There will be no return to normal. That's the starting point of the Tech For Industry Show panel hosted by Véronique Rollin, Managing Director (Accenture), with 25 years spent supporting industrial supply chains — including helping Airbus monitor its supply chain during the A320 ramp-up — and Céline Dumas, Strategy Principal Director(Accenture), a procurement and supply chain specialist in consumer goods and retail. Their thesis: in the age of permacrisis, supply chain is no longer just an operational efficiency topic but a strategic priority, and AI changes the very nature of the problem.

What's changing with permanent crises isn't so much their intensity — stronger, more frequent — but their combination and unpredictability: the playbook from the last crisis doesn't work for the next one. The new normal is structural instability; supply chains therefore need to be structurally adaptable, while balancing cost, service level, product availability and CO₂ footprint.

Digital twins, control towers, risk assessment systems: manufacturers have equipped themselves. But these solutions require massive upfront investment and, above all, continuous updating effort — maintaining a real-time multi-tier map (tier 1 to 4) of suppliers, correlating the right risks at the right time. "We can do it, but at enormous cost." Tomorrow's challenge: managing risk at a lower cost.

Generative AI can tap into dormant unstructured data — contracts, bills of materials, supplier specifications — that current systems, confined to structured data, ignore. Agentic AI adds goal-based reasoning: breaking down a problem, autonomously fetching information, cross-referencing it and proposing a solution. Where a capacity constraint today triggers days, even weeks, of manual coordination (work in progress, open orders, customer re-prioritization, stock transfers), an agent can form a first response. And tomorrow, why not have a company's agent negotiate — under agreement — with a supplier's agent to find the best option for the industry and limit bullwhip effects.

Supply chain isn't just planning: it's also warehouses, trucks, boxes. Physical AI connects the digital twin to robots, AGVs, AMRs and the WMS: the digital system continuously monitors via sensors, and the execution system self-adapts — a delayed truck, a demand spike, and replanning happens continuously. Logistics is where the first cases are becoming reality, ahead of assembly: the autonomous warehouse "is going to become the new normal."

A third breakthrough, of a different nature: the speed of building applications thanks to AI — vibe coding, low-code, and spec-driven methodologies such as the BMAD framework used by Accenture's data & AI teams, which "mirrors" a development squad (UX/UI, data engineer, product). This makes it possible to very quickly prototype decision-support tools for planners — without replacing core solutions (ERP, specialized planning tools), but by creating an agile orchestration layer on top of them.

Developed for a major pharmaceutical company, the platform monitors inventory by SKU and by market — with agents continuously navigating it. The first detects weak signals: overstocking trends, forecast accuracy drift signaling upcoming stockouts. A second agent tests planning parameters (order frequencies, safety stocks), reruns the model and triggers recalibration — something no planner could do for every micro-change, given parameters that today remain fixed within S&OP/S&OE cycles.

Key point: no black box. One agent traces everything the others do, and a decision agent explains what was done. Agents are autonomous on micro-adjustments; on major variations, they inform, and the planner makes the call. Applied to industry, this continuous monitoring would make it possible to increase safety stocks ahead of time — instead of waiting until it's too late, as with the shortage crises the aerospace industry has faced.

Why keep long planning cycles — S&OP, S&OE, MPS, MRP — when you can contextualize in real time and evaluate scenarios in seconds? Tomorrow's cycle: sense, contextualize, evaluate scenarios, simulate, decide, act. Performance will no longer be measured by adherence to the plan, but by the speed at which the industrial system reacts and adapts — a process that has become a learning one, deciding in context rather than replaying past playbooks.

Gone is the planner triangulating five Excel files between two S&OP meetings: tomorrow, they'll reason by exception, understand what the system is doing, arbitrate major changes and continuously improve the system. A new role also emerges — the system builder, a concept borrowed from Amazon, which has automated its entire chain of systems — responsible for continuously updating systems and algorithms to handle exceptions.

The target architecture: a stable, standard digital core (the ERPs), composable AI workblocks layered on top, finely modeled by business function — and, at the center, the decisive layer: the knowledge graph, the ontology linking products, parts, designs, customers and suppliers. "That's where the intelligence of companies resides." Not forgetting responsible AI: an autonomous system needs limits — tracing everything, and guaranteeing that humans can regain control at any time. A "bounded freedom."

The closing message: don't look for where to "sprinkle AI" onto your current processes — that's the recipe for POCs that never scale. Start from the big decisions that are hard to crack: the example cited, a client facing significant obsolescence on certain SKUs, shows that by starting from the decision, the agentic system goes beyond the single inventory use case to embrace related functions. And as one final exchange with the audience underscored: beyond cost and efficiency gains, AI reduces teams' mental load by absorbing the mass of information — and lowering the error rate.

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