The Factory Is Becoming a Software Routine

What if the most important consequence of AI is not that machines can think, but that work can finally run without waiting for someone to remember to start it?

That question connects two developments that are usually discussed in separate worlds. In one, an AI coding agent can be configured once and then invoked by a schedule, an API call, or a GitHub event. In the other, AI powered factories promise to turn inexperienced workers into productive machinists in roughly 30 days, while multiplying the output of scarce skilled labor.

At first glance, one is a software convenience and the other is an industrial strategy. In fact, they are instances of the same transformation: AI is converting expertise from a person you must continually coordinate into an operational system that can be triggered, repeated, observed, and improved.

That transformation matters far beyond coding or manufacturing. It may determine whether AI creates resilient economies with more capable workers, or merely concentrates power in a few companies that own the systems everyone else depends on.

The Hidden Bottleneck Is Not Intelligence. It Is Activation

Modern organizations are full of dormant capability. A company may have excellent engineers, a sophisticated factory, reliable suppliers, and valuable data, yet still perform poorly because every task depends on a chain of human activation.

Someone must notice that a pull request arrived. Someone must remember to run the tests. Someone must inspect the results. Someone else must decide whether the issue is serious. In a factory, someone must know which machine to use, how to configure it, how to interpret a deviation, and what to do when the experienced operator retires.

The organization does not lack intelligence in the abstract. It lacks a dependable way to turn intelligence into action at the exact moment it is needed.

This is why routines are more significant than they sound. A routine packages a prompt, a working environment, and the necessary connectors into a repeatable unit. It can be called by a schedule, an endpoint, or an event. The crucial change is not that an AI model has become smarter. It is that the model has been attached to the flow of work.

A GitHub webhook is a small example of a much larger principle. When a pull request arrives, the system does not wait for an engineer to open a dashboard and ask, “What should happen now?” The event itself becomes the trigger. The routine evaluates the change, runs the relevant checks, consults the repository, and returns a result in the same environment where the work occurred.

This is the software equivalent of a production line. Raw material enters, a sequence of operations occurs, and a usable output emerges. The work is not necessarily fully autonomous, but it is no longer dependent on a person manually initiating every stage.

The same logic explains why advanced factories can train newcomers so quickly. The factory is not asking a novice to absorb an entire craft through years of informal apprenticeship. It is embedding accumulated expertise into software, interfaces, machine settings, feedback loops, and standardized procedures. The worker still matters, but the worker is no longer expected to carry the whole operating system in their head.

AI becomes economically transformative when it is not merely available, but continuously activated by the environment in which work happens.

This distinction gives us a useful mental model. There are three levels of AI adoption:

1. Advice: A person asks an AI what to do.
2. Assistance: An AI helps while a person performs the task.
3. Activation: Events, schedules, and machine states cause AI to perform a defined operation inside a live workflow.

Most public discussion remains focused on the first two levels. The largest productivity gains may come from the third.

The Factory and the Code Repository Are Converging

The surprising connection between an AI coding routine and an AI powered factory is that both are becoming event driven systems.

In traditional software work, a repository is mostly a place where code is stored. In a more mature system, it becomes an environment that emits signals: a commit was pushed, a review was requested, a test failed, a deployment changed the system’s behavior. Each signal can initiate a specialized routine.

In traditional manufacturing, a factory is mostly understood as a collection of machines and people. In an advanced system, it also becomes an environment that emits signals: a tool is approaching a maintenance threshold, a component is outside tolerance, an order has changed, a material batch has arrived, or a downstream process is blocked. Each signal can initiate a specialized operation.

The common architecture looks like this:

Event, context, agent, action, feedback.

An event identifies what has changed. Context tells the system what the event means. An agent chooses or executes an action. Feedback reveals whether the action worked. The resulting data improves the next decision.

This architecture creates a different kind of scale. Human organizations traditionally scale by adding managers, specialists, and coordination meetings. Event driven AI systems scale by adding more reliable triggers, more relevant context, and more reusable routines.

Consider the difference between hiring another senior machinist and improving the system that helps every machinist diagnose a problem. The former adds one expert. The latter changes the productivity of an entire shift. Likewise, asking one elite engineer to review code faster is useful. Creating a routine that automatically evaluates every pull request changes the throughput of the whole engineering organization.

This does not mean expertise disappears. It means expertise can be multiplied rather than merely accumulated. The skilled worker becomes a designer of processes, a supervisor of exceptions, and a source of improvements that can be encoded into the system.

That is a crucial distinction in an economy facing shortages. If AI simply replaces workers, it may reduce payroll while leaving society with fewer pathways into meaningful work. If AI captures expert knowledge and makes it accessible to less experienced workers, it can expand the effective supply of capability.

The 30 day training model illustrates this possibility. A novice does not become an old fashioned master machinist in a month. Rather, the factory changes the definition of competence. The system narrows the set of knowledge the worker must personally memorize, provides immediate guidance, and makes correct action easier to repeat. The worker reaches useful productivity sooner because the institution has stopped treating every employee as an isolated apprentice.

The same pattern could apply to electricians maintaining data centers, technicians servicing advanced equipment, analysts monitoring energy systems, or operators managing logistics. In each case, the goal is not to pretend that expertise is simple. It is to externalize enough of the expertise that a larger population can safely participate in the work.

The Real AI Race Is a Race to Build Loops

National competition is often described as a contest over models, chips, or talent. Those matter, but they are inputs. The deeper contest is over who can build the fastest and most durable loops connecting intelligence to the physical world.

A model in a laboratory is potential. A model embedded in a factory, data center, supply chain, or engineering workflow is capacity. The difference between the two is infrastructure.

This is why innovation, infrastructure, and ecosystem cannot be separated. Innovation produces better agents. Infrastructure gives those agents energy, machines, data, networks, and places to operate. Ecosystems provide the developers, workers, suppliers, customers, and institutions that generate more applications and feedback.

A country may lead in software and still lose strategic ground if it cannot produce the physical goods required to deploy that software. It may attract brilliant researchers and still stall if there are not enough electricians, HVAC technicians, construction crews, machinists, or operators to build the surrounding systems. Conversely, a manufacturing base without modern software may possess machinery but lack the speed and flexibility required to compete.

The strategic unit is therefore not the AI model. It is the intelligence enabled production loop.

A production loop has four properties:

• It is repeatable: The same trigger can reliably initiate the same class of work.
• It is observable: The system records what happened and makes failure visible.
• It is extensible: New routines, tools, or workers can be added without redesigning everything.
• It is compounding: Each completed task creates data, learning, or infrastructure that improves future tasks.

An AI routine that runs once is automation. A network of routines that generates feedback and improves operations is an institution.

This helps explain why domestic manufacturing is a national security issue rather than merely an employment issue. Production capacity is a form of latent power. In a crisis, a country does not suddenly acquire the ability to make ships, drones, medicines, or precision components. It can only redirect capabilities that already exist.

The industrial strength that enabled large scale mobilization in the past came from a broad commercial base that could be redirected under pressure. A car company could produce aircraft. A watchmaker could produce navigation equipment. The important asset was not one defense contractor. It was a society with enough distributed production knowledge to change what it made.

AI may make that adaptability possible again, but only if it is used to widen the industrial base rather than narrow it. A factory that behaves like a data center can change product lines more quickly, train workers faster, and coordinate machines with greater precision. That makes manufacturing less like a fixed monument and more like a programmable capability.

The strategic advantage, then, is not simply having more factories. It is having factories that can be reconfigured, connected, and improved as easily as software systems.

The Danger of Mistaking Automation for Progress

There is an important tension here. The same systems that multiply workers can also centralize control over workers. A routine can liberate people from repetitive coordination, or it can turn them into passive executors of decisions they cannot inspect. An AI factory can create new jobs, or it can produce a narrow class of jobs while the highest value knowledge remains locked inside an opaque platform.

The difference depends on who owns and understands the loop.

A healthy system should not merely automate a task. It should increase the number of people who can comprehend, operate, modify, and improve that task. This suggests a better measure of AI progress than headcount reduction or raw output.

Call it the capability distribution ratio: the amount of organizational capability made accessible to ordinary workers compared with the amount retained exclusively by central experts or platform owners.

A high capability distribution ratio has recognizable features:

• New workers become productive quickly without being treated as disposable.
• Experienced workers spend more time improving systems and less time repeating instructions.
• Errors produce learning rather than hidden blame.
• Workers can see enough of the process to exercise judgment.
• Improvements made in one location can spread across the network.

This is also where software routines need careful design. A webhook that automatically launches an agent is powerful, but power without boundaries is dangerous. The routine needs permissions, audit trails, escalation rules, and clear definitions of what counts as success. In a factory, the equivalent is a system that knows which adjustments it may make independently and which conditions require human intervention.

The principle is simple: automate the activation of expertise, not the abandonment of judgment.

That principle also changes how organizations should train people. Instead of asking, “Which jobs will AI eliminate?” leaders should ask three more useful questions:

1. Which expert decisions are currently trapped in a few people’s heads?
2. Which parts of those decisions can be made visible, repeatable, and teachable?
3. What new judgment becomes important once the routine handles the predictable work?

The answer may reveal that AI does not remove the need for skilled workers. It changes where skill lives. Some skill moves into software. Some moves into system design. Some becomes the ability to handle exceptions, verify outcomes, and decide when the system is wrong.

Build Organizations That Can Start Themselves

The practical lesson for leaders is not to deploy more chatbots. It is to identify the moments when valuable work is waiting for a human to initiate it.

Start with a map of recurring events. A new customer request, an incoming shipment, a failed test, a machine anomaly, a compliance deadline, or a change in demand may each be the beginning of a routine. Then ask what context the system would need, what action it could safely take, and how a human would review the result.

A useful implementation sequence is:

1. Choose a narrow, high frequency trigger. Start with an event that occurs often enough to produce learning.
2. Give the routine a complete environment. Access to the relevant repository, records, tools, and permissions matters more than a clever prompt.
3. Define the boundary of autonomy. Specify what the system may do, what it may recommend, and what requires approval.
4. Capture feedback by default. Record inputs, actions, outcomes, exceptions, and human corrections.
5. Turn corrections into system improvements. Every repeated human intervention is evidence of a missing rule, connector, or training example.
6. Expand only after reliability is visible. A routine that works across many contexts is more valuable than a flashy demonstration that works once.

This approach applies to a small engineering team as readily as to a national industrial strategy. The goal is to create organizations that can start themselves, learn from themselves, and give more people access to the competence previously hidden in bottlenecks.

Key Takeaways

• Look for activation gaps, not just intelligence gaps. Find work that is delayed because no one has triggered the next step, then connect it to an event, schedule, or API.
• Treat AI as an operating layer. The largest gains come when models are connected to real tools, machines, data, and workflows rather than used as isolated advisers.
• Measure workforce multiplication. Ask whether a system helps more people become productive, not merely whether it allows fewer people to do the same task.
• Design for distributed capability. Make procedures, decisions, and feedback visible enough that workers can learn and improve the system.
• Build physical and digital infrastructure together. Software leadership cannot compensate indefinitely for shortages in energy, manufacturing, construction, maintenance, and logistics.

The deepest shift is easy to miss because it does not look like a single invention. It is a change in the unit of organization. We are moving from work performed by individuals inside static institutions toward work performed by connected routines inside adaptive systems.

The factory of the future will not simply contain robots, just as the software team of the future will not simply contain AI assistants. Both will contain networks of triggers, agents, machines, people, and feedback. The winners will be those that make these networks reliable while ensuring that their benefits spread through the workforce.

The decisive question is therefore not whether AI can do a task. It is whether a society can build systems that repeatedly turn intelligence into production, teach more people to participate in those systems, and retain enough human judgment to change course when conditions change.

A powerful AI model is like an engine sitting on the floor. A routine gives it a transmission. A factory gives it traction. An ecosystem gives it roads, fuel, mechanics, and destinations.

The future belongs less to whoever has the most impressive engine than to whoever can build the largest number of dependable loops between thought and reality.