Copilots Were a Waypoint, Not a Destination

The copilot era did exactly what it promised, and that is precisely why the industry stopped one stop short. Assistance beside a human was never the end of the road. It was the last place the road still looked familiar.

The emblem of the last few years is a screen split down the middle. On the left, a person doing their work — writing, analyzing, drafting, deciding. On the right, in a panel that has become as standard as a scrollbar, an assistant offering to help: finishing the sentence, proposing the next block, summarizing the thread, drafting the reply. The person glances right, reads the suggestion, accepts it or waves it off, and returns to the left, where the actual work still lives. Multiply that small gesture across a day and it feels, unmistakably, like the future has arrived — a fluent machine at your elbow making every individual keystroke lighter. And yet at the end of that day the volume of work that had to pass through this one person is roughly the volume a person of the previous generation moved, because every item in it still routed through the same single pair of hands. The assistance was real. The throughput barely moved. Those two facts sit together uncomfortably, and the discomfort is the whole story.

For three or four years the technology industry organized itself around that split screen, and it was right to be excited, because the copilot was a genuine advance — the first time software felt like it could reason alongside a person rather than merely execute a fixed instruction. But somewhere in the enthusiasm the industry did a subtle and consequential thing. It mistook a waypoint for a destination. A copilot was a real place on the map, a necessary one to pass through, but a waypoint is by definition somewhere you move through on the way to somewhere else, and the somewhere else was never a better-assisted human. It was software that does the work. The industry planted its flag at the waypoint, declared the journey complete, and has spent a surprising amount of energy since insisting that the view from here is the summit.

The waypoint was real, which is exactly why it got mistaken for the arrival

The confusion is easy to understand once you notice how rare it is for a technology to under-promise and quietly deliver. Most of what gets sold to enterprises overshoots and disappoints; the copilot did the opposite. It offered a modest thing — a suggestion in the margin, a draft you could take or leave — and it made good on that modest offer almost immediately, which built a reservoir of genuine, well-earned trust. When something actually works after a decade of things that mostly didn't, the natural human response is to assume you have found the answer rather than a better question. The copilot's reliability was its own kind of trap, because reliability at a small thing feels like proof of the large thing.

There is a deeper reason the form was so seductive, and it has nothing to do with the software and everything to do with us. A copilot changed nothing about who was in charge. The human stayed in the driver's seat with both hands on the wheel and simply gained a sharper navigator in the passenger seat, and that arrangement was profoundly comfortable. It demanded no reorganization of the work, no new division of responsibility, no act of trust larger than accepting or rejecting a single suggestion you could see in front of you. Every existing role, every existing process, every existing org chart survived intact; the only thing that changed was that each person got a little faster at the steps they were already doing. A destination that asks you to change absolutely nothing is a very easy thing to accept as the destination, and so the industry accepted it, and stopped looking down the road.

Assistance keeps the human on the critical path

What that comfort obscured is an architectural ceiling, and it is worth being precise about, because it is the reason the productivity gains kept coming in smaller than the excitement predicted. A copilot is, by construction, something that sits beside a person and proposes while the person disposes. That posture is exactly correct for the kind of work where a human wants to stay hands-on continuously anyway, and it is exactly why the tools felt so good in the moment. But the same posture guarantees that the human remains the load-bearing element of the whole system. Every suggestion still has to be read, judged, accepted, and integrated by that one person, which means the copilot can make the person faster at each step but can never lift the work off them. The constraint on how much gets done was never the speed of any individual step. It was that all the steps still funnel through a single human, and assistance, however good, leaves that funnel exactly where it was.

The data has been making this point quietly for years, most sharply in the one domain where copilots took hold first and hardest. If you look at how software engineers actually spend their days, a SonarSource survey found developers spend under a third of their time — around 32 percent — writing or improving code, and an IDC report put actual coding as low as 16 percent of the working day. The tool that finished your line of code was operating with real skill on the smallest slice of the job, and even a dramatic improvement to that slice leaves the larger reality of the day — the reviewing, the coordinating, the reconstructing of lost context, the moving of work and knowledge between people and systems — essentially untouched. That pattern is not peculiar to engineering. In nearly every knowledge role, the assistance landed on the part that was already the fastest, because that was the part a person kept their hands on, and the part a copilot could reach. The rest of the work, the connective labor that consumes most of the hours, was never something you could suggest your way through, because it does not happen at one person's keyboard at all.

The destination is software that carries the work

Getting to the actual destination requires a different posture in kind, not a better version of the same one. Not a system that suggests to a person and waits, but one that takes responsibility for the execution around the work — reading what comes in, gathering the context from whichever systems hold it, deciding what should happen next, doing it, and bringing a human in only at the decision points where judgment genuinely belongs. The distinction is not cosmetic and it is not a matter of degree. A copilot assists the human who is on the critical path; autonomous software takes the human off the critical path for the routine majority of the work and puts them back on it precisely where they add the most, at the gates that touch money, risk, or a commitment to someone else. This is the difference between a faster passenger and a vehicle that drives, and it is the distinction that separates real autonomy from the thing being widely sold as its equivalent. Gartner has predicted that over forty percent of agentic AI projects will be canceled by the end of 2027, naming among the causes what it calls "agent washing" — copilots and rule engines relabeled as agents without the underlying change in what they can actually carry on their own. Agent washing is what mistaking the waypoint for the destination looks like when it is turned into a product: a new sign posted at the old stop.

The road past the waypoint is what a growing number of operators mean when they describe the move toward an autonomous enterprise — organizations that no longer treat each new increment of work as a new increment of human attention, because the connective labor is carried by software that reasons rather than merely proposes. It is the premise behind platforms built around Autonomous AI Workers rather than assistants: a Reasoning Core that runs specialist agents across the real work of a function, taking observations in and producing finished outcomes, with Human-in-the-Loop wired into the decisions that warrant a person and nowhere else. The point of a system like that is not to make the human at the keyboard faster. It is to remove most of the work from the keyboard entirely, so the human's attention goes where it was always meant to — to the judgment that a copilot could only ever assist and never absorb.

The mental model worth carrying forward is a map with two dots on it. The first dot is the waypoint, the place where the road still looked like the old road with a helpful passenger riding along, and where a person carried every piece of the work while something clever whispered advice. The second dot is the destination, the place where the work moves without a person's hands on each piece, and where the people are freed for the decisions instead of chained to the execution. The industry has spent years polishing the passenger seat and calling it arrival. The organizations that pull ahead will be the ones who stop measuring progress by how good the assistance has become and start measuring it by how much of the work now completes without anyone assisting it at all — which is the only measure that was ever pointed at the destination rather than the stop before it.