The work didn't disappear. It moved.

This April, after a KSELUX seminar in Luxembourg, I ended up comparing notes on AI at work with Sunsick Yoon, a database engineer at Amazon, and his son, a student. I offered my example: I never liked preparing presentation slides. The narrative was fine - the production was a chore. Now, with Claude for structure and image models for the visuals, I actually enjoy making them.

Sunsick summed it up better than I could: "AI helps a lot with many things, but it's not one click. You still need to do a lot of work for the outcome."

His son's contribution was quieter and just as telling. Sunsick and I have decades of pre-AI work behind us and were comparing how the tools had changed it. His son has nothing to compare - for his generation, working with AI is the norm they're growing into, not a change to adapt to.

The one-click line stayed with me because it names the gap between how AI adoption is sold and how it actually feels. And I think the gap has a precise shape.

The one-click myth

Every AI pitch implies the same thing: work gets deleted. Write my report. Build my feature. One prompt in, finished output out. The demos are constructed to sustain exactly this impression - and demos can sustain it, because a demo is a task where the work has already been done. Someone assembled the context, curated the inputs and stood ready to judge the output against a picture of what good looks like. The click was never the work. It was the last step of the work.

Then people adopt the tools, discover the click doesn't stand alone, and conclude AI is overhyped. Both the expectation and the disappointment miss what actually happened.

Where the work went

The work didn't disappear. It moved - to both ends of the task.

Upstream, into specification and context. Before the model produces anything useful, someone has to define what's actually wanted, gather what the model needs to know and decide what to leave out. I spend real effort maintaining the context my tools work from - project conventions, domain constraints, what good output looks like in my setting. That effort didn't exist in my workflow three years ago. It's now some of the highest-leverage work I do.

Downstream, into verification and judgment. The model produces; someone decides whether it's right, whether it's safe to use and whether it should ship. That decision can't be delegated to the thing being verified. In regulated and enterprise settings it also carries a name: accountability. The output is fast. Owning the output is not.

What collapsed is the middle - the mechanical production between intent and result. The typing, the formatting, the first draft, the boilerplate. That middle used to be most of the visible effort, which is why its collapse looks like the whole job vanished. It didn't. The ends grew.

I see this everywhere I work with AI. Slides: the layout effort is gone, but deciding what each slide must do to carry the narrative is now the whole task. Software: the code gets written fast, but the specification before and the review after have expanded to fill the space - and they are where the quality is decided. Documents, including legal ones: the drafting is quick; knowing what a clause must protect stayed entirely with me. Even with strong context in place, I remind myself almost daily that my judgment is the load-bearing part.

Two true sentences

This is why two contradictory-sounding statements are both accurate. "AI didn't reduce my workload" - true, because the total effort often stays comparable. "AI transformed my work" - also true, because the composition of the effort changed completely.

What disappeared was the mechanical middle, the part that was never the point. What grew was the part that was. That's also why the slides became enjoyable: not because they take less time, but because the remaining work is thinking work.

Building for the ends

Here's where it stops being a personal observation and becomes a product one.

If you're building an AI product and your design target is one click, you've designed for a user who doesn't exist. Your real product surface is the two ends: how the user gets intent and context in, and how they verify and trust what comes out. The teams that treat specification and verification as the product - not as friction to be hidden - are the ones whose products survive contact with real users. The teams that hide the ends behind a magic button ship a demo, and the user discovers the missing work at the worst possible moment: after they've relied on the output.

The line

AI didn't remove the work. It moved the work to where judgment lives. Products, teams and careers built around the ends - specification before, verification after - will absorb the shift. The ones that keep optimising the middle are optimising something that's already gone.

The generation entering work now seems to know this without being told. They never expected the click to be the work.

Where has the work moved in your own workflow - and does your tooling admit it?