At the Tech For Industry Show, Vincent Wang, Chief Digital Officer (Legrand), a global specialist in electrical infrastructure for buildings, presented a strong conviction: plugging artificial intelligence into existing processes produces, at best, only incremental gains. Real transformation requires completely rethinking how work is organized.

Vincent Wang opens his presentation with a historical parallel: a little over a century ago, factories were organized around a central steam engine, with a network of transmission belts dictating machine layout. When electricity arrived, the first reaction was simply to replace the steam engine with an electric motor — without changing anything else. The result: barely a 10% productivity gain, a situation that lasted about twenty years. Real change came when manufacturers dared to ask a different question: what would our factories look like if every machine had its own motor? It was this break — motorizing every machine and reorganizing lines according to the manufacturing sequence rather than the belt network — that caused productivity to surge.

For Vincent Wang, generative and agentic AI is in a comparable situation today: most companies, Legrand included, started out by plugging a chatbot or an AI layer onto databases and processes designed for another era — with results limited to a few incremental optimizations.

The proposed shift in mindset is to ask not "where can we add AI to what we already do," but "what would this work look like if we redesigned it with agents in mind from the start."

To concretely answer this question, Legrand set up a system called the Agent Factory, structured around three elements.

A common foundation for all teams: the agents themselves, connectors to the company's major systems, orchestration allowing agents to communicate with each other and with employees, tools for evaluating response quality, and FinOps management to keep AI usage costs under control.

The Agent Factory runs on four seasons per year. Each season addresses challenges proposed by a business sponsor (improving financial reporting, plant performance, compliance levels, and so on). For each topic, a "business lead" — an employee who fully masters the process in question, able to challenge the status quo and curious to learn how agents work — is paired with business experts and agent specialists ("agenters"). This small team owns the outcome end to end: redesigning the process, actually building the agents, then returning to their home entity to continue the rollout and measure impact.

Every creation is integrated into the group's architecture standards and folded back into the shared production line, so that future teams can reuse existing components rather than starting from scratch. This governance also makes it possible to map agent coverage against business functions, to steer the transformation as a whole. The Dataiku platform plays a pivotal role in this shared production line and in governing the agents created.

Vincent Wang illustrates this approach with a concrete example. Legrand manages around 500,000 product references, sold in more than 90 countries, with presentation standards varying by market, in 40 languages. Workshops held with data managers in each country helped measure the scale of the problem: writing a single reference line takes an experienced person between 3 and 5 hours — work that must be multiplied by 500,000 references, by 40 languages, and by the various distribution channels (B2B, B2C, with content sometimes customized for distributors like Rexel or Sonepar, or for marketplaces like Amazon).

To address this challenge, Legrand set up a virtual team called the Product Data Office, pairing a central orchestrator with several specialized agents: context understanding, content enrichment, local regulatory compliance, translation and data consistency. This team of agents, named Kaya, cut the time needed to roll out a complete catalog in a new country — usually 12 months of intensive work — down to 3 weeks, a 98% reduction in time to market. The data produced can be used across traditional channels (web, distribution platforms) but is also directly usable by the future agents of Legrand's partners, who themselves build AI to serve their own customers.

Asked about the place of employees in this system, Vincent Wang is clear: the agent teams didn't replace human teams, they upskilled them. Rather than spending hours correcting spreadsheet rows, employees now validate and correct the work done by the agents, and can measure their impact in terms of revenue generated by content quality, rather than a simple file-completion rate.

Vincent Wang sums up the central lesson from this approach: the question to ask isn't "which use cases should we add," but "how do we redesign the company's processes, governance, organization and skills."

On the technical side, Legrand uses common tools for individual productivity (such as Microsoft Copilot), and standard multi-agent architectures for redesigned processes — small agents with a precise scope rather than agents given too much responsibility, with explicit collaboration rules between agents (agent-to-agent type protocols or the Model Context Protocol). The choice of AI model used varies with task complexity: a small model for a simple task, a more powerful model for long, sophisticated reasoning.

On the obstacles encountered, Vincent Wang first cites the difficulty of concretely understanding what these technologies enable, which led Legrand to set up a maturity-building system inspired by Six Sigma (yellow, green, blue, brown belts) to support the progression of all employees. Regarding connections to legacy systems (ERP, CRM), Legrand relies on a data-structuring program run over several years, with access rights management already in place, and on a governance catalog listing agents already created, to encourage reuse rather than duplicating effort.

Vincent Wang's session illustrates a simple but demanding conviction: the value of agentic AI doesn't come from adding it to existing processes, but from the ability to completely rethink how work is organized around it. Legrand's result on product data management — a 98% reduction in time to market — illustrates the scale of gains possible when this redesign is carried through to the end.

Watch the full session, along with the other Tech For Industry Show replays, on our dedicated page.