The story of how I had no time at the International AI Summit, so I had AI 'listen for me'

The dilemma of 'attending' a summit but 'not being able to attend' sessions

Last month, I participated in the India AI Impact Summit held in New Delhi, India.

It is one of the largest international conferences in the AI field, where government officials, major tech companies, and researchers from around the world gather. For several days, sessions are held in parallel across multiple venues from morning until night.

However, I must confess something honestly here.

I was unable to attend most of the sessions.

Actually, this time, apart from the main summit program, I was on the side of planning and managing our own side event. Coordinating with speakers, setting up the venue, and managing the day's proceedings—when you are busy organizing an event, you physically cannot attend other sessions happening at the same time. Balancing the success of the event you are involved in with learning from other sessions is the eternal dilemma of international conferences.

What I particularly regretted was missing the series of sessions on 'Agentic AI,' which is related to the project I am in charge of at work.

Why I am interested in 'Agentic AI'

Let me talk a little bit about my work.

I usually work on supporting the creation of international rules surrounding AI. Specifically, I belong to the Tokyo Center of an international framework called GPAI (Global Partnership on AI). GPAI was integrated with the OECD in 2024, and it is currently promoting policy recommendations and practical projects regarding AI as part of the OECD. And right now, one of the hottest themes in this field is Agentic AI—that is, AI systems that 'act autonomously' upon receiving human instructions.

Until now, AI has been like a 'consultant' that answers when you ask it questions. But Agentic AI is different. If you ask it, 'Book a flight for my business trip next week, reserve a hotel, and share the schedule with my boss,' it will actually do all of that. While convenient, the risks of 'AI making decisions and acting on its own' and the development of rules for it have become an urgent issue worldwide.

Looking at the program for the India AI Impact Summit, there were 14 sessions related to this Agentic AI. That is over 14 hours in total. Even if I could have attended all of them, it would take a considerable amount of time to organize the content and share it with others.

However, there was one saving grace. At this summit, most sessions were live-streamed on YouTube and made available as archived videos on the summit's official website. In other words, even if I wasn't at the venue, I could check the content via video later. The problem was how to secure the time to watch it all.

So, I decided to 'get help from AI' this time as well.

Automatically identifying 'related sessions' with Claude Code

The first thing I worked on was the task of extracting only the Agentic AI-related sessions from the entire summit program.

The India AI Impact Summit program lists hundreds of sessions. Searching for those related to Agentic AI from among them manually is a hassle.

So, I had Claude Code (Anthropic's AI coding tool), which I introduced in my previous article, read the summit program page and instructed it to 'extract sessions that include terms like "agentic," "agent," or "multi-agent" in the title or summary.' (By the way, Claude Code, which autonomously writes, executes, and repeats code corrections, is a typical example of the aforementioned Agentic AI.)

As a result, 14 sessions were listed, and each was organized into a CSV file with its date, time, venue, and even the YouTube live stream URL.

Generating a 'summary for 14 sessions' all at once with NotebookLM

The next step is the core of this experiment.

Out of 14 sessions, 13 had YouTube videos available. However, watching all 13 videos would take over 13 hours.

This is where Google's NotebookLM comes in. NotebookLM is an AI tool that structures and summarizes content when you feed it videos or documents. It doesn't just summarize; it also creates audio overviews, mind maps, infographics, slide decks, flashcards, and even quizzes.

I asked Claude Code to 'create an application that sends each YouTube URL listed in a CSV to NotebookLM and automatically retrieves the summary for each.' After some trial and error, I finally completed a system where NotebookLM generates a summary for each session just by passing the URL, and the results are compiled into a spreadsheet.

The processing time averaged about 21 seconds per session. I was able to build a system that delivers summaries for 13 sessions in less than 5 minutes in total.

13 hours of video summarized in 5 minutes.

Of course, I could copy and paste each video URL into NotebookLM one by one, but that would be a significant amount of work. Since I expect to be doing this kind of task a lot in the future, there is great value in building a system like this.

Of course, a summary is not a substitute for the video. Nuances and the atmosphere of the venue are lost, and there are limits to the accuracy of the summaries. However, for the purpose of 'getting an overview of what was discussed first,' the results were more than sufficient.

List: Agentic AI discussed at the summit — An overview of 14 sessions

Below is a list of the 14 sessions organized with the help of AI. I had Claude read the results above, and it even kindly categorized them by theme (lol).

5 emerging points of discussion: The current state of Agentic AI as shown by the summit

Reading across the summaries of the 13 sessions, the outline of the discussion begins to emerge. I used audio overviews and mind maps to grasp the details of the sessions I thought were particularly interesting, and if I felt they were truly meaningful, I watched the actual videos at double speed. From here, I will attempt my own summary.

1. The limits of the 'human-in-the-loop' model

'Human-in-the-Loop' was mentioned in almost every session. However, what is interesting is that this concept itself is being forced to evolve.

In one session, there was a discussion, using an analogy from the aviation industry and drone regulations, that it is unrealistic for humans to approve every single action of an AI agent, just as a pilot does not approve every single movement of the control stick. What was proposed instead was 'Human-in-Command'—the idea that humans should focus on overall direction and intervention in exceptional cases rather than individual decisions.

This resonates with my own experience. In the collaboration with Claude Code that I introduced in my previous article, I would reply to the policies proposed by the AI with 'Yeah, let's do that,' but it took a lot of time. There is no way I could approve everything during complex decision-making. Determining what humans should decide and what AI should decide may become important in the future.

2. Governance as 'part of the design' rather than an 'afterthought'

'Safety by Design,' 'Assurance by Design,' 'Governance as First Principle'—the expressions varied, but what was shared in almost every session was the recognition that safety and governance should not be added after a product is completed, but should be incorporated from the design stage.

One speaker systematized this as a 'governance stack' consisting of five layers (build-time -> deploy-time -> run-time -> remediation -> accountability). The idea of covering the entire lifecycle of an AI agent is still in its early stages, but it could become the foundation for future international discussions.

3. 'One-size-fits-all standards' will not work—Voices from the Global South

A unique feature of the summit held in India was the clear objection to applying standards originating from the West directly to the rest of the world, which was raised in multiple sessions.

Multilingual environments (hundreds of languages in India alone), infrastructure constraints, and the gap between urban and rural areas—the argument is that a 'one-size-fits-all' approach that ignores these contexts does not work. Concrete examples were shared where AI services designed for cities failed completely in rural areas.

This point is an unavoidable challenge in advancing the international harmonization of AI standards, and it is directly linked to my own work.

4. 'Know Your Agent'—The foundation of trust in the agent era

In the 'Agentic Commerce' session, where AI agents conduct commercial transactions on behalf of humans, it was reported that conversion rates for companies that introduced them jumped 7 to 9 times, and the concept of 'Know Your Agent' was proposed.

This concept, modeled after 'KYC (Know Your Customer)' in the financial sector, consists of three pillars: AI agent identity verification (cryptographic ID), permission settings, and tracking of behavioral history. In an era where agents act as human proxies in economic activities, a mechanism to verify 'is this agent trustworthy?' may eventually become the core of international rules.

5. The OECD as a 'venue' for international cooperation

In multiple sessions, especially in roundtables attended by US and Indian government officials, the OECD was repeatedly mentioned as a core framework for international cooperation on AI governance. Along with the G7 Hiroshima AI Process and the ITU's 'AI for Good,' it is expected to play a role as a venue for formulating practical-level standards.

While each country and region creates its own standards, they must be aligned with international benchmarks to ensure interoperability—this 'glocal' approach seems to be the basic direction for future Agentic AI governance. Listening to the discussions at the summit, I realized once again that the GPAI/OECD, to which I belong, is positioned exactly at that intersection.

Conclusion: An era of 'having someone listen' when you 'couldn't listen'

Looking back, this experience contains a bit of irony.

I had 'AI' listen to and summarize sessions at a summit discussing 'international rules for AI' on my behalf. And I am trying to apply the results to my work supporting the creation of 'international rules for AI'.

Understanding discussions about AI with the help of AI—if I had heard this story a few years ago, I might have thought it was some kind of joke.

Of course, there are limits. Video summaries do not convey the facial expressions of the speakers, the reactions of the audience, or the subtle temperature of the Q&A sessions. The accuracy of the summary is not perfect, and there is always a possibility that important nuances are lost. Also, in these types of international conferences, it is often important to know who said what and from what background, but at least with just a summary, that cannot be fully captured.

Even so, being able to grasp the overall picture of 13 hours of discussion through this work definitely moved my work forward.

There is often a debate that 'AI will take away human jobs,' but my experience is the opposite. AI significantly increases the amount and speed of information that a single person can absorb. It is not a threat, but a feeling of being able to reach things that were previously out of reach.

I would like to attend a few more sessions at the next summit—but even if I can't, I already know that I have a reliable 'proxy attendee'.

Tools used in this article: Claude Code (Anthropic), NotebookLM (Google), Claude Cowork (Anthropic). The automation tool created at this time is published on GitHub as 'notebooklm-auto'. It may take some time to set up, but I think you can set it up to run on your own PC. You can create NotebookLM notebooks, generate summaries, and get shared links all at once just by providing a list of YouTube URLs. If you are interested, please give it a try. https://github.com/s00048ri/notebooklm-auto

The session videos for the India AI Impact Summit are available on YouTube. Please watch the sessions you are interested in directly.