Redesigning (AI) Factories

Martin’s been turning over a question from this week’s reading: why doesn’t AI productivity show up on the bottom line? Azeem Azhar and Nathan Warren have been working through this in Exponential View, and the answer has everything to do with organizational structure, not technology choices.

The core finding: individual and workflow-level AI gains aren’t compounding into firm-level ROI. One public tech company’s senior exec put it bluntly — 1,000 engineers using Claude Code, and you can’t draw a line between that and meaningful business output.

The Installation Phase

Azhar and Warren build their argument on Paul David’s canonical account of electrification, which unfolded in two distinct phases. The Installation Phase ran from 1899 to 1919 — two decades of laying the infrastructure while the economy barely budged. Companies bought the technology, but used it to do what they’d always done, only slightly better.

Within this phase, David identified three sequential stages of adoption:

Stage 1 — The lightbulb: Individual task improvement. Electric lighting made factory floors brighter, but the layouts stayed the same — the same shafts, the same belts, the same steam-powered geometry. The technology bolted onto existing organizational logic. For AI, this is copilots and chatbots: faster writing, faster coding, faster research. The firm runs the same way it always did, only with faster people. Most companies are here now.

Stage 2 — Group drive: Workflow improvement. Individual motor drives replaced shared shaft systems, so you were free to arrange machines without respecting the old geometry. Production became more flexible, but the organizational logic still came first. For AI, this is agents automating multi-step processes — recruiting pipelines shrunk from weeks to hours, customer support handling two-thirds of chats autonomously (as Klarna does), procurement routing approvals without human touch. The firm still runs the same decision structures, but the structures operate more efficiently. A few companies have reached this.

Stage 3 — Unit drive: Reorganizing production around the new capability. Factories reoriented machines and workers around the workflow itself. For AI, this means agents that make decisions — signals observed, routed, and actioned without a human intermediary. The firm’s organizing logic changes. Almost no one is here.

The critical point: during the Installation Phase, productivity gains remained elusive. Companies spent heavily on electric infrastructure but saw flat output. The stages accumulated without compounding into growth.

Factory Redesign

The Factory Redesign Phase ran from 1919 to 1929 — half as long as the Installation Phase, but it was where the returns materialized. Ford’s Highland Park plant in 1913 had already demonstrated the new organizing principle: the assembly line, made possible because electric motors could be placed anywhere, driven by wire rather than shaft. By the 1920s, this logic spread.

US manufacturing productivity grew 5.4% annually through the decade. The technology hadn’t improved — it was already there. The growth came from reorganizing around what the technology made possible. The same electric motors that had merely replaced steam engines during the Installation Phase became the foundation for an entirely different way of working.

This is where most AI deployments are stuck right now. Companies are years into their Installation Phase — buying licenses, training employees, running pilots — but they’re still treating AI as a faster version of existing workflows. The Exponential View data confirms this: only 27% of executives say AI has met their ROI expectations. The spending is there. The compounding returns are not.

Where Companies Actually Are

The honest assessment: most firms deploying AI today are somewhere between Stage 1 and Stage 2 of the Installation Phase. They have copilots making individuals faster. Some have agents automating discrete workflows. Very few have reached the Factory Redesign equivalent — reorganizing decision authority so that AI can act without human intermediaries.

The congestion problem is real. A product team that can prototype in hours but needs days for sign-offs doesn’t ship faster — it builds a backlog. More AI workflows on a blocked decision pipeline makes congestion worse, not better. The Exponential View piece notes: “A firm running a fast loop against a bad model of the world loses money faster.”

Azhar and Warren call the missing structure the business operating graph — the machine-readable version of what middle management knows, some in well-organized systems, some in email, some in heads. One observation from Martin’s notes this week: Azhar’s team found their OpenClaw agents were dramatically more useful when they had this wider context. Not just “what task” but “what’s the current priority, who owns this area, where are the blockers.”

The bottleneck of a fast firm isn’t speed — it’s accuracy of the model. Get that wrong and you move faster in the wrong direction. Most companies haven’t reached the Factory Redesign Phase yet because they haven’t built the graph worth acting on.

The escalation structure matters as much as the execution structure. A support worker solves most complaints autonomously and escalates press or regulator issues — firms govern autonomy this way already, through permissions and thresholds. AI agents need the same. Until that structure exists, you’re still in the Installation Phase, burning money and waiting for the compounding to begin.

Sources: Why AI isn’t showing up on your bottom line · OpenClaw multi-agent architecture