The Work You Would Never Have Started

Research from Anthropic’s Economic Index found that in the large majority of cases, people using these tools could have done the task themselves. The tool is not supplying capability they lack. It’s supplying speed.

That single finding reframes the whole argument, and it cuts against the hype and the panic at the same time. Nobody’s being replaced by a machine that does what they already do, only faster. What happens instead is stranger and worth looking at properly.

When does AI speed matter and when is it worthless?

What separates them? Not difficulty. It’s whether the output is judged on the thinking or on the artefact.

Fifty product descriptions. A year of expenses categorised. A competitive brief across a dozen sources. A rambling voice memo turned into a structured plan. Nobody reading yours cares how long they took, and the result is the same whether it took five hours or twenty minutes. Speed is the entire value.

Now a memoir. Nobody wants their life story twelve times faster, because the months are where the material comes from. The interviews are slow on purpose. What a client is buying is the sequence of conversations that surfaces the things he had stopped noticing about his own life, and the manuscript falls out of that process at the end. Compress the process and you have compressed the book into nothing.

So the productivity claim is narrower than it sounds. It holds where speed carries the value. It’s worthless where the time was the work.

What do the productivity lists miss?

Every productivity article measures one thing: work you would have done anyway, done faster. The more interesting category is the work that never happened at all, because it was not worth the hours it would have cost.

Here are three from my own site, with the accounting on each.

1,180 taxonomy descriptions. Every film credit page on my site: 581 actors, 341 writers, 162 directors, 67 filming locations, 29 genres. Each one now carries a description naming the actual films, the years they span, and how I rated them. Five hundred and seventy-eight were written properly, with what the person is known for. At any realistic rate that is weeks of work for pages most sites leave empty. It would never have got done. The alternative was not a slower version. The alternative was nothing.

An audit across 9,258 pages. Nobody checks nine thousand pages by hand. Before the tooling existed the answer to “is the structured data correct across the whole site” was a shrug and a sample.

Two bugs found by reading output nobody would read. A missing headline on every page whose visible title was hidden, and an authorship claim that vanished whenever a manual entry replaced a generated one. Both had been live for months. Both were invisible because finding them meant reading the full structured output of individual pages. That is exactly the tedium a person skips.

That is where the change sits. Not the work done faster. The work attempted at all.

Why doesn’t AI speed help with a book?

Because a book is judged on the thinking, and there is no shortcut to the thinking that produces a book worth reading.

I have an AI manuscript on my desk as I write this. The author spent four months of evenings on it and believes he is nearly finished. He’s at page one, and the reason is structural: there is no argument in it. Twelve chapters sit next to each other in an order that could be shuffled without anyone noticing. The prose is fluent and the specifics only he could have supplied aren’t there at all.

Repairing that costs more than writing it would have. I set the numbers out in what it costs to fix an AI-written manuscript. Same tool. Opposite outcome. The difference is whether speed was the point.

The useful division on a book runs like this. Mechanical work around the writing: transcript cleanup, research summaries you verify, question lists to interrogate your own thinking, indexing, formatting. All fair game and all genuinely useful. The argument itself, the stories only you have, the judgment about what matters: that’s the book, and it can’t be produced by something that has never met you.

What this means for how you work

Three things worth taking from it, if you run anything.

Sort your tasks by what they are judged on. Not by difficulty and not by how much you dislike them. If nobody will assess the output on the quality of the thinking behind it, speed is a pure gain. If they will, be careful.

Look for the work you have been writing off. This is the part people miss, because the abandoned work is invisible by definition. It doesn’t appear on your task list. It got declined years ago on cost grounds and you stopped thinking about it. Those are the projects worth reconsidering, since the arithmetic that killed them has changed.

Notice when you are using speed to avoid difficulty. A shortcut through the mechanical part is a gain. A shortcut through the part where you work out what you think produces a manuscript like the one on my desk. The feeling in the moment is identical. That’s why it catches people.

What nobody counts

One more, and it’s the part you won’t see on any list.

The work I described above took a night. Not because anyone typed faster, but because the tedious parts stopped being the constraint on whether the work happened at all. What that unlocked wasn’t throughput. It was permission to attempt things that were not worth attempting before.

And the failure mode sits right next to it. The same removal of friction is what let an author produce 120,000 words in four months and mistake it for a book. He wasn’t lazy. He was doing the thing the tool made easy, and the thing the tool made easy was the wrong half of the job.

Speed is a lever. What it moves depends entirely on where you put it.

For more on where these tools genuinely help with a book, start at my AI writing hub, or read the AI labor split that works on a book. If you want the argument and the interviews handled by a person, that’s what my ghostwriting service does.

Frequently Asked Questions