"Agentic AI is everywhere, we talk about it a lot, but I don't yet see a concrete impact on my P&L." This line from a CIO, reported in the closing minutes, sums up the stakes of this keynote at Tech For Industry Show 2026 (Paris Expo Porte de Versailles). Moderated by Maxens and featuring Philippe (Microsoft) and Kenaz (Stellantis), it offers an honest status check on agentic AI: what is buzz, what is in production, and what remains to be solidified to scale.

According to Gartner, 80% of enterprise applications delivered in 2026 already embed at least one AI agent. But when it comes to production at scale, only one agent solution in nine is actually deployed — the rest is still in testing.

Agentic AI doesn't replace the existing stack, it complements it. Predictive AI (predictive maintenance, forecasting, control) remains in use. Generative AI shifted computing from the world of technicians to the general public: we now talk to machines in our own words, including by voice, and "old-style" human-machine interfaces (menus, sub-menus, clicks) are starting to disappear. The agent, finally, is a generative model with an identity (subject to the same authorization rules as a human), able to navigate the information system and use tools via a standard protocol, MCP. You can delegate tasks to it, which it plans and executes.

Agentic AI is an "orchestrator" agent able to talk to other specialized agents — accounting agent, compliance agent, forecasting agent, predictive maintenance agent, risk analysis agent… — assembled like a team of experts. Two cardinal principles: it's collaborative, and humans remain always responsible. The agent acts with the identity of the person delegating, who stays accountable for its actions.

Formulated in the 1960s by Melvin Conway, the law states that a company creates a product in the image of its organization. This produced "monolithic" architectures: for any transformation (changing ERP, PLM, CRM), you had to transform both the IT system and the organizational model, team after team, each limited to about ten screens. Agentic AI "shatters" this constraint: agents can navigate across silos, discover where data and the right functions are. The consequence is strategic: you can decouple work reorganization from IT transformation — start by reinventing how you work (business agility), then transform the information system. Whereas fifteen years ago, SOA and microservices failed for lack of making both levels agile at once.

For Stellantis, the agentic transformation is "not a choice." It serves the strategic plan (presented a month before the show), summed up by one requirement: do more, faster, with better quality and at lower cost.

The plan rests on portfolio simplification (from 14 brands toward 4 global brands and 6 regional ones), made possible by common foundations: a quality database and a business context layer on which agents rely. So the agent answering on Citroën's website does so in the Citroën context; on Leapmotor's site, in the Leapmotor context. With $60 billion of investment and 60 new products by 2030, automotive development cycles drop to around 2 years — a pace "almost impossible" to sustain without AI. Humans stay at the center, augmented by a "workforce" of agents under their responsibility.

Stellantis walks through its value chain:

The plan targets notably $6 billion in cost reduction and better quality. Key point: Stellantis created an agentic AI organization (started in 2024) and relies on "empowered" regions — AI teams close to the field to ensure adoption, the only real source of value.

Microsoft's message, as "customer zero" of its own technologies: to capture ROI, don't do incremental work, but rethink processes from the constraints and the target objective.

The first major transformation, nearly three years ago: consumer customer support (Xbox environment). Starting from strong constraints — keep the same people, improve all indicators (first-call resolution, customer and employee satisfaction, attrition), turn agents into advisor-sellers, and a challenge of €250M in annual savings within 18 months — Microsoft first identified the useful information, deployed agents to fetch it (improving source quality along the way), and built "collaborator" agents that listen to the conversation (with customer consent) and advise in real time. Result 18 months later: all indicators improved, beyond €250M. Only in a second phase, once the process was validated, were the now-useless intermediate interfaces and systems decommissioned — hence most of the savings (avoided maintenance, security, integration costs).

A second register, research: the multi-model agentic platform Discovery accelerates scientific discovery. Cited examples: a new solid-state battery designed in 3 months with 70% less lithium (16.7 million candidate molecules generated then filtered by chemist agents and AI simulation, before high-performance computing on about ten candidates); optimizing PC fan form factors (improved dissipation, zero decibels); and, in 8 weeks, a PFAS-free cooling fluid for data centers, a problem researchers had struggled with for 10-15 years. Another illustration: an "organic" cooling radiator, unattractive but far more efficient than the engineer-designed model, 3D-printed.

Asked about ROI, Stellantis distinguishes two measurable categories: cost reduction (decommissioning) and revenue generation (doing things faster). Human-related efficiency gains, however, don't translate directly into dollars: "when you're told 40% reduction, that's not how it works."

A major watch point: cost. A simple agent costs $0.5 to $1 per interaction; a human, $5 to $8; but complex orchestrated agents can reach $10. So the agent must be treated as an additional labor cost: decentralize responsibility (each domain manages its agents, not IT alone), and have very detailed FinOps dashboards to decide whether to "keep alive" or "shut down" an agent based on the value created.

Three improvement levers on Microsoft's side: hybrid models (models able to run on-edge in factories via Azure Local, at fixed cost; 11,000 models available, including open-source and fine-tuned, complemented by large models on prepaid consumption); human learning (stop defaulting to the best model for everything — "taking a Ferrari to buy bread 300m away" — in favor of an orchestrator that calls a local model first); and observability (KPIs on all agents, and ownership governance: at least two owners per agent, an alert if one leaves, a block if both leave the company — "we don't leave an agent without governance").

What's the difference between generative AI and agentic AI? Generative AI predicts words or images from a prompt. Agentic AI goes further: it's a model with an identity, able to navigate the information system, use tools (via the MCP protocol), plan and execute tasks, and collaborate with other agents — under human responsibility.

Why are so few AI agent solutions in production? Per the cited figures (Gartner), 80% of 2026 enterprise applications embed an agent, but only one agent solution in nine is in production at scale. Scaling requires quality data foundations, process reinvention and rigorous governance.

How much does an AI agent cost? According to Stellantis, a simple agent costs $0.5 to $1 per interaction, a complex orchestrated agent up to $10, versus $5 to $8 for a human — hence the need for precise, domain-by-domain FinOps.

How to measure the value of agentic AI? By distinguishing cost reduction (notably system decommissioning) and revenue generation (execution speed), without mechanically converting human efficiency gains into dollars.

This conference was recorded at Tech For Industry Show 2026, the B2B trade show for industrial transformation (Smart Manufacturing, AI, Supply Chain) at Paris Expo Porte de Versailles. Watch the full keynote video above, and join us on May 19–20, 2027, Hall 7, for the next edition.