Lots of people are asking what I do at IDC, so let me tell you the plan. The research program I lead focuses on two opposite ends of a spectrum. On one side: the newest, most innovative ideas in AI. I won’t look only at the technology layer, but also at how that layer affects business and organizational models, professions, welfare strategies, the economy, and more. Think of this as an incubator program for bold new concepts. We want to capture what’s next in AI before it becomes obvious to everyone. If and when these ideas become mature enough, we’ll create new vertical research programs to fully explore how the market absorbs the technologies and the models behind them. On the other side of the spectrum: how large end-user organizations adopt AI, starting with CIOs, CISOs, CEOs, and boards. I will look at how they make decisions, how they perceive the AI opportunity and the market landscape, who is doing something differently, and how those decisions make them successful. I will also look at how they rethink the role of their organizations to address the challenges AI creates. Think of this as an attempt to capture the playbook of AI winners, rather than focusing only on the challenges posed by new, disruptive technologies. To move between these two extremes, my team and I will use the invaluable data IDC allows us to collect at a global scale. We can offer a macro view of who is doing what, where, and at what scale. And that view is not limited to North America. My team is in Europe, the Middle East, India, Singapore, and South Korea. What I described is a very ambitious, monumental mission. Obviously, I can’t do this alone. Beyond my team, I can count on an amazing group of IDC analysts all around the world. I’ll work with all of them to understand how the world is changing and what our clients must do to win in the age of AI. Isn’t this the best job in the world?

.@thsottiaux Passkey on X prevents login in Codex browser. Workarounds (other than using Codex extension in a Chrome instance)?

Recursive self-improvement going from being a ridicule-worthy fringe sci-fi concept to a completely normalized part of the discourse which is “obviously the plan” is one of the more dramatic Overton window shifts I’ve experienced There’s something disorienting about it, like if the sky suddenly turned red, and everyone acted like it had been that way all along

Thinking about how GenAI is affecting our attention and decision-making capabilities, and wondering whether there’s a point of no return. Printing press → Internet → Mobile phones At every step, the increase in information availability has made us more distracted: less time to absorb all available information, and shallower in how we process it, with less time to review each piece carefully. Then, GenAI arrives. It is not just driving the cost of producing information toward zero. It is also driving the cost of producing complex information toward zero. So, is there a point of no return? A point where information becomes so abundant, so overwhelming, and so complex that our decision-making capabilities become catastrophically impaired? Why? Why not?

Agentic Sunday, as usual. Three years ago, quite a few people were certain that developing prompting skills wouldn’t be necessary. Not only did the need for prompting skills not go away, but today, it has become critical to make the jump from AI assistants to AI agents. How you articulate your instructions inside agents[.]md, skills[.]md, and every other file you use to define guardrails, state management, etc., as part of your process orchestration framework, makes a huge difference. And even when you stay focused on AI assistants, your prompting skills are the difference between AI slop and good generated content. In completely unrelated news, we have a looooong way to go before Codex and Cowork generate really good presentations.

Controlling AI agents remotely today looks clunky and inaccessible to non-technical users. You have to type and review outputs on a tiny phone screen, while your computer has to stay on, at home, all the time. But these are just the early stages. We already have all the technology we need to audio- or video-call our agents, speak to them in any language, and hear back what we need in a perfectly human-sounding voice. And what about the need to keep your computer on at home at all times? If I tried to convince you to pay an annual subscription for access to Windows or macOS instead of owning them, I’d probably fail. But what if I told you that a slightly more expensive tier of your AI subscription, the one you already pay for, came with a synthetic assistant you can speak to, and that this assistant also had access to all your files, calendar, email, and more? Would that sound more tolerable?

"How do you forecast token consumption in your organization?" "No"

Today, current UIs significantly constrain the potential of AI agents. Too many interactive processes are too slow simply because a button, a pane, or a feature is not where it should be. Eventually, we’ll get to a point where we’ll be able to say: “Based on our previous conversations, morph your UI in a way that maximizes the speed of my workflows.” But then, our inability to express our intent as fast as we formulate it will become the new bottleneck. Tomorrow, future UIs will significantly constrain the potential of AI agents.

One of the many paradoxes of AI: business and IT leaders I meet around the world constantly talk about the need to trust AI. And yet, unlocking the full potential of agentic AI will require the biggest leap of faith we’ve ever made. Human attention is such an inescapable bottleneck.

In Paris, today and tomorrow, for the Capgemini CxO Forum. Great conversations so far, fully aligned with our position on AI adoption and where the world might go next.

My biggest hope is that, at some point, Codex will have a peer review UI where the user can comments on specific sections of an answer and modify some of its text (just like in Word). When the review is done, Codex would take comments and edits into account and rewrite whatever it has to rewrite to incorporate the feedback. Too much to ask @thsottiaux ?

I spent yesterday in NYC speaking at our IDC CIO Summit. Lots of @lmstudio users. @yagilb should be pleased. I had no doubts.

The general understanding of how AI agents work, and how they differ from AI assistants, remains very nebulous for many people. From my vantage point, a lot of workers have yet to discover what Skills are and how they can influence agents.

AI models’ decision-making will be one of the most important areas to unlock business value in the future. I’ll try to explain why with a personal story. Joining IDC has not changed the fact that I am an AI practitioner at heart. Over the past few weeks, I’ve been working to turn a key business process into an advanced AI framework, chock-full of goals, skills, and guardrails. The scenario I’m focused on involves turning the main session of the agentic AI platform I’m using into an orchestrator. That orchestrator must be able to spawn 10 subagents, each specialized in highly detailed jobs, and manage them reliably over multi-hour working sessions. As expected, refining the framework has taken many iterations. Along the way, minor limitations in the UI of the AI platform I’m using became very evident (a topic for a future briefing with the vendor!) But the hardest part remains guaranteeing that the orchestrator successfully manages its subagents with minimal to zero supervision from my side. To improve my chances, I had to define a watchdog loop that constantly checks on and nudges the subagents. All in plain English. No specialized programming language involved. For me, it’s a very fun project. But it confirms, once again, the enormous challenges non-technical users face when designing and managing business processes executed by multiple agents. The only possible way to unleash the true potential of agentic AI for those users seems to be to completely abandon the task-oriented approach to work we are used to, and embrace a purely outcome-based approach. The challenge is that it’s not very easy for humans to explain the outcome they have in mind. Even through an iterative process, it might take hundreds of interactions to see the ideas in our minds fully realized. Very few people will have the patience or the time to do that. Most will probably settle for whatever the AI recommends after a few interactions. The dear old “good enough.” And if that’s true, how AI models make decisions about what direction to take, what tool to use, and what trade-offs to accept becomes one of the most important areas to unlock business value. “Good enough” better be really, really good. (As my agents would say: "You are right to push back on this.")

On human supervision: we will auto-approve those AI agent requests. The imperative is to conserve energy, and those requests are already becoming too complex to fully understand.

The fastest way to realize you were asking the wrong questions? Watch a swarm of AI agents deliver in minutes the answers you’d have taken a week to find.

During a recent trip to participate in a CIO council, I was asked which blind spots CIOs are least prepared for over the next 1–2 years. The steep learning curve of designing AI agents was one of the three things I mentioned. It’s only when you move from interacting with an AI assistant to programming a swarm of AI agents that you realize how incredibly sophisticated and nuanced your prompts can and must be. Natural language makes the interface feel simple, but shouldn’t deceive us. The jump is huge and, as of today, still too advanced for many LOB workers.

If you abstract skillfully, the management/governance stack of every major new technology wave looks the same. Virtualization > Cloud > GenAI. Is it because the roles of each box in the diagram are universally important? Or is it because we cannot leave behind a certain mental model of how things work?

The helicopter pilot skill was an .md file. Google’s release of DESIGN[.]md makes the analogy inevitable. For a long time, I’ve suspected that the workflow is where the ultimate competitive advantage lies. I mean “workflow” in the broadest sense of the word, and well beyond the boundaries of Information Technology. Now, the workflow is codified in an .md file. Does this mean we are moving toward a workforce powered by a dynamic, composable matrix of skills? Does it mean that we, as humans, gain interchangeable superpowers as we lean on agents to multiply our impact? Does it mean that, going forward, there will be business value in customizing the standard skills that are beginning to emerge today? I don’t have the answers. Not yet. What do you think? youtube.com/watch?v=SoAk7zBT…

Even after joining @IDC, I spend my Sundays designing workflows for my agents. It’s the slowest moment of my week. It feels both contemplative and mystical: sitting in front of the screen, trying to imagine what an agent will do in one situation or another, involving it in the design decisions, watching it execute the plan, and then ironing out the details together. Somewhere in the middle of all that, I detach from the moment just enough to realize that what I’m doing is completely surreal and, somehow, it actually, really works.