Industrializing AI for Software Development: a roadmap applicable to every sector

This article was written by Claude, Anthropic's artificial intelligence service, based on the transcript of an interview by the Lean Enterprise Institute (LEI) with Fabrice Bernhard, co-founder of Theodo.

When Fabrice Bernhard describes his company's approach to AI-driven software development, he turns to an unexpected metaphor: manufacturing engineering. It is a comparison that reveals something fundamental about the transformation under way in software development — and about the surprising relevance that lean principles offer in the age of artificial intelligence

Over the past two years, Theodo, an international technology consultancy with 700 professionals, has invested heavily in understanding how to use AI not only to write code faster, but to industrialize the entire development process. The results are impressive: in legacy system modernization projects (old systems), the company has been achieving three times the speed of the traditional approach.

But the real story is not about speed; it is about how they got to this point and what it means for the future of work in software and, eventually, for other organizations and industries.

The parallel between coding and manufacturing

As Theodo began leveraging AI to upgrade its programming, the breakthrough came from a simple insight: large language models (LLMs) are exceptionally good at translation; not only between languages, but at transforming any well-defined text into another format, as long as there is adequate context and clear instructions.

The same principle applies to code. Legacy systems written in outdated languages or frameworks can be systematically translated into modern versions. But this only happens if the problem is approached in the right way. This is where the parallel with manufacturing becomes illuminating.

Theodo's process for upgrading systems follows three distinct phases:

• Discovery: AI agents recursively analyze legacy systems, creating graph visualizations that highlight dependencies between components. This mapping helps engineers understand what the system actually does and identify which parts need to migrate together.

• Translation: Engineers craft precise, context-rich prompts (instructions) that allow the AI to transform the code consistently and reliably. This is where most of the intellectual investment takes place and where the principles of manufacturing engineering prove most relevant.

• Industrialization: Once the transformation works reliably, agentic tools such as Claude Code can run the process at scale, applying it to hundreds of similar pieces of code.

The critical conclusion is that productivity gains do not come from asking the AI to build something "almost right" and then fixing it manually. It is about investing heavily up front to achieve high precision and then running that process at high volume — as on a well-designed production line.

Designing the production line

The real art — and the real work — lies in developing what Fabrice calls the "precise prompt". This requires technical sophistication and a complete change of mindset.

“The difference between something reasonable and something truly precise — precise enough that I can apply it to 200 versions of similar code — in practice takes a few hours or even days of investment”, Fabrice explains.

The process calls for what Fabrice calls a “jidoka approach”¹: when something does not work as expected, you should not simply iterate on the result until it is acceptable. Instead, you need to stop, analyze the deviation, learn from it and improve the system — revising the prompt, the context and the quality-check mechanisms — before trying again.

“If there is an error, we learn how to improve the system as a whole and go back to refining the system, instead of iterating on the final result until it is 'ok'”, he says. “The risk of the alternative is that you do not incorporate your learning into the initial prompt, and then you cannot reuse it. You cannot scale.”

This is where the parallel with manufacturing becomes most evident. In lean manufacturing, jidoka means building quality checks directly into the production process, structuring mechanisms that make the operation stop automatically when a deviation is identified, through error-proofing devices (poka-yoke).

In Theodo's AI-driven development, this translates into consistently investing in comprehensive automated tests capable of identifying failures at the very moment they occur.

“If you invest in test data and scenarios that can be run automatically, you end up with andon systems (signals for help)”, Fabrice explains. “The AI calls you as soon as it identifies that something is wrong.”

This lean practice applied to AI profoundly transforms the human role. Instead of writing code manually or acting as reviewers who exhaustively check everything the AI produces, engineers start designing robust systems, with clear criteria and structured quality checkpoints (quality gates).

When well designed, the system allows the AI to operate autonomously — sometimes even in YOLO mode (you only look once ), without requesting validation at each step —, but always within a flow that was designed to stop automatically when quality standards are not met.

Without these quality systems, giving autonomy to artificial intelligence creates a trap: the AI produces in high volume, but humans end up overloaded with the tedious task of reviewing and correcting, instead of devoting themselves to the higher-value work, which is designing and improving the system.

This iterative approach to developing reliable prompts requires deep technical expertise. Someone needs to be able to notice when the result is not right and, above all, understand the cause of the deviation. That is why Theodo always has an expert in the target system on the team responsible for structuring these transformations. 

The expertise paradox

This is where the sociotechnical challenge arises: how will this expertise be developed in the future if writing code by hand becomes increasingly rare? This is the question that haunts every automation revolution: if machines do the work, where will human competence come from?

Fabrice acknowledges the challenge, but offers a historical perspective: software engineering has already gone through several waves of abstraction over the decades. Developers who soldered circuit boards in the 1970s had knowledge that many of today's engineers do not have. Even so, expertise did not disappear; it simply shifted to higher levels of abstraction.

The key, according to Fabrice, lies in a practice widely known to lean organizations: kaizen. “We need to invest time, on a regular basis, to step back and analyze how we are working, starting from a clear problem and deepening our understanding of the system.”

In other words, kaizen offers structured opportunities to go beyond the surface layers of abstraction — ensuring that, even with AI taking over routine coding tasks, engineers retain the deep understanding needed to architect solutions, solve complex problems and craft the precise prompts that make the industrialization of the process possible.

From craft to engineering

Perhaps Fabrice's most provocative claim is that software development without adequate testing is closer to craftsmanship than to engineering — and that AI is forcing the industry to face this reality and adopt structured processes, standardization and consistent practices.

“If you do not have this design step that automatically tests the result, you are not really applying the scientific method”, he argues. “You cannot industrialize craftsmanship; but you can industrialize engineering.”

This changes how we interpret what is happening. AI is not replacing software engineering — it is exposing where software development had not yet fully structured itself as engineering. The organizations getting the best results with AI are those that treat it as an industrial tool, with all the discipline, systems thinking and focus on quality that this implies.

The skills that are coming to the fore reflect this evolution. Breaking down problems, architecting solutions, designing flows, establishing quality checkpoints (quality gates) and practicing continuous improvement — these are precisely the capabilities developed by lean practitioners.

As Fabrice observes: “Engineering is migrating to manufacturing engineering. This requires even more ability to abstract what is happening, to understand and design the system.” And, in the lean community, “practitioners are extremely well prepared for this AI era.”

The production line of the future

Looking ahead, Fabrice sees software development moving toward ever greater interaction with AI through natural language — but not in the naive way often imagined, as if it were enough to “chat” with the AI for complex systems to emerge automatically.

“The central point is learning to take advantage of LLMs by breaking complex problems down into smaller problems that the model can understand”, he explains. This mirrors the way lean organizations work: identifying problems, breaking them down and quickly solving increasingly specific problems, so that major failures rarely occur.

In other words: designing the production line itself becomes the core skill. “Those who know how to create a production line in which LLMs handle small, specific problems, with well-defined quality checkpoints between each step — that is what we are going to learn over the next 10 years.”

Technology embedded in human systems

What emerges from Theodo's experience is a genuinely sociotechnical approach to AI adoption. The technology enables significant productivity gains, but only when it is embedded in well-designed human systems:

• Technical excellence in crafting prompts and building quality-check mechanisms frees people from the repetitive work of reviewing.
• Automated tests create the conditions for trusting AI with a greater degree of autonomy.
• A lean culture of stopping and learning when problems arise (andon) turns iteration into structured improvement rather than continuous rework.
• Expertise is no longer centered on writing code by hand and shifts to systems design and to manufacturing engineering applied to development.
• The practice of kaizen preserves the ability to develop deep knowledge of complex processes, even in contexts of increasing abstraction.

On their own, neither the human factor nor technology generates sustainable value. It is the careful integration of both, guided by principles consolidated over a century of manufacturing excellence, that unlocks the true potential of AI.

As organizations rush to adopt AI, Theodo's experience offers an essential lesson: the companies that succeed will not necessarily be those with the most powerful AI models, but those able to build robust sociotechnical systems, in which people and technology work in a complementary way — each doing what it does best — with lean principles as the reference for designing and managing this integration.

The future of software development may indeed take on industrial characteristics. But if organizations bring the same rigor in thinking about human work, the same discipline in continuous improvement and the same care in systems design that characterize lean manufacturing excellence, that future can be an environment in which human creativity and machine capability evolve in a complementary way — becoming a reference for organizations in any industry.

Receive insights on AI applications from the perspective of Lean Thinking directly in your inbox. Subscribe to our monthly newsletter.