Jean-Marc Frangos spent 23 years running British Telecom’s venture arm in Silicon Valley. Now he drives digital transformation at Shimizu, a 220-year-old Japanese contractor and one of the country’s “big five.” In this episode he lays out how Shimizu put AI in the hands of thousands of staff in about seven months, the six-step playbook he wrote along the way, and why the oldest building on his campus is the one pointing at the future. If you run tech at a contractor, this is a rare look at what an at-scale rollout actually takes.
Most AI-in-construction stories are told from a startup’s side of the table. This one is told from the buyer’s side, at a scale almost nobody gets to see up close. Shimizu is a general contractor founded in 1804, listed on the Nikkei 225, with roughly 20,000 people on its books. It is exactly the kind of large, careful, margin-aware business that tech vendors say they cannot move.
Frangos is the person moving it. He came from telecoms, ran BT’s 250 million dollar corporate venture fund out of the Bay Area, and got hunted into construction almost by accident. What makes his account useful is that he has seen both playbooks up close, the Silicon Valley “break it and apologise later” model and the Japanese habit of refining something until it is close to perfect, and he is running an AI program that borrows from both.
For anyone weighing an AI rollout at a contractor, the value here is specific: what order to do things in, where the accuracy traps hide, and how to keep the whole thing from getting shut down on cost.
A veteran Silicon Valley exec is running AI at a 220-year-old Japanese contractor, and he has turned it into a repeatable playbook: fix your data, make many models easy to reach, get retrieval accuracy up, add agents with governance, and track cost before your board does it for you.
Two ways to build the future
Silicon Valley breaks things. Japan refines them.
Ask Frangos where he would rather run an innovation function, Silicon Valley or Tokyo, and he splits the answer. The Valley is unbeatable for anything software. Tokyo, he argues, is quietly excellent at physical AI and things that have to work in the real world. Neither is inferior, they just optimise for different things.
His favourite illustration is the Tokyo metro ticket gate. At rush hour, a river of commuters moves through the gates without slowing down, because the latency between the card reader and the gate opening is one of the fastest anywhere. Nobody rebuilt the system around iris scanning or face recognition. They perfected the boring parts: the read, the logic, the exception handling, the mechanics of the gate. That is Kaizen, improvement by repetition, applied to a problem most people would never notice.
The Valley reflex is the opposite: tear it down and ship something disruptive, sometimes without thinking about the 2,000 people queued behind you if it runs a second too slow. In a country of earthquakes, where a contractor builds things that must not fail, the careful instinct earns its keep. Frangos wants both, Valley speed for software and Japanese discipline for anything with real-world consequences.
How Shimizu actually rolled AI out
From a cautious Copilot trial to thousands of daily users
When the ChatGPT wave hit in late 2023, Shimizu did what most large firms did first: worried about security and data leakage. Once that was checked, the team started using OpenAI models inside their own Azure environment, a safe and familiar place to learn. Microsoft Copilot came next, and it is still in the mix, though on a reduced trial because it is expensive and Shimizu watches cost closely.
The turn came from an unexpected place. Shimizu’s venture arm had earlier taken an equity stake in Lightblue, a University of Tokyo spinout focused on computer vision. When the dust settled on the first ChatGPT frenzy, Lightblue offered Shimizu a single, easy front end that reached every major model, OpenAI, Anthropic, Google, Mistral, from one secure window. Trust was near-instant because the two companies already knew each other, and both are Japanese.
The takeup surprised even Frangos. Of the roughly 20,000 staff, about half suit these tools, and Shimizu converted around half of that group, near 5,000 people, in seven to eight months. For a construction company, where new tools usually crawl into use, that is fast.
If someone told you five years ago that we would do what we do today with AI, would you have believed it? Probably not.Jean-Marc Frangos, Digital Transformation and Open Innovation Adviser, Shimizu
The six-step AI playbook he wrote along the way
A field guide for anyone doing this at a real company
Frangos keeps a guidebook he rewrites almost weekly, because the ground keeps shifting. Stripped down, it is a sensible order of operations that any tech leader at a contractor can steal. The through-line is that the exciting parts, the agents and the chatbots, come after the unglamorous groundwork.
That last point is where a lot of programs quietly die, and it lines up with what we hear across the market. The ceiling on AI in construction is rarely the model. It is whether the buyer can absorb it, document it, and justify it, a dynamic we dug into in the quiet brake on ConTech AI.
The playbooks contractors are actually using
Deals, signals, and operator stories from inside construction tech, once a week.
Join 3000+ ReadersPast the chatbot: routing, agents, and the cost reckoning
The stuff that separates a demo from a strategy
One idea Frangos keeps circling is model routing. Inside a front end like Lightblue, you do not have to send every question to the most powerful, most expensive model. The system can pick the right model for the job, or split a longer task and run the pieces in sequence or in parallel. You do not reach for a bazooka to answer a simple question, and that discipline matters as costs climb.
He also splits two kinds of AI that often get blurred. There is chat-style help any employee can use by just talking to it, the “superpower for everyone” story. And there is predictive machine learning, which needs clean data and data scientists but can forecast things like a safety incident risk on a given day, or a supply chain problem before it lands. Both matter. They are not the same job, and getting retrieval trustworthy on construction documents, full of diagrams, tables, and BIM files rather than flat text, took months of tuning.
The robots are the other half of the story
Why Japan has been at physical AI since the early 90s
Software was where the conversation spent most of its time, but Frangos is clear that physical AI is the bigger prize for a builder. Japan has been working on construction robotics since the early 1990s, and the big contractors even share parts of their robotics research through a consortium so they can agree on common standards.
He frames robotics as three layers: the physical machine and its balance and strength, the operating system that lets a robot navigate a site or work as a fleet, and the “brain,” where a lot of progress is coming from simulation software from the likes of Nvidia. Today Shimizu’s painting and welding robots are commanded by tablets. The near-term goal is to talk to them in plain language, give them more autonomy, and have them report back that a job was done well.
That mirrors a wider land-grab. Silicon Valley money keeps circling heavy machinery and autonomy, a pattern we tracked when Uber’s founder bought an autonomous mining company. The difference in Japan is that this work has decades of patient engineering behind it, not just a fresh round of funding.
The oldest thing on campus points at the future
Onko Chishin, and why roots matter to a 220-year-old firm
Shimizu’s innovation center opened in 2023. In its courtyard sits a modest wooden residence the company built long ago for Shibusawa Eiichi, the Meiji-era figure widely called the father of Japanese capitalism, who mentored Shimizu’s early leaders. Every day, staff walking past it get a quiet message: remember where you came from, honour the craft of the people before you, and make them proud by building what comes next.
Frangos ties it to Onko Chishin, an old Confucian idea about learning from the past to shape something new. He likes it more than a mission statement written once and never read again. A forest, he says, draws its strength from its roots and grows toward the sky. For a company that has survived 220 years, that is not nostalgia. It is how they think about the next stage.
The lesson for the rest of us is not to copy Japan wholesale. It is that a serious AI program can sit comfortably inside a deeply traditional business, as long as you respect both the speed the technology demands and the caution the work deserves.
Copilot, Lightblue, and the tools behind the rollout
How the pieces of Shimizu’s AI stack compare
Shimizu did not bet on one product. It layered a few, each doing a different job. Here is how the main pieces stack up, based on Frangos’s account.
| Tool | What it is | How Shimizu uses it | Strength | Watch-out |
|---|---|---|---|---|
| OpenAI in Azure | Foundation models in a private cloud | Early, safe experimentation | Familiar, controlled environment | A starting point, not the whole answer |
| Microsoft Copilot | Assistant across Microsoft apps | Reduced trial for some staff | Deep Office integration | Cost per seat adds up fast |
| Lightblue | University of Tokyo AI front end | Main daily tool for thousands | One window to many models, easy to adopt | Needs tuning for high-accuracy retrieval |
| Predictive ML | Custom models on historical data | Safety and supply-chain forecasting | Sees risk before it lands | Data-scientist heavy, not plug-and-play |
He is Shimizu’s Digital Transformation and Open Innovation Adviser, based in Tokyo. Before construction, he spent more than two decades as a senior innovation executive at BT (British Telecom), where he built and ran the company’s 250 million dollar corporate venture fund and open innovation function out of Silicon Valley. (Source)
Shimizu was founded in 1804 and is one of Japan’s “big five” super general contractors, alongside Kajima, Obayashi, Takenaka, and Taisei. “Genicon” is Japanese shorthand for general contractor. Shimizu has roughly 21,000 consolidated employees and is a constituent of the Nikkei 225. (Source)
Lightblue is a University of Tokyo startup that builds AI tools for image analysis and language, with a stated aim of AI democratization. For Shimizu it acts as a single, secure front end to many foundation models. Shimizu had earlier invested in the company through its venture arm. (Source)
RAG, retrieval augmented generation, points an AI at a controlled set of documents, like your project files, so answers come from your material rather than the open web. It is hard in construction because the source documents carry diagrams, tables, drawings, and BIM files, not just plain text, which makes accurate retrieval difficult. Frangos says raw accuracy starts around 60 to 70 percent and needs tuning to clear 95.
Shibusawa Eiichi (1840 to 1931) is widely known as the father of Japanese capitalism. After travelling to Europe in the 1860s, he introduced joint-stock companies, modern banks, and the Tokyo Stock Exchange to Japan, and helped found hundreds of firms. Shimizu built one of his early residences, which now sits at the heart of its innovation center. (Source)
Do the boring work first. Frangos puts data cleanup, an easy multi-model front door, and honest cost tracking ahead of the flashy agent demos. Adoption at Shimizu moved fast because the groundwork made the tools easy and trustworthy to use, not because the technology was novel.
Shimizu Corporation profile and history
Jean-Marc Frangos, INSEAD bio
Lightblue, University of Tokyo AI startup
Shibusawa Eiichi, Japan’s moral capitalist
Japan’s big five super general contractors
Related Articles
“You Can Only Go As Fast As Your Customers”: The Quiet Brake on ConTech AI
Uber’s Founder Just Bought an Autonomous Mining Company
