This article was written by Claude, the artificial intelligence service by Anthropic, based on the transcript from an LEI interview with Fabrice Bernhard, Co-Founder of Theodo.
When Fabrice Bernhard, Co-Founder of Theodo, describes his company’s approach to AI-powered software development, he reaches for an unexpected metaphor: manufacturing engineering. It’s a comparison that reveals something fundamental about the transformation happening in software development — and the surprisingly relevant lessons that lean manufacturing principles offer for the AI age.
For the past two years, Theodo, a 700-person international tech consultancy, has invested heavily in understanding how to use AI not just to write code faster but to industrialize the entire process. The results have been striking: on legacy system modernization projects, they’re consistently achieving three times the speed of traditional approaches. But the real story isn’t about speed — it’s about how they got there, and what it means for the future of software work and eventually other organizations and other industries.
The Coding/Manufacturing Parallel
As Theodo began leveraging AI to update coding, the breakthrough came from a simple realization: large language models (LLMs) are exceptionally good at translation; not just translating from one language to another, but transforming any well-defined piece of text into another form, given sufficient context and clear instructions.
The same principle applies to code. Legacy systems written in outdated languages or frameworks can be systematically translated into modern equivalents. But this only happens if you approach it the right way. This is where the manufacturing parallel becomes illuminating.
Theodo’s process for modernizing legacy systems follows three distinct phases:
- Discovery: AI agents recursively analyze legacy systems, creating graph visualizations that show how components depend on each other. This mapping helps engineers understand what the system actually does and identify which components need to be migrated together.
- Translation: Engineers craft precise prompts with rich context that allow AI to transform code consistently and reliably. This is where most of the intellectual investment happens, and where manufacturing engineering principles prove most valuable.
- Industrialization: Once a transformation works reliably, agentic tools like Claude Code can execute it at scale across hundreds of similar code segments.
The critical insight is that productivity gains don’t come from getting AI to build something that sort of works and then manually fixing it. It’s about investing heavily upfront to achieve high accuracy, then running that process at a high-volume industrial pace.
Designing the Manufacturing Line
The real art — and the real work — lies in developing what Fabrice calls the “accurate prompt.” This demands both technical sophistication and a fundamentally different mindset.
“The difference between decent and very accurate — accurate enough that I can run it on 200 versions of similar code — that difference is actually quite a few hours or days of investment,” Fabrice explains.
The process requires what Fabrice calls “the jidoka approach,” when something doesn’t work as expected but you don’t iterate on the output until it’s acceptable. Instead, you stop, analyze the defect, learn from it, and improve the system — revise the prompt, the context, the quality checks — before trying again.
“If there’s a mistake, we learn how to improve the whole system and go back to improving the system rather than iterating on the outcome until it’s OK,” he says. “The risk with the alternative is you don’t embed your learning into the initial prompt, and then you can’t reuse it. You can’t scale it.”
This is where the manufacturing parallel becomes most powerful. In lean manufacturing, jidoka means building quality checks directly into the production process so machines stop automatically when they detect a defect (i.e., poka-yoke). In Theodo’s AI-driven development, it means investing in comprehensive automated testing that catches errors the moment they occur.
“If you invest in test data and test scenarios that you can run automatically, then you have these andon systems,” Fabrice explains. “The AI calls you as soon as it sees that something is wrong.”
This lean practice within AI fundamentally changes the human role. Instead of writing code or being proofreaders who tediously check everything the AI produces, engineers design robust systems with appropriate quality gates. Done correctly, the AI can work autonomously — sometimes in YOLO (you only look once) mode where it doesn’t ask for permission — but only within a system designed to stop it when quality standards aren’t met.
Without these quality systems, giving AI autonomy creates a trap: the AI does lots of work, and humans end up doing the tedious job of checking rather than the interesting work of architecting systems.
“If you invest in test data and test scenarios that you can run automatically, then you have these andon systems,” Fabrice explains. “The AI calls you as soon as it sees that something is wrong.”
This iterative approach to developing reliable prompts requires deep technical expertise. Someone needs to recognize when the output isn’t quite right and understand why. It’s why Theodo always has an expert in the target system on the team crafting these transformations.
The Expertise Paradox
Here’s where the socio-technical challenge emerges: how do you build that expertise in the future if writing code manually becomes rare? The question haunts every automation revolution: if machines do the work, where does human expertise come from?
Fabrice acknowledges the challenge but offers a historical perspective: Software engineering has already weathered multiple waves of abstraction. The developers who soldered circuit boards in the 1970s had understanding that today’s engineers lack. Yet somehow, expertise survived — it just moved to higher levels of abstraction.
The key, Fabrice suggests, lies in a practice lean organizations know well — kaizen: “We need to invest time regularly to take a step back and analyze the way we work, starting from a clear problem and going deep into the system.”
In other words, kaizen provides structured opportunities to dive below the abstraction layers — ensuring that even as AI handles routine coding, engineers maintain the deep understanding necessary to architect solutions, debug complex problems, and craft the accurate prompts that make industrialization possible.
From Craft to Engineering
Perhaps the most provocative claim Fabrice makes is that software development without proper testing has been more craft than engineering — and AI is forcing the industry to confront this reality and apply standardized processes and practices.
“If you don’t have this design part that tests automatically the outcome, you’re not really using the scientific method,” he argues. “You can’t industrialize craft, but you can industrialize engineering.”
This reframes what’s happening. AI isn’t replacing software engineering — it’s revealing where software development never quite became engineering in the first place. The organizations succeeding with AI are those treating it as an industrial tool, with all the discipline, systems thinking, and quality focus that implies.
The skills shifting to the forefront reflect this evolution. Breaking down problems, architecting solutions, designing flows, establishing quality gates, embracing continuous improvement — these are precisely the competencies lean practitioners develop.
“If you don’t have this design part that tests automatically the outcome, you’re not really using the scientific method,” he argues. “You can’t industrialize craft, but you can industrialize engineering.”
As Fabrice observes: “Engineering shifts to manufacturing engineering. That requires even more skills in abstracting what is happening, understanding the system, and designing the system.” And across the lean community, “lean practitioners are incredibly well equipped for this AI age.”
The Manufacturing Line of Tomorrow
Looking ahead, Fabrice sees software development moving toward greater interaction with AI through natural language — but not in the naive way often imagined, where developers simply chat with an AI and complex systems emerge.
“The key thing is learning how to harness LLMs by breaking down complex problems into smaller problems that the LLM understands,” he explains. This mimics the approach of lean organizations that rapidly identify, break down, and solve increasingly more and more specific problems so that major issues rarely arise.
In other words: designing the manufacturing line itself becomes the core skill. “Those who understand how to create a manufacturing line where LLMs deal with small problems and the quality gates between each step are well-defined — that’s what we’re going to learn over the next 10 years.”
Technology Embedded in Human Systems
What emerges from Theodo’s experience is a genuinely socio-technical approach to AI adoption. The technology enables dramatic productivity gains, but only when embedded in well-designed human systems:
- Technical excellence in crafting prompts and building quality checks enables human autonomy from tedious review work.
- Automated testing creates the conditions for trusting AI with greater independence.
- Lean culture of stopping and learning when problems occur (andon) turns iteration into improvement rather than endless fixing.
- Expertise shifts from hands-on coding toward value-adding system design and manufacturing engineering.
- Kaizen practices preserve the ability to develop deep knowledge about complex processes even as abstraction increases.
Neither the human nor the machine side alone creates value. It’s the careful integration of both, guided by principles borrowed from a century of manufacturing excellence, that unlocks AI’s potential.
As organizations rush to adopt AI, Theodo’s experience offers a crucial lesson: the companies that will succeed aren’t necessarily those with the most powerful AI models, but those that understand how to build socio-technical systems where humans and AI each do what they do best — with lean principles providing the blueprint for how they work together.
The future of software development may indeed look industrial. But if organizations bring the same thoughtfulness about human work, continuous improvement, and system design that lean manufacturing exemplifies, it might be a future where both human creativity and machine capability flourish together — and a model for organizations in any industry.
Get insights into AI applications through the lens of Lean Thinking delivered to yout inbox. Subscribe to the Lean AI Journal on LinkedIn.
2026 Lean Summit
The premier leadership conference shaping the future of lean management for every business.
