, and anyone who leaves it out is selling you something.

Security firm Veracode tested more than a hundred language models on eighty coding tasks and found that 45 percent of AI-generated code contains a known vulnerability. That figure has been essentially flat for two years: given the choice between a safe and an unsafe route, a model picks the wrong one about half the time. It is a benchmark from a vendor with a stake in the outcome, so read it as direction, not as a final score. But the direction is confirmed elsewhere. A study of 302,600 verified AI commits across more than six thousand GitHub projects, shared as a preprint and not yet peer reviewed, found that for every tool tested, more than fifteen percent of commits introduce at least one demonstrable issue, that AI commits add one and a half times as many security problems as they fix, and that nearly a quarter of those issues were still unresolved at the last measurement. Technical debt at machine pace. And Daniel Stenberg, the maker of curl, software that runs on billions of devices, shut down his bug bounty program in early 2026 after six years: a fifth of submissions had become AI-generated, each one could cost three to four reviewers hours of work, and not a single AI submission ever produced a valid finding.

This is why speed alone proves nothing. AI lowers the barrier to building, but not the barrier to building something sound. Without craftsmanship, the new method produces the same mistakes as the old one, only faster and at greater volume. MIT researchers who studied why enterprise pilots stall arrive at the same point: it is rarely the model that fails, it is the organization that never learns how to build with it. I wrote about that gap between demo and service earlier in the production gap; everything in it has only become more true since. How sturdy the failure numbers themselves are is something I weigh separately in the 95 percent myth.

who the new builders are

Add the findings together and a profile appears. The new builders are not big teams with an AI license bolted on, and they are not prompters without a craft either. They are small, senior builders who bring two things at once: domain knowledge deep enough to direct a system, and the discipline to distrust machine work. Specification first. Agents checking each other’s work. Every change tested before it ships. And because the methods shift every month, a serious share of the time, in my case about half, goes into keeping up: testing what proves itself, and leaving the rest.

The shift is not coming, it is underway. The composition of the work has already changed, and the market is already repricing the old equation. Which means the question has changed for anyone commissioning software. Not: does my supplier use AI? Everyone uses AI. But: is there someone at the wheel who genuinely understands my process, who checks every change before it goes live, and who is still there when the system behaves differently on day ninety than it did in the demo? Ask that question, and you will recognize the new builders on your own.