How work has to change around AI

In my last essay I argued that AI adoption mostly fails because nobody inside the company benefits from it. But suppose you fix that. Suppose the incentives are right and people actually want to use AI. I think most companies would then run into a second problem, which is that they'd put AI into the work they already have, and the work they already have was designed for people.

You can see this in almost any process. Take the way a marketing team produces content. Someone writes a brief, someone writes a draft, a manager reviews it, legal checks it, and then someone publishes it. The obvious place for AI is the drafting step, so that's where it goes. Drafting gets ten times faster, and the team produces about the same amount as before, because now everything piles up at review and legal. The slowest step decides how fast the whole thing moves, and speeding up one of the other steps just makes the queue in front of it longer. Eliyahu Goldratt wrote a whole book about this in the eighties, and it still describes what I see in most AI rollouts. [1]

So adding AI to a step usually doesn't change much. What changes things is asking what the process would look like if you designed it today, knowing that execution is now cheap and fast, and that the scarce things are deciding what to make and checking whether it's good.

In the last essay I used the example of factories and electricity. The gains came when factories stopped using one big motor and put a small one at each machine. That change also made the next big idea possible. Once power could go anywhere, Ford could lay out a factory around a moving line in 1913, and the work was organized around the line instead of the other way around. [2] People stood at fixed points along it and did one thing to each car that came past.

I think a lot of knowledge work is going to end up looking like that line, and it's probably worth being honest about it. AI agents will do most of the actual execution, like writing the code, producing the designs, drafting the documents and running the analysis. A smaller group of people will decide what goes onto the line, which is where strategy and creativity come in, and I'm fairly sure those stay human for a long time. And a large group of people will stand along the line checking what comes off it.

That last group is the part people don't like to talk about, but if you look at how agent-based work actually runs today, most of the human time already goes into reviewing. Someone reads the output, fixes what's wrong, approves it and passes it on. My guess is that for the majority of office workers, some version of this becomes most of their job, at least for a while.

When you think it through, I think roles split in two directions.

The first direction is much broader roles. When execution is cheap, the expensive thing is judgment, and judgment doesn't respect department boundaries. Someone who can tell whether a design is good, whether the code does what it should and whether the product makes sense can now do all three, because the agents do the building. A few years ago you needed a designer, a developer and a product manager, and a lot of meetings between them. Increasingly one person with good taste in all three areas can do it, and they can move a lot faster, because there's no handoff and nothing gets lost between people. This is very good news for entrepreneurs and founders, who were always expected to wear many hats anyway. The generalist used to be at a disadvantage compared to specialists, and I suspect that's reversing.

The second direction is much narrower roles, and they're mostly about checking. Part of the reason these roles exist has little to do with quality. When something goes wrong, a company needs a person who signed off on it, and an AI can't be held responsible for anything. Regulation points the same way. The EU's AI Act, for example, requires human oversight for high-risk systems. [3] So a lot of organizations will want a human in the loop partly so there's someone to blame, even when the AI would be right most of the time without them.

The interesting thing about these checking roles is what happens to the corrections. Every time a reviewer fixes an output or rejects it, that's a labeled example of what good looks like. Collect enough of those and you have exactly the dataset you need to make the model better at the thing the reviewer was checking. So the people in these roles are, often without realizing it, training their own replacement.

Self-driving cars went through this already. For years the robotaxi companies had a safety driver behind the wheel, whose job was to watch the car and take over when it made a mistake. Every takeover was data. In a growing number of cities those cars now drive with nobody in the driver's seat, because the safety drivers did their job well enough that it stopped being needed. [4]

I think the same thing will happen to a lot of checking work, and faster than most people expect. The checking produces the training data, the models improve, the error rate drops, and at some point it becomes hard to justify a full-time person reviewing output that's almost always right. Liability will probably follow the track record too. Once a system has been right a few million times, it becomes easier to insure it than to keep paying someone to watch it. So I expect these roles to grow quickly for a few years and then shrink even more quickly.

For companies, this means the real work is redesigning processes around the line and being deliberate about which roles sit where. It also connects back to the incentive problem. If you ask people to check AI output, and everyone understands that checking is training the system that replaces them, you need to be honest about that and have a plan for where those people go next, or you'll get the same quiet resistance I described last time.

For individuals, I think it means moving toward the broad side if you can. The useful skills are the ones that let you judge work across several fields and decide what's worth making in the first place. Learning to do one narrow task very well, in a way that's easy to check, is probably the riskiest position to be in right now.

If most of your week today is reviewing and approving things other people or systems made, it's worth thinking about what you'd do if that review stopped being necessary. For a lot of roles I suspect that will happen sooner than people think, and it's much easier to move while you still have the choice.

Notes

[1] Eliyahu Goldratt, The Goal, 1984. The idea is usually called the theory of constraints.

[2] Ford introduced the moving assembly line at its Highland Park plant in 1913.

[3] The EU AI Act includes human oversight requirements for high-risk AI systems.

[4] Waymo, for example, runs fully driverless robotaxi services in several US cities.