Reinventing business processes through Artificial Intelligence : Danone’s progressive strategy

In this interview, Axel Droin shares the Group’s vision : Artificial Intelligence serving sustainable performance, operational efficiency and controlled autonomy, supported by a rigorous and evolving methodological framework.

How do you define an AI Agent at Danone ?

We adopt a fairly broad definition of an AI agent. It is the combined interaction between Artificial Intelligence, autonomous action capabilities and a user interface.

For example, a simple chatbot connected to a retrieval augmented generation (RAG) system can already be considered an agent. What matters is not necessarily the technical complexity, but rather the agent’s ability to execute a task within a reusable framework.

How are AI agents perceived within your organisation ?

We approach AI agents with significant caution, particularly to anticipate as early as possible the risks inherently associated with any emerging technology. This mainly stems from the fact that our understanding of the agent concept, and its potential impacts, remains imperfect.

What is certain, however, is that agents exist to support teams, and that they deliver their full value when integrated into a multi agent architecture orchestrated by a specialised coordinating agent whose performance is validated and monitored by human teams.

The orchestrator agent, responsible for coordinating and activating specialised agents, simultaneously represents the main technical challenge and the greatest source of value creation.

This approach transforms the linear execution structure typically associated with traditional automation methods such as RPA into an exponential efficiency model.

Could you walk us through the history of AI agent adoption within the company ?

Our journey began in the first quarter of 2024 with the implementation of chatbots equipped with RAG capabilities.

This pragmatic approach has provided us with two major advantages :

• Demonstrating value for business functions : we proved the added value of these solutions for global functions, particularly within controlling, internal audit and “Design Authority” type functions
• Revealing the underlying ambiguities of these models : the experimentation highlighted existing informational ambiguities. When information assumed to be present in a document is actually missing, the agent logically cannot use it, thereby exposing inconsistencies in document management. The challenge then becomes determining whether the performance drop is caused by missing information or by insufficient accuracy and performance of the language models themselves

We then decided to expand our experimentation by giving our agents action capabilities within a controlled autonomy framework.

‍

How do you concretely deploy these solutions ?

Our current methodology consists of analysing operational processes and identifying opportunities for partial or full automation.

This approach, inspired by RPA principles, has proven effective for initial testing but also reveals limitations when scaling deployment.

Our immediate priority is now the development of an Agent Framework designed to formalise the functional, technical and architectural principles governing our agents.

More specifically, we want to build the capability to reuse individual agents across multiple business processes, in the same way that data itself is reused across numerous use cases.

We first establish a coherent operational logic before addressing technical considerations.

For scaling, which is planned over the coming months, we rely on a three pillar approach : framework, training and functional scope.

At the same time, we are evolving our methodology to focus on atomic activities that can be “agentified” and then orchestrated, enabling scalable deployment from both a technical and financial perspective.

Where do you see the greatest potential for these AI agents ?

The application potential is universal, but certain areas already present immediate opportunities, particularly in application debt management.

Costly user interfaces, technologically obsolete environments and proprietary languages such as SAS, whose value often relies primarily on the scarcity of associated expertise, could see their management, maintenance and evolution significantly simplified, as agents make it possible to abstract the way organisations interact with these capabilities.

The immediate priority is therefore to understand how this potential will materialise within the company ecosystem.

To achieve this, the functions that must most rapidly adapt and anticipate these changes are architecture and cybersecurity teams. They must quickly rethink the organisation’s application and functional integration approach, while also redefining how concepts such as service quality, application performance and user support should evolve with the arrival of AI agents within enterprises.

To learn more, visit the website of Danone.

To explore these topics further and identify the impacts and opportunities relevant to your organisation, contact our partners and experts :

• Morand Studer
• Simon Georges-Kot

‍