80% of SAP users in France have done nothing with AI yet, and AI agents will have to wait

“I don't know of any large-scale project that has been deployed. There are attempts.” This is not the view of an AI skeptic. It comes from Emmanuel Ruez, co-president of AUFO, the French Oracle users' association, interviewed by LeMagIT in an article published on October 16.

Yet enterprise software vendors have been promising for months AI agents able to automate entire processes. But the two major customer associations describe the same reality in France: so-called “agentic” AI remains at the prototype stage. That holds true on both the Oracle and SAP sides.

Prototypes, but no scaling up

AUFO has about 1,000 members from 150 companies. According to its co-presidents, large groups are prototyping, sometimes at an advanced stage, but struggle to scale up. Mid-sized companies are only just discovering the topic. This is not a quantified survey: these are members' accounts, relayed by their representatives.

What exists rarely resembles the autonomous agents seen in demos. “It's not fully agentic, it's generative AI connected with tooling and MCP servers,” explains Yohann Garcia, the other co-president. MCP (Model Context Protocol) is a protocol that lets an AI model call tools and data sources. The agents are semi-autonomous and limited to specific tasks. Most often, a human stays in control of each step or supervises the whole process.

On the SAP side, the figure is even starker

USF, the French SAP users' association, surveyed 276 of its members from 154 companies, roughly one third of its membership. According to LeMagIT, 80% of respondents say they have done nothing on AI. Their priority lies elsewhere: migrating from the legacy SAP ECC software to S/4HANA, a heavy undertaking that ties up their teams.

At SAP, the tone is different. Olivier Nollent, head of SAP France, believes the market is moving “from promise to reality.” He hopes to present customers next year who have deployed these agents at scale. As LeMagIT points out, this amounts to saying that these solutions have not yet been purchased, and therefore not yet deployed.

The obstacles cited by companies

If you are hesitating to launch an agent project, the open questions identified by LeMagIT make a good checklist:

• Autonomy and security: how far should an agent be allowed to act on its own in a system that handles orders, payroll or accounting?
• Access rights: an agent inherits the permissions it is given.
• Governance and compliance: who is accountable for a decision made by an agent?
• Explainability: being able to reconstruct why the agent acted.
• Cost: token consumption climbs quickly when agents chain calls together.

One bias should be kept in mind, pointed out by Yohann Garcia himself: companies share their successes more readily than their failures. Abandoned projects probably exist without being made public. The question now is whether the customers SAP showcases next year will be numerous enough to change the picture.