Digital & AI Transformation Intern · LISI Group · January – July 2026
Deploying AI across a global industrial group, from the shop floor to the executive board.
- 0
- technical documents made searchable by AI
- Hours → minutes
- to find a maintenance procedure
- 5 days → < 5 min
- to approve a purchase request
- 0
- executives the robot was presented to
The context
LISI Group is a listed French industrial group, a leader in fasteners and components for aerospace (LISI Aerospace: Airbus, Boeing, Safran, Dassault Aviation) and automotive (LISI Automotive: Stellantis, Renault, Toyota).
I joined the group's Digital & AI Transformation department and worked for both divisions and the central functions: a cross-group vantage point.
The platform
Agents were built on Dust, an aggregator of LLM-based AI agents. My role: design the agents (prompts, system instructions), iterate them with the business teams and connect them to data sources or APIs.
Projet phare
The AI agent at the service of maintenance technicians
The problem
Facing a breakdown, technicians had to search for the right procedure across 45,000+ documents (manufacturer notices, wiring diagrams, manuals, service bulletins). It sometimes took hours, and some ended up repairing without consulting the documentation — with quality and safety risks.
The solution, in 3 pillars
Rethought document base
A SharePoint architecture based on tags and metadata (machine, document type, category, site) rather than folders, so the AI can navigate 45,000 documents efficiently.
AI agent
It understands a natural-language question ("Machine X, main motor overheating, which procedure?"), runs a semantic search and answers with the relevant documentation excerpts.
Mobile interface
Built with the plant's IT team so technicians can query the agent from their phone, right on the shop floor.
Technician (question)
Mobile interface
AI agent
45,000 tagged documents
Answer + excerpts
Before
- Manual search →
- Sometimes hours →
- Procedures not always followed
After
- Precise answer in minutes
- Better adherence to documented procedures
Result
In production at a pilot plant of the group, used daily by technicians. Roll-out to other plants was decided by the group, with the aim of being generalized.
Project management
About 2–3 months: weekly meetings between the AI team, the plant's IT team and technician representatives, plus 2 days on site to observe technicians and adapt the agent to field constraints (shop-floor network, speed, ergonomics).
The other projects
Checking contract compliance
- Problem
- Lawyers manually reviewed every clause of every contract to check compliance with group policy.
- Solution
- An agent automatically analyzes clauses, assesses their compliance and suggests a rewording or a negotiation strategy for risky clauses. Iterated weekly with the lawyers over several weeks.
- Result
- Contract-analysis time significantly reduced.
Labour-market analysis
- Problem
- Quickly assess how tight the market is for each open role.
- Solution
- API connections with France Travail and APEC (completed by a Power Automate flow), then an agent generates a report: talent-pool tightness (rich / tight / very tight), typical profiles, skills requested by competitors, HR recommendations.
- Result
- Being finalized at the end of the internship.
Before
- Employee →
- Department manager →
- Finance manager →
- Entry into the ERP →
- Purchase order
After
- Digital form
- One-click approvals
- Automatic notifications & archiving
Digitizing purchase requests
- Problem
- An approval flow by email chain, slow and hard to track.
- Solution
- A digital form (browser or mobile) with one-click approvals, automatic notifications and archiving. Built on Power Apps + Power Automate, self-taught, in about 2 weeks after scoping with purchasing.
- Result
- From several days to under 5 minutes for the whole approval flow, in a pilot phase with about ten employees; generalization considered.
Autonomous AI-driven robotic arm
- Problem
- Prove that repetitive operator gestures can be automated with a low-cost robot.
- Solution
- A "leader" arm handled by a human, a "follower" arm that learns by imitation, cameras for vision. Within the team: built the data-processing logic and the datasets (100 to 200 demonstrations per gesture), then automated training of 5 to 12 h.
- Result
- Presented at the LISI Talk to the group's 150 top executives, then shown at the Spring of AI trade show.
My method
Gather the need
Start from the field, understand the real problem.
First agent
A first working agent, quickly.
Test with users
Confront the agent with real usage.
Iterations
Refine prompts, data and ergonomics.
Production
Deploy in the field, measure, support adoption.
Spreading AI across the group and beyond
Talk at UTBM
Spoke to around a hundred students on AI agents, prompting and concrete use cases.
AI training
Co-led half-day sessions (~30 employees each) on AI, its ethical and confidential use, and getting started with Dust.
Deloitte contest — finance innovation
Pitched an AI agent to a jury of experts: 2nd place.
Trade shows
Industry Show (Paris-Villepinte) and Spring of AI show (robotic-arm demo on the LISI booth).
Stack & tools
AI agents on Dust, the Microsoft 365 foundation, low-code and API connections.
AI agents
DustLow-code & docs
Power Apps
Power AutomateEcosystem
What I took away
Use case first
Technicians don't care about the model: they want to find the right procedure faster.
Adoption matters as much as the tech
A tool is useless without support, teaching and trust.
Pilot, then roll out
A pilot site that proves the value, then generalization.