How to Build AI Solutions That Actually Work on the Factory Floor

Eighty-five percent of manufacturing AI projects fail. But one venture builder operating inside factories has cut that failure rate to 30%, and is still improving. The difference isn't better algorithms. It's a fundamentally different approach to finding problems worth solving.

‍

In my conversation with Renan Devillieres on the AI in Manufacturing podcast, the founder and CEO of OSS Ventures laid out a framework for manufacturing AI that challenges nearly everything the industry assumes about how to adopt these technologies. With 22 companies spun out, 3,800 factories served, and 200,000 monthly users across his portfolio, Renan brings a rare perspective: he's been a McKinsey consultant, a factory director, a quality director, a tech startup founder in San Francisco, and now a venture builder focused exclusively on manufacturing. His thesis is simple and uncomfortable—most manufacturers aren't failing at AI because the technology doesn't work. They're failing because they don't have the right people, the right problems, or the right design.

‍

‍

Why Are 95% of Factories Still Operating Like It's 1995?

‍

The gap between tech-enabled factories and everyone else isn't about budget or ambition. It's about how the business is structured. Renan frames it through a comparison that stings: a Tesla car made in Austin costs roughly the same to produce as one from the Shanghai Gigafactory. Labor cost differentials barely matter when five people can produce what used to require fifty, because each person has enormous leverage through technology.

‍

Consider industrialization—essentially, coding the machines. At a traditional BMW factory, fifteen engineers program thirty machines. They're not particularly well paid, and they repeat the same programming work for every new model. At Xiaomi, five engineers paid half a million dollars each write code that deploys across twenty factories simultaneously. Divide those salaries by the number of cars produced, and the per-unit cost is lower than BMW's approach. The math is straightforward, but the organizational implications are profound.

‍

The pattern Renan sees across 900 factory visits is consistent: when you look at the executive committees of companies like Tesla, Xiaomi, or Michelin, at least two or three people understand code. At the other 95%, the person leading digitization has often never written a line of code in their life. As Renan puts it, "Would you give a factory director position to someone who never read a plan? So why is your digitization leader someone who has never written a line of code?"

‍

What Changed That Makes Factory Transformation Possible Now?

‍

Renan sees a direct historical parallel to the early twentieth century, when machines first entered factories. Before mechanization, three hundred people performed physical labor. Within a decade, factories operated with a fraction of the workforce, but those remaining workers programmed and managed the machines—and were paid more because they created more value.

‍

AI is creating the same dynamic, except the leverage is intellectual rather than physical. The examples are already live. At a factory in Istanbul, Renan's team is replacing fifty-five people performing research and development activities with five people and an AI system that handles the work project by project. The Chinese fast-fashion company Shein runs a supply chain department of roughly one hundred people for a thirty-billion-dollar company because AI handles the bulk of the work. These aren't pilot projects or conference demos. They're production systems operating at scale.

‍

The technology exists. The question is no longer whether this transformation is possible but how fast it spreads—and which factories survive the transition.

‍

What Does "AI Pilot Purgatory" Actually Cost Manufacturers?

‍

The hidden cost isn't the failed pilot itself. It's the compounding opportunity cost of solving the wrong problems. Renan is emphatic on this point: AI projects in production are expensive, which means you should choose a limited number of big problems, not scatter resources across small ones.

‍

He describes a conversation from three days ago with an aeronautics CEO running ten factories. The CEO had good news—two new clients. The bad news: thirteen thousand hours per month of industrialization work, including new machine programming and standard operating procedures. He couldn't hire two qualified people, let alone thirty. The AI opportunity wasn't abstract. It was the difference between fulfilling the contracts and losing them.

‍

The failure pattern Renan sees most often is what you might call the "chocolate-covered broccoli" problem—a term he picked up in San Francisco. Companies take an old system designed twenty years ago and slap a chatbot or copilot on top. "That is not innovation," he says. "It's laziness." The new paradigm isn't asking questions of a legacy system through a RAG interface. It's working with a digital worker that needs to be managed, taught, and trusted. That's a fundamentally different experience, and treating it as a feature addition to existing software is a recipe for rejection.

‍

How Should Manufacturers Actually Approach AI, And Why Do Most Get It Backwards?

‍

The most counterintuitive insight from Renan's experience across a hundred-plus AI deployments in the last two years is about where the real knowledge lives. Thirty to forty percent of the data points critical to an AI system exist only in the heads of frontline workers—not in ERPs, not in MES platforms, not in any system.

‍

At a socks factory where OSS Ventures is automating research and development, the team discovered 850 rules that lived exclusively in the heads of the R&D staff. These rules constituted the algorithm those people ran mentally to design new products. No system captured them. The only way to extract that knowledge was to go to the shop floor, work alongside those people, and document what they knew. The result: what used to take four to six months, eight iterations, and thirty-five thousand dollars in human time now takes one week, with 80% of new designs untouched by any human, at a cost of two thousand dollars.

‍

That's the 10x threshold Renan requires before launching any venture. If the technology isn't ten times better than the alternative—measured in numbers, not feelings—it doesn't get built. And critically, the team doesn't build what factory directors ask for. They build what they've validated through their own discovery process, then go back to factory directors with a specific proposition and specific numbers. If at least three out of ten agree to pay when it works, they launch. Fewer than three, they kill it. "People are nice," Renan warns. "They will say your idea is amazing. But what about the money? No money, but amazing. Don't do that."

‍

The other critical design principle: you never go from fifty-five people to zero. You go from fifty-five to five. Those five people become managers of AI agents. Their job is understanding the system, running quality gates, approving outputs, and nurturing the rules as the factory evolves. Designing that experience—making someone feel in control of a system of AI agents—turns out to be harder than building the AI itself. Renan thought AI might mean he needed fewer designers. He needs more.

‍

How Do You Actually Scale AI from One Factory to Hundreds?

‍

Scaling requires doing two contradictory things simultaneously and doing both well. The first is resisting the urge to scale prematurely. Renan advises founders to spend more time than feels comfortable on early-stage product quality—creating something users genuinely love that can be deployed easily. Trying to hyperscale too soon is a trap.

‍

The second phase is disciplined deployment at a level Renan compares to Toyota or the US military. His best-performing companies operate with granular deployment playbooks: on day one, these twelve data points are required; training happens in this sequence; version one deploys with this configuration; follow-up happens on this schedule. One portfolio company, Mercateam, went from zero to six hundred factories using exactly this combination—obsessive product quality followed by ruthless deployment discipline. The critical judgment call for founders is knowing which problems to solve with better product and which to solve with better process. Getting that wrong in either direction kills the scale effort.

‍

Underlying all of it is shared infrastructure. Every company spun out of OSS Ventures inherits a common technology layer that's already cybersecurity compliant and already connects to major ERPs. When an IT director asks whether the product integrates with their systems, the answer isn't "let me prove it over the next year." It's "we've deployed in 3,800 factories—here's how it works." That shared infrastructure removes what would otherwise be a twelve-month barrier for a two-person startup trying to sell into enterprise manufacturing.

‍

What Question Should Manufacturing Leaders Actually Be Asking About AI?

‍

The question isn't "how do we implement AI?" It's "where are the pockets of thirty-five people working with massive Excel files who are bottlenecking everyone else?" Renan's heuristic is disarmingly simple: open your eyes. Do a gemba walk. Find the manual bottlenecks where skilled people are doing repetitive intellectual work that prevents the organization from moving faster.

‍

The deeper reframe is about talent. Manufacturing leaders consistently tell Renan they have a talent attraction problem. He tells them they have a system problem. If the job involves programming three machines for mediocre pay, no one talented will take it. If the job involves managing AI agents across four hundred machines, creating millions in value, and paying half a million dollars—people will line up. The prediction Renan makes is striking: within five years, median manufacturing pay in the US rises 25%, and a quarter of new MIT graduates choose to work in manufacturing. That future doesn't arrive by making factories slightly more digital. It arrives by making them operate like tech companies—where fewer, better-paid people create dramatically more value through intelligent systems they understand and control.

‍