The digital transformation of industrial environments, driven by increasingly sophisticated enterprise resource planning (ERP) software, now requires new paradigms of interaction. The development of the future “Co-Pilot Assistant” by Infologic is part of this strategic initiative. The thesis, planned in collaboration with the company, involves developing an original hybrid agent-based AI architecture that relies on the combination of two pillars, following a “Generation and Verification” paradigm:
- Learning action policies from historical records (Generation): Rather than coding rigid rules, the first research focus aims to transfer the tacit expertise of the ERP’s best users to the agent.
- Secure planning through state exploration (Verification): To reduce the error rate in the face of task variability, the second research focus will equip the agent with a deliberative exploration capability.
In summary, this combination enables the agent to learn from past expert trajectories while drastically reducing hallucinations in the present through systematic validation via simulation. This architecture meets Infologic’s requirement for sovereignty while opening up prospects for major publications on the hybridization of neurosymbolic methods and the alignment of decision-making agents.