.svg){kind=small}
.svg)
.svg)
Sovereign AI Transformation for National Media Heritage Modernizing the Institut National de l'Audiovisuel (INA) by converting "Shadow AI" risks into secure, high-value operational assets.
The Result: Eradicated unsecured AI usage and achieved a massive 80% reduction in indexing time (from 4 hours to 40 minutes). This flipped the internal culture from fear of replacement to active "AI Copilot" collaboration across the institution.
How do you modernize an institution guarding 80 years of media history without compromising data security or alienating its experts?The Institut National de l'Audiovisuel (INA), France’s national media archives repository, faced a critical dual challenge: rampant "Shadow AI" usage risking copyright infringement, and deep internal anxiety among archivists fearing they would be replaced by algorithms.My mission was to turn this defensive posture into a competitive operational advantage. Moving away from theoretical training, I led a pragmatic transformation program across 15 business units. We replaced unsecured tools with a sovereign "AI Toolbox" built on Azure, deploying specialized AI Agents co-constructed with the teams. These were not generic chatbots, but vertical tools designed for specific heritage workflows: automatic synthesis of radio archives, transformation of editorial content into social media assets, and B2B newsletter drafting.The results are measurable and transformative :Massive ROI : Editorial indexing tasks dropped from 4 hours to just 40 minutes, an 80% time reduction reinvested in high-value analysis.Security : We eradicated Shadow AI risks by providing a safer, better internal alternative.Cultural shift : The narrative flipped from "AI will replace us" to "AI is our copilot," with teams now actively requesting new agents.This project proves that even the most heritage-centric institutions can lead the AI revolution when technology is perfectly aligned with human expertise.
The Institut National de l'Audiovisuel faces a unique paradox: it is the guardian of France’s audiovisual memory (24 million hours of archives) but must remain at the forefront of digital innovation.Before this project, the institution was trapped in a critical bottleneck :The "Shadow AI" Risk : Faced with increasing workloads, employees across editorial and documentation teams had started using public tools (like ChatGPT) without supervision. This created severe risks regarding data sovereignty and copyright infringement, a non-negotiable red line for a public heritage institution.Cultural resistance & fear : There was a deep-seated anxiety among experts (archivists, journalists) that Generative AI was a threat to their jobs. The prevailing sentiment was defensive: "AI will replace our expertise."Operational Iiefficiency : Highly qualified staff were drowning in low-value, repetitive tasks. For instance, writing a standard documentary index or summarizing a radio show could take up to 4 hours of manual labor, preventing experts from focusing on high-value analysis and heritage preservation.The client needed to urgently move from an uncontrolled, fearful environment to a secure, mastered, and productivity-driven AI ecosystem. They needed to prove that AI could be a compliant "copilot" rather than a replacement.
The technical infrastructure was built on a sovereign Azure OpenAI instance, deployed by the client’s IT department to guarantee data security and GDPR compliance.My role was to engineer the "Intelligence Layer" and the User Experience to turn this raw technical capacity into operational business tools. I acted as the bridge between the technical potential and the business reality.1. Architecture of 75+ specialized agents : Instead of a single generic chatbot, I architected a granular ecosystem of 75+ vertical AI assistants. I configured each agent to solve a specific business friction (e.g., "Radio archive summarizer" vs. "Social media caption generator").2. Advanced system configuration & Prompt engineering: For each assistant, I engineered the "System Instructions" to ensure reliability and quality. I utilized techniques such as :- Role Prompting : Defining strict personas for each agent (e.g., "Act as a senior archivist").- Few-Shot Prompting : Feeding the agents with specific INA examples (past summaries, editorial guidelines) to align the output with the institution's tone of voice.- Formatting constraints : Forcing structured outputs (bullet points, JSON, specific character counts) to fit existing workflows.3. UX/UI & Workflow optimization: I redesigned the navigation logic of the Toolbox. I organized the assistants by "Business Vertical" rather than by "Model Capability," significantly lowering the barrier to entry for non-technical staff. I ensured the interface guided the user from the problem to the right AI solution intuitively.
The project delivered immediate, measurable ROI on both operational efficiency and organizational transformation :1. Hard ROI : Massive Productivity Gains on Core Tasks The most spectacular result concerns the heritage documentation teams.Editorial indexing : Writing complex archival notices previously took 2 to 4 hours per unit. With the specialized "Archivist Agent," this dropped to 20–40 minutes (including mandatory human verification). This represents an 80%+ reduction in processing time, allowing experts to tackle the archive backlog.Content repurposing : For the Communication department, transforming long-form articles into social media assets (carousels/threads) now takes 50–70% less time, increasing content output volume without increasing headcount.2. Soft ROI : Adoption & Cultural shiftScale : We successfully onboarded 15 distinct business units (Editorial, Tech, HR, Comms) and trained a network of 50+ active power users.Deployment : We moved from zero to 75+ specialized assistants deployed in production within 6 months.Security : We achieved a near-total reduction of "Shadow AI" risk in the target groups, as employees switched from unauthorized public tools to the secure internal Toolbox.3. Long-term strategic value I delivered a reusable "AI Evaluation Matrix" (assessing reliability vs. token cost vs. human time saved). This gives the INA a permanent framework to autonomously qualify and deploy future AI use cases, ensuring the transformation outlasts my mission.
This project stands out because it succeeded where most AI initiatives fail : bridging the gap between radical technological innovation and strict institutional constraints.1. Proof of scale in a Risk-Averse environment : Implementing GenAI in a startup is easy. Implementing it in a National Heritage Institute, obsessed with copyright, sovereignty, and job preservation, is a massive challenge. I proved that with the right "Pragmatic" methodology, even the most protective organizations can embrace AI without compromising their values.2. ROI beyond the hype : I am not presenting a theoretical roadmap or a "fun" marketing POC. I am presenting a production tool that slashed core business processing times by 80%. Saving 3+ hours per document for hundreds of documents is a tangible, massive financial and operational win for a public organization.3. The human victory : The greatest excellence of this project isn't technical; it's cultural. I took teams that were genuinely afraid of being replaced by ChatGPT and turned them into power users who now design their own prompts.This project is the perfect benchmark for Responsible & Efficient AI : secure, verticalized, and strictly serving human expertise rather than replacing it.
.svg){kind=icon}