I've Been Wrong About What Should Be Automated

A coworker of mine is friendly with AI. She uses it, and she's automated parts of her own job without anyone pushing her to. There are other parts she still does by hand, though, and those parts could be automated just as easily.

That puzzled me. I didn't think less of her for it. I just couldn't work out why. I've been chewing on it for a while now, and it's turned up some assumptions of mine I'm starting to question.

The automation imperative

Here it is: anything that can be automated or sped up should be, and the hours that frees up go to better work.

It's a default assumption, and I've very much had it myself. I've been treating the benefits of AI as self-evident, as if more productivity were an obligation in every situation, for every person, without exception.

Look at "better work." I haven't really been defining what better work is. I've been assuming that any work that can be automated or sped up is lower value by definition, or has no value at all. That's uncomfortably presumptive.

And then the question I skipped completely. Just because work could be automated or sped up, should it always be? I think there are several good reasons to say no.

Working by hand is how some people think

The reason she goes manual, as best I understand it, is pace. Doing it by hand keeps her inside the material long enough to process it, connect it to other things she knows, and let it settle. The thinking happens while the task is happening.

That cost doesn't show up anywhere, because the thinking was never a scheduled item. It happened during the work. Automate the work and those hours don't come back to you as thinking time. You reach the same decision sooner, and you understand it less well when you make it.

There's a stronger version of this. Some people arrive at their judgment through the act of producing, not before it. Change how the work gets done and you can close off their route to their own thinking. There may be another route. That doesn't mean it's easy, or that they'll like it, or that what comes out the far end is as good.

People inhabit their jobs

A lot of people aren't optimizing their job. They're inhabiting it. The doing is where the meaning lives for them.

I've made the pitch plenty of times: automate the doing so you can focus on higher value work. Put that way, I'm asking someone to hand over the part of the job they actually like, in exchange for something I've decided is better for them.

Imagine a portrait painter who spent twenty years becoming a master. She studied for it and sacrificed for it, made choices and investments around it, and built an identity and a sense of pride on top of all that. She loves the craft. Then photography arrives, and it's better at the one thing she staked those twenty years on.

I have a lot of sympathy for that painter. She isn't at fault for not wanting to become a photographer, and wanting to keep painting doesn't make her backward.

Her pride was never only in the likeness. It was also in what the likeness cost her: the twenty years, the training she paid for, the other paths she turned down to stay on this one. Photography doesn't make painting foolish. But for some people it puts a question mark over that whole investment and what it was worth, and that's a different injury than losing the income.

Production work is how people become senior

The junior version of almost every job is production. First drafts, first decks, first models, the pass a senior person sends back covered in edits. It's slow and expensive. It's also how a person gets good enough to stop needing the edits.

Automate all of the first passes and this year's throughput problem is solved. But the training pipeline goes with it. Keep having AI do all your junior work, and soon you won't have anyone with the skills to manage the AI.

When everyone runs their thinking through the same few tools

Route everyone's thinking through the same handful of tools and the output converges. Not bad output. Competent output, all pointed in roughly the same direction.

One of the things that makes people useful to each other is that we're not all the same, with the same strengths. Diversity gets better outcomes. Sand off everyone's rough edges and idiosyncrasies and you end up with a more homogenized system. Maybe we need humans for a little noise in it.

Work the machine can do that a person should keep

Up until now I've been sorting the work into two bins. Things a machine can't do, which stay with people. Things a machine can do, which get handed over.

I'd been missing a third bin. Work the machine can do that a person should keep doing anyway, for any of the reasons above. The thinking that happens while the task happens. The route some people take to their own judgment. The meaning that lives in the doing. The expertise that only gets built by producing. The variety that keeps a company from converging on one answer. There are probably more that I haven't gotten to yet.

A doctor could have a model draft the diagnosis. We'd still want the doctor to have actually thought it through. A correct diagnosis is not the only thing we're asking a doctor for. It's also the reasoning behind it, still live in her head, so she can catch the detail that doesn't fit the pattern, answer the question you ask next, and change her mind when the test comes back.

What a company gets from keeping people in the work

None of this is only about being good to people. There's a benefit to a company in thinking more broadly about when and why to keep a person in the work, and it shows up past the present moment.

Speed to market is usually a good thing. It gets customers and the business what they need. Focusing on what's needed right now is important too.

And right now, companies are well set up for exactly that. The knowledge equity they're holding, meaning the expertise sitting in their people plus the corpus, the written record of how the work actually gets done, is still a largely untapped resource. It's like sitting on a big oil reserve. They also have senior people with years of lived experience, and that experience is what lets them get a lot out of AI today.

Look past this moment, though, and there are other things to care for. Knowledge equity isn't like oil, in that it can regrow. It behaves more like a tree, reborn and growing every year, and it needs new growth going in to stay alive.

The risk is dilution. Take away the work where new expertise gets made and less goes in each year than went in the year before. It gets worse if the corpus starts feeding on itself, because a record built from its own earlier output lowers the bar a little every time around. Fresh tradecraft and fresh lived experience are what keep it healthy, and those come from people doing real work. Being more thoughtful about where a human stays in the loop is how a company mitigates that risk, and it may be part of the answer to the next-generation problem.

The upside runs the same way. People are a precious resource, and they deliver value in many different and bespoke ways. That variety is part of what drives quality and good business outcomes. People being happy enough to stay is part of the same value chain, because a person who leaves takes their share of the knowledge equity out the door with them.

So this is a system question, looked at across time. Where and how do people keep their hands on the wheel.

What I'm changing

I still believe there's a lot of value in maximizing business automation, and I'm excited about where it's going.

There's also a practical point here that has nothing to do with sentiment. It doesn't much matter whether the automated version really is better. If the person whose job it is doesn't feel that, then for them it isn't better.

So telling someone that every task they can hand off, they should hand off, is too broad a claim to make about somebody else's work. The narrower claim holds up fine. Some of the work is just friction, and no pride lives in it. That work should go. The rest is a real question, and the person doing it knows things about it that I don't.

What I want now is a clear set of criteria for deciding when a person keeps doing work a machine could do, decided up front and for stated reasons. I don't have the full list yet.