The Orchestrator

You do not need a proprietary survey to see the antipatterns, because they are visible in almost any organization that has started using generative AI. Somebody is shipping twice as much and understanding half as much of what they shipped. Somebody is paying a token bill that grew faster than any forecast and cannot say what the spend bought. Somebody is holding a meeting about “AI strategy” that never names a decision boundary. Somebody is tired in a way that looks like burnout and feels like review — not because they stopped caring about craft, but because they spent the entire day deciding which almost-right outputs were safe to keep. Technical debt is rising while the room stays irrationally exuberant, to borrow Alan Greenspan’s phrase from an earlier boom that also confused motion with progress.¹

I mean that last part literally. Repositories fill with generated code nobody quite owns, evaluation sets lag the prompts that keep changing, and “temporary” agents become permanent paths through production while the all-hands slide announces that the company is AI-native. Both things are true at once — the tooling is real and the debt is real — and the exuberance is precisely what makes the debt hard to name without sounding like you are against the future.

Klarna is the public case that already appears in this book’s introduction and in the chapter on delegation: announce that an assistant is doing the work of seven hundred agents, let cost become the dominant measure of the operation, discover that volume is not judgment, and then rebuild the human layer when complex interactions degrade and customers notice.² It has company now, and the pattern is public enough to study.

Duolingo announced an “AI-first” company in April 2025 — contractors phased out wherever AI could handle the work, new hires permitted only where a team could not automate — and spent the next few weeks absorbing the backlash until the CEO walked the framing back: AI was not replacing what employees do, the company was hiring at the same speed as before, and AI was a tool to accelerate the work.³ McDonald’s switched off its automated drive-thru ordering pilot at more than a hundred restaurants after two years of customers posting videos of the machine adding bacon to their ice cream and hundreds of dollars’ worth of chicken nuggets to their orders — while saying it still believed voice ordering belonged in its restaurants’ future.⁴ Amazon retired Just Walk Out from its US grocery stores — the “fully automated” checkout that, it emerged, had relied on roughly a thousand people in India reviewing transactions by hand, a characterization Amazon disputed, saying its associates annotated shopping data to train the models rather than watching shoppers to generate receipts.⁵ None of these are stories about models that were not smart enough. They are stories about orchestration that never happened — no escalation paths worth the name, no authority boundaries protecting the hard cases, nobody whose actual job was deciding what the machine could decide. The antipatterns below were visible inside every one of those reversals.

So if your company has not done anything as dramatic as Klarna, do not congratulate yourself too quickly, because each of these patterns has a modest costume as well as a theatrical one. Premature overadoption can look like a mandate and a press release, and it can also look like forty unofficial agents, three overlapping “AI working groups,” and a finance person who has started forwarding the token bill with the subject line left blank, because the honest subject was a curse word. The number on that bill has that effect on people. Solo Sovereign can look like a CEO who codes, or like a director who “just quickly” ships the prototype the team was still designing. Cost without purpose can look like a million-dollar line item, or like a week of agent work that produced a product nobody asked for and a demo everyone clapped for.

Several of the anti-patterns in this chapter first appeared, under sillier names, in two satirical field guides this author published earlier in the AI years.⁶ The names were jokes; the patterns were not, and they reappear here in plainer clothes because teams kept using them.