The factory as institutional fact
What Searle's social ontology explains about factories, factory software, and the AI agents now entering both. Cross-listed in Business; builds on the summary of The Construction of Social Reality and on themes from my public writings.
1. The thesis
A factory looks like the most physical thing in the economy: steel, heat, torque. My thesis is that this is an optical illusion. A factory is mostly an institutional edifice in Searle's sense — a web of status functions layered on the physics: work orders, quality releases, batch records, inventory positions, routings, certifications. The physical plant executes; the institutional layer decides what the execution counts as. And almost everything we call “operations software” is, precisely, machinery for creating, storing and transferring institutional facts.
2. The shop floor, read through Searle
A work order is not a document
It is a declaration: this metal counts as job #4711, releasing these materials, obligating that machine and this operator. Delete the paper and nothing physical changes — yet production stops, because the institutional fact is gone.
The quality gate is a status function
Nothing chemical happens to a batch when it passes inspection. What changes is its status: 'conforming product, releasable to customer' — X counts as Y in context C. The physical batch is the X; the sellable product is the Y; the quality system is the C.
Inventory is collective belief
The ERP quantity is not a measurement, it is an institutional record — which is why cycle counts exist: they are reality checks on a belief system, and why 'inventory accuracy' is a meaningful KPI in a way 'reality accuracy' would not be.
Deontic powers run the plant
Who may release the batch, who must sign the deviation, who is authorized to override the schedule: strip away the machines and a factory is a dense web of rights, duties, permissions and obligations — Searle's deontology, wearing safety shoes.
The ontology is the hard part
Factory software fails not on algorithms but on institutional mapping: what counts as a batch here, when is an order 'complete', who can declare scrap. Every plant's C is different — which is why deployment is anthropology before it is engineering.
AI agents are new declarers
When an LLM closes a work order or validates a record, it is not producing text — it is exercising a status function. The question is not whether the output is correct but whether the declaration is authorized, and who bears its deontic consequences.
3. Why factory software is hard — restated
I have argued in my writings that the bottleneck of industrial software is the ontology — that deployments succeed or fail on whether the system's objects match how this plant actually works. Searle explains why that is the hard part. The physics of two factories are identical; their institutional constitutions never are. “What counts as a completed order” is not discovered by sensors — it is a local constitutive rule, negotiated over years, partly tacit, enforced by people with deontic powers the org chart only approximates. An MES rollout is therefore not digitization; it is a constitutional reform of the plant — which is why it meets the resistance constitutional reforms meet, and why the vendor's “best practices” land like a foreign legal code.
4. Enter the machine declarers
Until now, every institutional fact in a factory was ultimately declared by a human — software recorded declarations, it did not make them. Agentic AI breaks this. An agent that closes work orders, auto-validates quality records against a specification, or reorders material is exercising status functions: it performs Searle's status function declarations without being the kind of entity to which the plant's constitution ever assigned that power. Three problems follow. Authorization: nobody has decided, explicitly, which declarations an agent may make — the deontic grant is silently inherited from whoever wired the integration. Accountability: deontic consequences (the batch shipped, the deviation accepted) need a bearer, and “the model” cannot bear obligations — someone must own the agent's declarations the way a manager owns a delegate's signature. Drift: a stochastic system asked to apply a constitutive rule will, at some rate, declare X to count as Y when it does not — and unlike a wrong sentence, a wrong declaration changes the institutional world, releasing real material into real trucks.
5. The engineering consequence
This is why I have argued for deterministic code as the output layer of industrial AI: let models interpret, draft and propose in language, but let every status-changing act pass through a typed, auditable, deterministic interface — a registry of status functions that says which declarations exist, which principals (human or agent) hold the power to make them, under which conditions countersignature is required, and that logs every declaration with its bearer. In Searle's terms: keep the Background probabilistic, keep the declarations formal. The factories that get this right will get agents that compound; the ones that let free-text intelligence write institutional facts directly will get the industrial version of a constitutional crisis — some quietly, batch by batch.
6. The wider claim
None of this is special to factories — money, contracts and corporations face the same arrival of machine declarers. The factory is simply where the stakes are visible: the gap between “the record says” and “the metal is” is measured in recalls and injuries, not just restatements. Manufacturing, which formalized physical work a century before the office did, may end up writing the constitutional law of agentic AI first — the discipline that invented the quality gate now has to invent the deontic gate.