Agentic AI refers to artificial intelligence programs, called agents, that pursue a goal, use tools and carry out multi-step tasks, instead of just answering a question. In the supply chain, an agent can for example monitor supplier delivery delays, build a transport plan or prepare a purchasing tender, with humans keeping the decision. The field is still young: Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027.
Agentic AI: definition
There is no single definition of an AI agent. The recurring features are:
- goal-directed behaviour: the agent is given an objective, not a sequence of instructions;
- use of external tools: the agent queries a database, calls software, runs a calculation;
- action on its environment: it creates an order, a ticket or a plan, and does not just produce text;
- multi-step tasks, carried out without intervention at each step.
Most current agents run on a large language model, with memory, planning logic, connections to the company's software and orchestration software that coordinates several agents. The word "agentic" only became widespread in 2024. The Model Context Protocol (MCP), published in late 2024, has become a standard way to connect an agent to tools and data.
Predictive, generative and agentic AI and RPA: what is the difference
Speakers at Tech for Industry Show describe three layers that add up rather than replace each other:
- predictive AI (machine learning) forecasts: a failure, demand, a quality drift;
- generative AI produces text or images and lets people talk to software in plain language;
- agentic AI combines both to choose how to solve a problem, then act.
RPA (robotic process automation), for its part, executes fixed rules written in advance: it changes who performs a task without changing the task. An agent handles the steps that required judgement, such as reading a non-standard document or choosing between two suppliers. Gartner warns against "agent washing", the rebranding of assistants, RPA robots or chatbots as agents when they lack the capabilities: according to the firm, only about 130 of the thousands of vendors claiming agentic AI are real.
What agentic AI does in the supply chain
The use cases presented by manufacturers mainly concern repetitive, numerous decisions that are costly to coordinate by hand:
- supplier monitoring: spotting delays and weak signals across all sites, then proposing an adjustment of lead times or orders;
- transport planning: building the inbound flow plan from costs, routes and plant constraints;
- purchasing: preparing a tender, comparing bids with supplier data, proposing a decision to the buyer;
- planning parameters: recalculating safety stock and order frequencies as conditions change;
- product data: writing, translating and bringing into compliance the data sheets of thousands of items.
Gartner distinguishes decisions by risk: replenishment and order prioritisation lend themselves more to automation, while logistics network design and inventory policy will continue to require human judgement. For classic planning and its tools, see our guide Supply chain software.
Where companies stand
Analyst forecasts remain cautious:
- over 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls (Gartner, June 2025);
- at least 15% of day-to-day work decisions will be made autonomously by agents in 2028, up from 0% in 2024 (Gartner, June 2025);
- in supply chain planning, only 5% of organisations automating their planning will make at least 10% of their decisions autonomously by 2030 (Gartner, September 2026);
- in the same survey, carried out in late 2025 among 243 senior leaders of large companies, 83% had already spent at least 3 million dollars on automating their planning, including AI.
In other words, budgets are committed, but autonomous decision-making remains rare. The reasons cited: data quality, decision ownership, skills and system architecture.
What manufacturers said at Tech for Industry Show
At Tech for Industry Show 2026, at Paris Expo Porte de Versailles:
- Stellantis presented its agents across the value chain. In the supply chain, the Optima tool builds the inbound transport plan for around 40 plants and hundreds of suppliers, a planning task that took days by hand and caused errors. In purchasing, agents prepare tenders and compare bids with supplier data. In quality, an agent reads job sheets, spots anomalies and creates tickets, with 11 plants lined up for deployment. The group puts a simple agent at 0.5 to 1 dollar per interaction, against 5 to 8 dollars for a person, and up to 10 dollars for complex agents working together: hence cost tracking agent by agent. Read the article
- Sanofi is moving step by step: a correlation tool monitors the delivery delays of all suppliers across its 38 production sites and proposes actions, with users keeping control until trust is established, before one day letting the system adjust a lead time or place an order on its own. Read the article
- Michelin has been automating its shared services since 2018, with around 170 RPA robots in production. Some steps stayed manual because they required judgement, for example in purchasing; agents make it possible to cover them, provided the process is redesigned and people, robots and agents, including those of the group's other software, are orchestrated together. Read the article
- Legrand gave a team of agents the job of writing, translating and bringing into compliance the data of its roughly 500,000 items, sold in more than 90 countries: rolling out a full catalogue in a new country went from 12 months to 3 weeks. Teams validate and correct the agents' work. Read the article
The talks can be watched again as replays: Agentic AI: what concrete applications for industry, what ROI?, From RPA to agentic automation.
Risks and safeguards
An agent acts inside the company's software: its mistakes have real effects. The safeguards described by manufacturers:
- an identity and access rights: the agent acts with the permissions of the person who delegates the task, and that person remains responsible;
- named owners: at Microsoft, every agent has at least two owners, and an agent whose two owners leave the company is blocked;
- traceability: keeping a record of what each agent does and being able to explain each decision;
- graduated autonomy: autonomous on small adjustments, the agent raises an alert on large deviations and a human decides;
- data quality: false production data can steer an agent's decisions, a data poisoning risk pointed out by Velotic;
- cost: every interaction has a price, which rises with the complexity of the agents.
On regulation, the European AI Act (Regulation (EU) 2024/1689) does not create a category specific to agents: obligations depend on the use. Its rules for high-risk systems were postponed by Regulation (EU) 2026/1744 of 8 July 2026: they will apply from 2 December 2027 for the uses listed in Annex III and from 2 August 2028 for AI built into regulated products listed in Annex I, such as machinery.
The 2027 exhibitors offering AI agents
To meet at Tech for Industry Show 2027 (in alphabetical order):
- Aras Software: product lifecycle management (PLM) and digital thread platform, connecting teams to product data and AI agents.
- Celonis: process mining and process intelligence platform that gives AI agents the context of the company's processes.
- IFS: publisher of a modular platform (ERP, asset and service management) with built-in agentic AI.
- KBRW Adventure: French publisher of an agentic execution engine for orchestrating supply chain flows in real time.
- Neptune Software: a platform for building applications, processes and AI agents connected to the SAP ERP.
- Poka: a platform for frontline workers, with governed AI agents.
How to start an agentic AI project
1. Start from a decision, not the technology: pick a frequent, costly decision (replenishment, prioritisation, supplier follow-up) and the expected outcome.
2. Decide the level of autonomy: the agent proposes, assists or acts alone, depending on value, risk and the share of human judgement.
3. Prepare data and context: clean master data, a shared vocabulary, source systems (ERP, MES, WMS) that are actually used.
4. Redesign the process: an agent added to an unchanged process only brings marginal gains.
5. Set safeguards: access rights, named owners, traceability, human takeover.
6. Measure better decisions, not the number of agents: lead times, errors, avoided costs, and the cost of each agent.
Frequently asked questions
What is agentic AI?
AI programs that pursue a goal, use tools and act inside the company's software in several steps, instead of just answering a question.
What is the difference between an AI agent and a generative AI assistant?
The assistant answers a request. The agent plans and carries out a task, for example creating a ticket or preparing an order, with the rights it is given.
What is the difference between agentic AI and RPA?
RPA applies fixed rules written in advance. An agent also handles steps that require judgement. The two combine: the agent decides or proposes, the robot executes.
What are the use cases of agentic AI in the supply chain?
Supplier monitoring, transport planning, purchasing preparation, recalculation of planning parameters and product data management are the use cases presented by manufacturers at the show.
Is agentic AI covered by the AI Act?
Yes, depending on its use: the European AI Act applies to agents as to other systems. The rules for high-risk systems will apply from 2 December 2027 or 2 August 2028 depending on the case.
How much does an AI agent cost?
Stellantis puts a simple agent at 0.5 to 1 dollar per interaction, and up to 10 dollars for complex agents working together.
