Bespoke at Scale: How Generative AI Is Rewriting the Economics of Manufacturing Automation

Author: Fabien Cros, Chief Data and AI Officer, Ducker Carlisle
Publication: Automotive Engineering (SAE Media Group)
Date: September 1, 2026

---

For the past four decades, automating a factory has meant one thing: buying into a large industrial software platform. Manufacturers turned to established suppliers such as Rockwell Automation and Siemens for the control systems, execution layers, and quality software that run their plants — millions of lines of rigorously engineered code, developed under strict standards and wrapped in guardrails. That rigor was never optional. When software commands a robot arm, a stamping press, or a paint line, a defect is not a bug ticket; it is a safety incident. The engineering discipline behind these platforms reflects that reality, and has served the industry well.

But discipline at that scale comes at a price. Platform deployments are measured in millions of dollars and years of integration work, which means they only make economic sense for the largest, most repeatable problems a plant faces. Everything else — the hundreds of smaller, plant-specific pain points that erode margins every day — has largely gone unaddressed. Running a modern factory became expensive by design.

That economic logic is now breaking down. Not because the incumbent platforms got worse, but because generative AI has collapsed the cost of building custom software. And that collapse is opening an enormous market that the big platforms were never priced to serve.

The Gap: High-Value Problems Below the Platform Threshold

Walk any production line and the opportunities are easy to spot. Quality engineers want automated weld inspection tuned to their specific joint geometries and acceptance criteria. Operations managers want to know which station is actually pacing the line this week, not last quarter. Safety teams want compliance verified continuously across every shift, not sampled during audits. Maintenance planners want early warning from equipment that predates the industrial IoT era.

Individually, none of these justifies a multimillion-dollar platform deployment. Collectively, they represent a significant and largely untapped source of productivity gains. Traditionally, plants had two options: stretch an off-the-shelf system to fit a problem it was not designed for, or accept the pain and move on. A bespoke solution, software written for this line, this product, this defect mode, was technically ideal but financially absurd. Custom industrial code was priced the same as custom industrial anything: high.

Generative AI changes the arithmetic. When a small team can generate, test, and document production-grade code in weeks instead of quarters, bespoke stops being a luxury.

The Proof Point: A 10x Cost Collapse

Ducker Carlisle saw this shift firsthand on a recent weld quality assessment project. A manufacturer needed automated inspection matched to its own welding processes and defect catalog. The quotes it received from established vendors were defensible pricing for heavily engineered, general-purpose systems, but far beyond what the business case could carry.

A bespoke alternative, built around pre-trained vision models, AI-assisted software development, and cameras already installed on the line, delivered the same capability at roughly one-tenth the cost. That is not a 10 or 20 percent saving that procurement can negotiate; it is a 90 percent structural cost reduction that changes which problems are worth solving at all. At platform prices, only the most severe quality problem justifies automation. At a tenth of the cost, many more become viable.

Three factors drive the collapse. First, foundation models arrive pre-trained: the computer vision that once demanded a bespoke research effort is now a starting point, not a destination. Second, AI-assisted development compresses the engineering itself: code generation, test generation, and documentation that consumed most of a traditional quote now take a fraction of the hours. Third, the hardware is already amortized. Commodity cameras, edge compute, and the plant network are sunk costs; the bespoke solution rides on infrastructure that platform vendors would have replaced.

The pattern repeats across use cases: line bottleneck analysis built from cycle-time data the plant already collects, safety checks running at scale on ordinary CCTV feeds, scrap-cause tracing stitched together from quality logs no packaged system was ever going to ingest. The common thread is specificity. These are not cheaper versions of platform software; they are solutions to problems platform software was never priced to touch.

The Data Is Already on the Floor

The second shift matters as much as cheap code: the raw material for these solutions already exists. Most plants are sitting on years of photographs taken at inspection stations, video from cameras installed for security or traceability, and operator notes buried in quality systems. Until recently, that archive was effectively dead weight. Extracting insight from it meant hiring a data science team, labeling data for months, and spending millions training models from scratch, an investment only the largest OEMs could contemplate.

Large language models, vision language models, and their smaller cousins have changed that. A vision language model can read an image from the line and describe a weld bead, a missing fastener, or a blocked walkway without a custom training program. Small language models, compact enough to run on edge hardware next to the line, handle the latency-sensitive and data-sensitive cases without sending anything to the cloud. Fine-tuning on a few hundred plant-specific examples now does what bespoke model development once did but at a fraction of the cost, in a fraction of the time.

The practical consequence: the first question in a plant AI project is no longer “what sensors do we need to install?” but “what do the cameras and systems we already own already see?”

Rigor Still Matters

None of this replaces the deterministic core of the factory. PLCs, safety instrumented systems, and certified motion control remain exactly where they are, doing what they do under standards written in hard experience. The new bespoke layer sits above that core — inspecting, analyzing, flagging, and advising where a wrong answer costs a re-check, not an injury.

That boundary makes the economics work. Software that observes and recommends can be validated proportionately to its risk: human-in-the-loop review, shadow-mode trials against known outcomes, staged rollouts line by line. Discipline is still required, and model drift, data governance, and cybersecurity do not disappear because the code was cheap to write. It is discipline sized to the application, not inherited from the safety-critical stack.

The builders change too. These solutions do not require a 200-person software division. A small team pairing process engineers who understand the defects with developers fluent in AI-assisted tooling can carry a project from pilot to production in a quarter. The scarce ingredient is no longer programming capacity; it is knowing exactly which problem on the floor is worth solving, knowledge that manufacturers have long had in abundance.

When Manufacturers Become Technology Companies

The most interesting consequence may be what happens after these tools are built. A manufacturer that develops its own weld inspection system, its own bottleneck analyzer, its own safety monitoring stack, owns something with value beyond its own four walls. Its peers likely face similar pain points, similar legacy infrastructure, and similar platform quotes. The natural next step is commercialization.

Normally, selling operational tooling to competitors would be strategically fraught. Here, it mostly is not, because these solutions are about making manufacturing better, cheaper, and more sustainable, gains the entire industry is under pressure to deliver. A supplier that helps its peers cut scrap or energy use is not handing over its product roadmap; it is monetizing a capability while strengthening the industrial ecosystem its own supply chain depends on.

Industrial companies have tried this pivot before. For a decade, CES keynotes have featured manufacturing executives announcing software platforms and digital ventures, and most of those efforts quietly fell short of the ambition. Perhaps the strategy was not wrong, but merely early. It needed a disruptive, general-purpose technology to make software creation cheap enough that manufacturers could build for themselves first and sell second, rather than trying to out-platform the software industry at its own game.

That technology has arrived. The next generation of manufacturing software may not come from software companies at all. It may come from the plants that needed it most, built it for themselves, and discovered they had something worth selling.

This article was first published in Automotive Engineering, SAE Media Group, September 2026.