On 13 October 2026, AI agents were the focus of the user day held by AUFO, the French-speaking Oracle users’ association, in Paris at the LAHO Business Center. The stated theme was “Intelligence and data”. The afternoon included a workshop on agentic AI and the Model Context Protocol. Three days later, LeMagIT published an interview with the two co-chairs. Emmanuel Ruez said he knew of no large-scale project in production. There are attempts. The association has about 1,000 members across 150 companies. This is not a national survey. It is a user group’s report, and it is clear.
Prototypes, and a gap that will last
Ruez describes large groups that prototype, sometimes in an advanced way, and mid-sized companies for whom agents are still unexplored. Moving to scale, he said, remains difficult. Yohann Garcia, the other co-chair, who also directs Data and AI programs at École Polytechnique Executive Education, adds that strong heterogeneity will persist. A few of the vendor’s partners demo their agents, and likely use them in their own sales or engineering teams. That is not a deployment at the customers.
Garcia’s wording fits what actually exists better than the word “agent” does. He called it generative AI connected to tooling and MCP servers, close to an agent, and short of the autonomy that vendor roadmaps stage. The topics still open, as both co-chairs listed them, are degrees of autonomy, security, governance, compliance, access rights, token cost, and explainability. Ruez added a caveat: calling the phase purely exploratory would be excessive. No one, he said, is ready to conclude on what agents become.
Two barometers, two populations
On 1 October 2026, Forvis Mazars and ESSCA published the second edition of their AI barometer, based on 826 professionals in France, the United Kingdom, Germany, and Spain. In that sample, 35% of companies already use autonomous agents, and 7% of the whole sample use them intensively. Four companies out of five use language models for office tasks. Two thirds of respondents see shadow AI, meaning tools used outside what the organization has declared, and only 44% have formalized ethical guidelines. 65% have a dedicated AI team, and 57% train their staff. In France and the United Kingdom, the share of companies with a dedicated team moves from 49% in 2024 to 71% in 2026, and the share that train staff from 40% to 62%.
Workday’s AI@Work barometer for October 2026 surveys AI users, not every employee. In the French edition, 74.9% of the users surveyed say they hand complete tasks to AI with no intervention of their own. Trust in the employer’s handling of AI is 36.2%, the lowest level among the markets studied. Reported shadow AI falls from 66.8% to 46.2% in a month. Across the global barometer, 64.7% of executives say they understand autonomous agents well, against 25.3% of individual contributors. These percentages describe each study’s respondents. They do not compare directly with AUFO’s report. The difference in population is the finding: where AI is already in people’s hands, delegation is high and trust is low. Where the system of record is at stake, scale deployment is not there.
Frame the object that exists
A committee that asks for “an agent” after a vendor announcement often starts from an object that ERP users do not see in production. The field describes tooled generative AI, wired to tools and to MCP servers, with a human still in the loop because rights, cost, and error are not stable. Industrializing that object is narrower, and more tenable, than announcing general autonomy. The blockers are those of moving from PoC to production: ownership, data, integration, maintenance. At Programini, that reading matches the discipline of an AI Solution Factory: a written scope, a measure, a human recovery path.
The layer most often missing is the business meaning the tool relies on, the semantic layer. There are still cases where the right choice is not to deploy an agent. The deployment frame is on the AI and agents page. The questions the AI Office put to providers on 14 October, containment and monitoring after release, meet the same frame: EU AI Act, the first requests for information.
