Two days ago, one of our portfolio startups we work with produced something we believe might be a first in its category. A complex physical product — with a full bill of materials, manufacturing logic, and digital passport — generated end-to-end by AI and sent directly to production. No human intervention between the initial prompt and the machine program that produced the part.
Prompt → product.
Not a demo. Not a prototype. A real product running in a real factory.
This article is not a victory lap. If anything, it’s closer to a field report — a description of what it actually took to make this work. As usual, I cannot disclose the company names or the client involved, but the mechanics are worth sharing for the builders working on similar problems.
Because the reality of AI in industrial environments is both simpler and harder than most people imagine.
The Product Was Not the Hard Part
The product itself was not trivial.
It involved a one-layer bill of materials with around ninety objects per product and multiple constraints: machine capabilities, material properties, production tolerances, and user experience requirements.
In other words, a design space with many degrees of freedom and non-obvious interactions.
But the true complexity was not the product. It was the knowledge embedded in the people who knew how to build it.
To make the system work, our engineers literally flew across three countries to capture the expertise required to manufacture this class of products.
Three main sources of knowledge emerged.
First, the designers’ intent. These were not CAD drawings or documentation, but rules in people’s heads: “if this product is used in this context, you must consider this constraint,” or “if the user experience needs to feel a certain way, that parameter must remain within a certain band.” There were 300 pages PDFs somewhere — they were conflicting, outdated, and definitely not used.
Second, the knowledge of a machine maker whose family had been building the machines for nearly a century. This was first-principles understanding of how the machine actually behaves — the difference between what the manual says and what physics does.
Third, the operational know-how of the factory itself. Thirty years of accumulated experience producing these products for a specific client.
All together, we extracted roughly 800 operational rules.
Not abstract ideas — concrete constraints.
And those rules turned out to be the real raw material of the system.
Reconstructing the Industrial Brain
Once the rules were gathered, we needed a way to represent them.
The solution was a knowledge graph.
Every component of the product — materials, machine parameters, tolerances, interactions — was represented as nodes and relationships.
We then mapped the historical product universe onto this graph: roughly 50,000 products, each with an average BOM of around 90 objects.
This reconstruction process was largely automated using agents.
Agents parsed historical BOMs, inferred relationships, validated compatibility constraints, and progressively rebuilt the causal network linking parts, materials, and machine programs.
What emerged was not just a database of parts.
It was something closer to a map of how the product category works.
Iteration: Where Reality Bites
Once the graph existed, we began running design scenarios with the teams.
Predictably, the first results were wrong.
Rules conflicted.
Some were incomplete.
Others were outdated.
Some were simply myths passed down through the organization.
Every day we iterated with designers, machine specialists, and factory operators.
The process looked less like coding and more like industrial archaeology.
Slowly the rule system converged.
Contradictions disappeared. Edge cases were added. Gaps were filled.
Only then did the system begin to behave predictably.
This stage matters more than people think. AI systems in industrial contexts are rarely limited by model capability. They are limited by the quality and consistency of the operational knowledge they encode.
The Unexpected Control Point
At first we thought the key object in the system would be the bill of materials.
It wasn’t.
The true control point turned out to be the machine program.
Once the machine program was generated correctly, a cascade of other parameters became trivial to infer.
Cycle time.
Machine utilization.
Raw material consumption.
Human labor time.
All of those are direct consequences of the machine instructions.
In other words, the machine program is the compiler output of the factory.
Once the AI could reliably generate that program, the rest of the operational logic fell into place.
This insight simplified the architecture dramatically.
Standard Work, Generated
Another surprising result was how easily standard operating procedures could be generated.
Once the machine program and process flow were defined, AI-generated code could automatically produce the associated SOPs.
Setup instructions.
Operator guidance.
Quality checkpoints.
In traditional manufacturing environments, these documents are written manually and maintained inconsistently.
Here they were generated directly from the process logic.
The SOP became a derivative artifact of the system, not a separate piece of documentation.
Sixty-Four Days
From the first site visit to the first working implementation, the entire process took 64 days.
This includes:
Knowledge extraction across three countries
Reconstruction of the rule system
Creation of the knowledge graph
Iterative validation with designers and operators
Generation of the machine programs
And production of the first product
Two months earlier, this system did not exist.
Two days ago, it produced its first physical output.
The Real Disruption Is Organizational
Now comes the interesting part.
The technology is not the hardest question anymore.
The organization is.
The client involved has roughly:
200 designers
55 full-time R&D engineers
For decades, their role has been to design products and encode the logic that leads to manufacturing.
When an AI system can generate designs, BOMs, machine programs, and SOPs directly from a prompt, the structure of that work changes dramatically.
Nobody — including us — fully knows what the future process should look like.
But this is not necessarily a negative story.
The client involved is one of the most innovative companies we know, and they care deeply about their people. If anyone can adapt intelligently to this shift, it is them.
Still, the magnitude of the change should not be underestimated.
What Happens Next
The system today works for one product.
Based on what we have seen so far, we believe the full product family could be automated within eight weeks.
Once the rule system exists, extending it becomes mostly an exercise in expanding the graph and refining the constraints.
The marginal cost of new products drops dramatically.
At that point, the factory begins to behave in a new way.
Products are no longer designed and then manufactured.
They are compiled and executed.
A Builder’s Take
After the last three years of hype around AI, it is tempting to oscillate between exaggerated optimism and exaggerated skepticism.
The truth, as usual, is more boring.
AI does change things.
But only when combined with painstaking work:
Extracting real operational knowledge
Structuring it into systems
Iterating with the people who actually run the machines
It is harder than most demos suggest.
But when it works, the effect is profound.
For the first time, we are beginning to see factories behave less like static installations and more like programmable systems — where intent can be translated directly into production.
We are still early.
But two days ago, a machine somewhere in the world produced a product that started as a prompt.
And that felt like a glimpse of the future.