The Autonomous Enterprise Won't Replace You. It'll Promote You to Proofreader.

I came back from two weeks without a laptop. The maintenance window closed without incident. A couple of messages and that was it, zero stress, even though the Garmin said otherwise — but that was the holiday's own fault.

I caught up with the feed and ran into a number that looked like a victory. I pulled the thread and ended up somewhere else. This edition was going to be about something else, but I got something silly wrong and it ended up being about that.

The mistake

The CEO of Microsoft AI summarized on his feed a terrific essay about why AI will take far longer than people say to replace jobs. I read it, I was interested, I clicked the link and read the whole thing.

The essay wasn't his. It was by Daron Acemoglu, the MIT economist who won the Nobel in 2024. I knew that. I saw it with my own eyes when I opened the link. And when I sat down to write this edition, I had it noted as being by the Microsoft guy.

I checked, and the wrong name had survived anyway. The article had been filed in my head with the face of the person who showed it to me, not the face of the person who wrote it.

Why that stuck with me

Because that same week I saw the big version of the same phenomenon.

SAP published results from its models for making small, fast decisions: where to route an order, which invoice matches which purchase order, whether to escalate something or not. These are thousands of decisions a day currently resolved by a rule or by a person.

The technical blog is careful. It says it's early, that it's uncharted territory, and that the next step is testing it in production with data that changes and harder cases.

And it shows a table comparing itself against the competitor's model. Four rows:

Guessing at random gets you 47%.
The competitor's model reached 72.70%.
SAP's reached 76.75%.
And with some proprietary embeddings, 78.75%.

The table also has a column in the middle that says something important: the competitor's model answered without seeing any prior examples, and SAP's answered with examples in front of it. Like comparing two students when one studied with worked solutions and the other walked in having seen none.

In the summary that circulates, that column isn't there. What's left is the two numbers and the headline: the decision layer has arrived.

Nobody lied. The blog tells the complete truth and the summary tells a shorter truth. It's just that the version that travels is always the short one, and the first thing to fall when you shorten is exactly the part that let you interpret it.

The chain

Follow this to the end. A lab result gets summarized into a post. The post gets summarized into a conversation. The conversation reaches a committee as "this already works". And somebody approves automating an entire function.

At the end of that chain there's someone explaining why it didn't work out. And it isn't the person who wrote the blog. It's you, unless you saw the whole play before it reached the committee.

What the essay I almost misquoted actually says

Acemoglu argues that AI is going to take time (and that in ten years we shouldn't expect much more than 5% of what humans do to be replaced, an estimate of his from two years ago that he still holds), and not for lack of capability.

Electricity had plants generating in New York and London by 1881. It wasn't until the 1920s that half of factories used it. The bottleneck was never the technology: it was reorganizing the company to be able to use it.

And he adds the line that made me sit up: many companies are going to try to automate entire occupations, and that will remain impossible in the short term. Around here that has a commercial name and you probably know it: Autonomous Enterprise.

The 1% that ruins your day

Go back to the best number in the table. 78.75%. Read forwards it sounds good. Read backwards it says it's wrong one time in five.

Acemoglu tells a story that explains why that matters more than it seems. He injured both arms in 1997 and used dictation software for thirty years. The latest version promised 99% accuracy.

That 1% still drove him crazy, because it isn't a small error: it's an error that changes the meaning. A similar word in the wrong place and the sentence says something else. So you end up rereading everything, always, because you don't know which part is the 1%.

Now take that to one in five decisions in Finance. It doesn't eliminate the person approving payments. It turns them into the person reviewing every payment without knowing which of the five is wrong.

That's the real risk, and it isn't extinction. It's degradation. They don't take you out of the process: they leave you in it, but on the boring side. From builder to proofreader. And proofreading pays less, is valued less, and leaves you nothing demonstrable to take elsewhere.

The good part, which almost nobody says

The same thesis has a twist.

If the bottleneck isn't the model but reorganizing the company — reassigning tasks, redesigning processes, getting managers to understand what to ask AI for and how to ask for it — then the bottleneck is an exact description of our job.

The S/4 migration already proved the same thing: the technology had existed for years and the consulting market grew because of the migration, not in spite of it.

What changes is which side of the project it catches you on. Whoever signs off on automating an entire function eats the failure. Whoever says "this yes, this not yet, and this part with a control outside the model" gets paid the same and is also right.

Two things to keep in mind

Getting an author's name wrong is not the same as approving the wrong payment. Almost everything we shorten has no consequence at all, and living in a state of checking every source is an elegant way of getting nothing done. The point isn't to distrust everything: it's knowing which three or four figures of the year are actually going to reach you as a decision.

And Acemoglu has been forecasting slowness for a while. He's been fairly right, but he himself admits the models advanced faster than he expected. A forecast of slowness can be wrong too, and it leaves you looking as bad as the person announcing the apocalypse. So don't get comfortable thinking we're untouchable. There's no AGI but we're closer than we were a couple of years ago.

What's left

This week I lost an author's name after having read him, and a table lost a column on its way to a headline. Both are the same gesture: shortening something true until it can no longer be interpreted.

The autonomous enterprise is coming, slower and uglier than it's being sold, probably. But it's the roadmap and the "logical" thing given the technological advance we're living through. And like it or not, you have to prepare for it.

What decides which side it catches you on isn't how much you know about SAP. It's whether you open the source before forwarding the chart, whether you understand it and do things properly.

Pablo / En Producción

Sources:

SAP Community, "Jev vs. RPT: The Decision Layer Is Here".
Typed-decisions, huggingface.co/datasets/LocalLLaMA/typed-decisions
Daron Acemoglu, "Will AI Replace Workers? Not If We Build It Right", Humanist Review.