What Holds the Bundle

01 · The apocalypse didn’t come. Something quieter did.

Joshua Rothman opens his New Yorker piece with three rival teams, each with an AI-fluent recent graduate.1 All three proposals come back fast, polished and full of data. Then two things go missing. The boss loses the signal that used to show who was good, because when everything is good, quality stops telling you anything. And the proposer loses conviction, because speed removed the slow working-through that made them sure they were right.

The aggregate numbers say the apocalypse hasn’t arrived. The Yale Budget Lab found no meaningful change in the occupational mix, or in how long people stay unemployed, for workers in highly AI-exposed jobs from ChatGPT’s release through March 2026.3 In May, Sam Altman said he was “delighted to be wrong” about how many entry-level white-collar jobs would be gone by now.4

But look closer and the entry door is narrowing.5

So the economy looks calm, the first rung is thinning, and inside firms the old ways of reading people are breaking.1 Rothman ends by saying the best we can do is ask the right questions. Judgment Is the Scarce Resource made the case that production got cheap and judgment got scarce. This piece tries to answer his questions.

02 · Work is a bundle, and bundles have glue

The book behind Rothman’s column, Messy Jobs by Luis Garicano, Jin Li and Yanhui Wu, treats a job as a bundle of tasks.2 Its key variable is the cost of separating those tasks.

In a weak bundle, you can pull tasks apart without losing anything. In a strong one, pulling them apart destroys value. The doctor who diagnoses should also prescribe, and answer for both. The vendor rep you negotiate with should still be there in two years, owning the implementation. The self-test is simple: if you’re better at some parts of your job because you do the others, your bundle is strong.

Here is the claim worth adding. In both of those examples, the glue isn’t skill. It’s continuity and accountability: the same party, answerable over time.

The question isn’t whether AI can do the task. It’s whether accountability can travel with the work when AI does part of it.

03 · The first fear: hollowing out

The first worry is bosses treating AI as cost-cutting. That is “bolted on” applied to people: automating the old org chart, task by task, without asking what held the tasks together. It unbundles strong jobs by accident. The work still gets done, and nobody answers for it anymore.

Four questions prevent that, and any leader can run them before automating anything.

1. Commodity or star work?Decide before you start. AI gets you to 90% quickly and often strands you there. In one study Rothman cites, artists using AI reached two-hour quality in thirty minutes, then stalled, because they polished before they had worked out the composition. That’s a method failure, not a ceiling: compose first, then let the machine finish.
2. Weak bundle or strong?If separating the tasks destroys value, automate inside the bundle, not across it.
3. If we could do three times the work, would anyone buy it?Upskilling saves jobs only where demand can grow. Health care can absorb more; many back offices can’t.
4. Who are our insiders?The authors argue the scarcest people right now aren’t CIOs or consultants. They’re employees who combine technical fluency, detailed knowledge of how the work really gets done, and a feel for the politics that resist change. BBVA gave AI first to its enthusiasts, rewarded good uses, and let change come from inside.

The method behind these is in The Loop Moved and The Gate Dissolves.

04 · The second fear: strong bundles going weak

The darker possibility is that AI keeps improving until strong bundles become weak ones. The way that happens is specific: tasks get cheap, and nothing carries identity and liability across the handoff.

An anonymous agent unbundles. The task gets done, and nobody answers for it. A known agent, bound to a person who does answer, can join the bundle instead. That is what the authors call “robots below”: the agent runs the routine path, and exceptions route to someone accountable.

Governance Comes Home shows the failure in miniature: a land seller whose agent answered buyers, while the person accountable for the price was no longer the one speaking. The fix there and here is the same: identity and authority that travel with the work. Neither You Nor a Stranger lays out what that requires.

05 · The unanswered question: where the next experts come from

Juniors earned trust by doing the easy work and being seen doing it. Along the way they built a social map: who is slow, who’s worth asking, when not to pitch the CFO. “Robots above,” where AI supplies the expertise juniors once escalated for, removes both the easy work and the escalations. The Stanford data suggests the effect is already showing up as fewer entry-level hires.5

There’s no full answer yet, but there are partial ones. Give juniors the composition, not the polish, so they work out the structure and let the machine finish. Route exceptions to a junior paired with a senior, so the hard cases become the training ground the easy ones used to be. A junior and an agent are the same design problem: a capable actor not yet trusted, placed at a seam. Bringing one along teaches you how to admit the other.

The IBM i world shows the stakes clearly. Its seasoned practitioners are textbook strong bundles, holding all three of those rare competencies at once, and the talent pipeline behind them is thin. When they retire, the glue goes with them unless someone has been deliberately handed it.

The sequel to Judgment Is the Scarce Resource, which argued that judgment became the only scarce input. This one asks what holds judgment to the work. The commercial crossing version of the same problem is Governance Comes Home; the companion on the human side is The Human Factor.