Act 3 · The Machine · Station 09
You are HQ
Run a café chain: delegate or centralize each store's hidden state, spend profits on AI upgrades, and find out when central control finally wins.
Every store has a real local state you cannot see — a neighbourhood's shifting taste, a flaky supplier, a barista carrying the rush. Delegate a store and its manager acts on all of it, but keeps some of the upside and ignores the chain. Centralize it and you decide — from a dashboard that is coarse, stale, and, early on, mostly wrong. Spend your profits on AI upgrades and find the week when the dashboard finally tells the truth.
Plan week 1
0 centralized · 2 delegated“Steady week. Read the room, adjusted the menu, kept regulars happy.”
No dashboard here — their knowledge is tacit. You see only the result.
“Steady week. Read the room, adjusted the menu, kept regulars happy.”
No dashboard here — their knowledge is tacit. You see only the result.
AI upgrades
budget $0Monthly coarse reports
Analyst (K̄=1) — refresh 1/week
What this is really about
The gap you feel between the dashboard and the result is the centralization frontier, lived from the inside. Early, with coarse telemetry (φ low) and a headquarters that can only act on one store at a time (K̄ = 1), the codified view is worthless and delegation wins — you are the Lange/Lerner planner, and you lose. As AI raises both channels, the frontier moves under your feet until central control wins strictly. That reversal is the whole thesis of Brynjolfsson & Hitzig, and it is not hypothetical.
The café chain is the paper's own anchor. In the 1980s Mrs. Fields Cookies built a headquarters expert system that prescribed store-level actions in real time. As co-founder Debbi Fields put it:
"We have removed the decision-making process from the store level. The manager's responsibility is to execute the plan – not to plan." Debbi Fields, quoted in Richman (1987); via Brynjolfsson & Hitzig, 2025, §7
The system “dictated when to mix dough, when to bake, which varieties to emphasize, whether to call in extra labor, and even when to hand out free samples, all based on live traffic forecasts” (Harvard Business School, 1990) — so that, in the contemporary write-up, “the system not only tells her what's happening, it tells the stores what to do about it” (Richman, 1987).
Walmart ran the same playbook at continental scale. “Beginning in 1987, Walmart linked every store to Bentonville via what was then the nation's largest private satellite network, giving headquarters real-time visibility into SKU-level sales. By 1991 it had launched Retail Link, an extranet that auto-generated store-specific replenishment orders and shared live data with suppliers.” The result: “Central category managers – not local store managers – decided exactly which items each outlet would stock, in what quantities…” (Lee, 2006; Fishman, 2006). Across U.S. retail, the four-firm concentration ratio “has risen from less than 15% in the 1970s to over 40% today” (Brynjolfsson et al., 2023).
And the comic relief that keeps the thesis honest: Anthropic's Project Vend “has already let an LLM make every key decision for a small online shop — albeit with mixed results. While the project failed to make a profit, making some comically bad choices along the way, one can imagine a future version doing much better” (Anthropic, 2025). Which is exactly what the next station is about: where centralization stops.