The race for AI is on, but the gap between ambition and reality remains wide: plenty of POCs, few production deployments. At the Tech For Industry Show, André Queiros, Manufacturing Industry Lead | AI & Agentic(OutSystems) and Floriane Mignon, Head of Products (Suez Digital Solutions) showed how a governed low-code platform can bring AI into the core business of a major French industrial company.

To open, André Queiros sets the scene with two figures cited during the session: 71% of CEOs are said to have asked their CTO for a clear AI roadmap, and $2 trillion is said to have been invested by companies in AI technologies in 2025. The finding is twofold: the race is on, but moving from POC to production remains the real challenge.

According to OutSystems, the problem doesn't come from the models themselves, which have been available for a while, but from how AI lands inside the enterprise: rigid core systems that are hard to connect to LLMs and agents; business semantics (organizations, processes, data models) unknown to general-purpose LLMs; insufficient governance — only 21% of agents in production are said to genuinely comply with governance, compliance and security rules, according to an analyst figure cited during the session; and the risk of vendor lock-in, since betting on a single LLM is risky when the market shifts every six months.

OutSystems' stated mission: enabling French companies to become "AI native" while retaining their industrial sovereignty, through an open, governed application and agent development platform that decouples business logic from AI vendors — allowing companies to keep control over their data, governance and processes, regardless of the model used.

Suez, a French multinational with 40,000 employees present in 40 countries, is structured around two divisions: water services (treatment plants, distribution networks, sanitation) and waste (collection, sorting, recycling, energy recovery). Its Digital Solutions unit brings together around 600 data and AI experts, a European leader in remote meter reading (connected meters) and leak detection.

Across the waste value chain, digital technology is everywhere: trackers on waste containers, fill-level sensors to avoid unnecessary collection rounds, traceability of recyclable material flows (from copper cable to electric vehicle batteries) with 40% efficiency gains in some segments, and computer vision on conveyor belts — over 85% accuracy — to control the quality of incoming and outgoing flows and reduce downgrading.

In 2021-2022, facing a need for flexibility (test, fail, retry), adaptability to harsh industrial environments (dust, gloved hands, multiple sites) and a nearshore strategy, Suez ran a comparative test: the same POC built in hard code and on the OutSystems platform, with supporting metrics.

The verdict: 3 times faster with OutSystems, and, notably, 10% maintenance anomalies versus 30 to 40% observed with hard code. The result: from 2 applications on the platform in 2024 to 20 today — a tenfold increase — without adding developers, thanks to a factory in Portugal, and with declining maintenance costs.

The flagship use case presented: letting users freely query internal and external data about industrial sites through a secure internal LLM (Suez relies on GPT and Mistral), while guaranteeing that every answer complies with the regulatory framework — an absolute priority at Suez.

A concrete example: "Simulate closing a given site in 2028 and its impact on sites within a 50 km radius." The agent enriches the question with business rules — household waste cannot be redirected to certain categories of sites, nor beyond 40 km, nor above a regulatory maximum capacity — before sending the request to the LLM, then returns the answer visually on exportable maps. The agents natively embed the LLM connection, endpoint management, authentication and permissions: developers can focus on business rules.

First result: the first queries were placed in users' hands within 3 months. At 6 months, the project entered a verification phase: checking that agents stay within bounds and rating the relevance of the LLM's answers for the business. Other AI successes: OCR now reliable to 98% for customer billing.

Floriane Mignon stresses three safeguards: maintaining human oversight and a critical eye (never treating AI as 100% reliable), cybersecurity (covered here by Suez's foundational stack), and controlling LLM token consumption costs at scale — to the point of considering a dedicated agent for financial tracking.

Responding to audience questions, OutSystems clarifies its positioning: a layer that sits "on top" of every development approach — low-code, hard code, vibe coding (Claude, Cursor, Codex), private LLMs, Mentor (the platform's native AI) — to bring standardization, governance and compliance against shadow IT, whose risks lie in cost and security. The platform can connect to any MCP server or expose one, and orchestrates agents specialized by function, combining the probabilistic (LLM) with a deterministic layer that enforces the framework. One final prerequisite, stressed by Suez: upstream data quality and governance, worked on for a long time despite a history of growth through acquisitions.

The joint account from OutSystems and Suez Digital Solutions illustrates a concrete path out of the POC stage: a governed low-code platform, able to constrain AI agents with deterministic business rules, while keeping control of costs, security and regulatory compliance.

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