Bruno Maillot, "Après les LLM, l’IA agentique"

Bruno Maillot, "Après les LLM, l’IA agentique"

Title history
  • Bruno Maillot, "Après les LLM, l’IA agentique" — observed on 20261110-0712 UTC (YouTube)

by Bruno Maillot · View the channel on PSI

Applied Sciences & Engineering AI — News & Applications UYComputer scienceUYQArtificial intelligence (AI)
🎙 Bruno Maillot 👥 1K 📅 November 6, 2026 ⏱ 31 min 👁 159 📄 expert opinion 🧭 2026-11-10
Available in: English (current) Français

Keywords

agentic AIAI agentslarge language modelsAI governanceuse cases

Summary

In this expert interview, Bruno Maillot, partner at Sopra Steria NEXT, discusses the emergence of agentic AI after the initial wave of LLMs. He clarifies the distinction between a virtual assistant and an autonomous agent: an agent receives an objective and tools, chains tasks autonomously, and can retry if needed, whereas an assistant only answers questions. He highlights the evolution that made agents possible: LLMs providing the ‘brain’, then the ability to interact with tools (the ’eyes’ and ‘USB’ via MCP), and finally agent-to-agent communication. Maillot emphasizes that choosing the right use cases is critical, focusing on processes with high frequency, accessible data, and the ability to measure errors. He presents several examples: Sosh’s Charlie, a speech-to-speech customer service agent that can act on systems; Canal Plus’s gradual upgrade from decision trees to conversational agents with RAG and action capabilities; and a counter-example, Klarna, which over-automated and faced quality erosion after layoffs. He also discusses the Air Canada chatbot case where the company was held liable for an aggressive online discount, underscoring that whoever controls the agent bears responsibility. He stresses the importance of cybersecurity, human-in-the-loop oversight, and continuous evaluation before production. Finally, he describes the value chain: models are commoditized, while value resides in proprietary data, governance, and final decision-making, indicating job opportunities for data professionals in these higher layers.

221 words

Critical Evaluation

Value of the Information & Strength of the Argument

The talk offers valuable insights into the practical deployment of agentic AI, backed by real-world examples and cautionary tales. Maillot argues convincingly that success hinges on measuring errors and maintaining human oversight, as illustrated by the Air Canada liability case and Klarna’s quality issues. His pragmatic approach is grounded in his consulting experience and directly addresses implementation challenges. However, some claims are presented broadly without detailed evidence or statistical support, and the tone remains rather optimistic about the technology’s potential.

Scientific Rigor, Source Quality, Title Accuracy

The video is not a formal scientific presentation but an expert practitioner’s talk. The speaker cites real companies and incidents (Air Canada, Klarna, Sosh, Canal Plus) without providing academic references or links to primary sources. The title is accurate and matches the content well. The description includes only institutional links (ENSAI’s website and social media) rather than sources for the cited examples, reducing traceability. Nevertheless, the reasoning is coherent and the examples are relevant.

169 words

Title / Content Match

The title accurately reflects the content, focusing on the evolution from LLMs towards agentic AI with definitions, practical use cases, and challenges.

Quality & Reliability

7/10

The speaker is an industry expert with substantial practical experience in AI deployment and provides concrete examples and lessons learned. However, the presentation is opinion-based and lacks citations to peer-reviewed research or formal sources.

Key Moments

Markers derived by PSI from the transcript: the creator did not define chapters.

Cited Sources

  • ENSAI official website — Institutional site of the school hosting the event.
  • ENSAI LinkedIn — Institutional LinkedIn page provided in the description.
  • ENSAI Instagram — Institutional Instagram page given in the description.
  • ENSAI Facebook — Institutional Facebook page from the description.

Concurring Sources

  • Large language model — Background reference on LLMs, the foundational technology discussed.
  • Intelligent agent — General concept of intelligent agents, consistent with the definition provided in the talk.

Contribution & Novelties

The video provides a practitioner’s view on how to select and deploy agentic AI in enterprises, emphasizing the necessity of evaluation and human control. It offers a clear framework for judging whether a process is suitable for an agent, using frequency, data access, and error measurement. The speaker’s distinction between the ‘wow’ effect and long-term reliability is a practical contribution often overlooked in academic debates.

Pour aller plus loin :

  • Model Context Protocol (MCP) — Official protocol for connecting AI to data sources and tools; directly relevant to the ‘USB’ mentioned in the talk.
  • Retrieval-Augmented Generation — Technique combining retrieval with generation, used in the Canal Plus example and key for grounding agents.
  • AI governance — Discusses frameworks for responsible AI, aligning with the talk’s emphasis on oversight and accountability.

130 words

Radar Profile

The radar profile shows a high quantity of information with moderate technical depth and a decent reliability score for an expert opinion. This suggests a broad, accessible overview rather than a deep technical or highly rigorous analysis.

Reliability 7/10