01From 3,000 to Approximately 20,000 Agents
Around the time of Davos in January 2026, McKinsey CEO Bob Sternfels mentioned a figure in describing his firm's current state. There are roughly 40,000 people inside McKinsey. And roughly 20,000 AI agents. A year and a half ago, the human headcount was nearly identical, but agents numbered only about 3,000. Some outlets reported the figure as 25,000, but what is certain is that this change, from approximately 3,000 to approximately 20,000, happened within 18 months.
When I heard this, what struck me first was the ratio. Nearly 2 to 1. One AI agent for every two humans. That proportion materialized from almost nothing in 18 months. This is not simply a story about automation tools being adopted. This is a story about entities counted as organizational members reaching half the size of the human workforce.
The internal platform Lilli, rolled out firm-wide in July 2023, is the foundation. It integrates vast institutional knowledge with external data, and roughly 72% of staff use it in daily work. The surge in agent numbers reflects the practical depth that has been built on top of this platform.
02The Week-One Answer Became the Hour-One Answer
Kate Smaje, McKinsey's Global Leader for Tech & AI, described the shift in concrete terms. Consultants used to need a week to produce an initial hypothesis. Now it takes an hour. She calls this the move from the "week-one answer" to the "hour-one answer."
This is not only a statement about speed. It carries the implication that the sequence of thinking has changed. Previously, a substantial portion of time was spent gathering, cleaning, and visualizing data before a hypothesis could even be formed. With AI handling that groundwork, humans can start from the question itself. The volume of work has not simply decreased; the nature of the work has changed.
The week-one answer has become the hour-one answer. ── Kate Smaje, McKinsey Global Leader, Tech & AI
03PowerPoint Has Retreated to Being a Delivery Format
Smaje also noted that PowerPoint usage dropped substantially within a matter of months. The nuance matters here: this is not a claim that PowerPoint has disappeared. Decks remain as the final artifact delivered to clients. What has changed is that the actual working layer, where thinking happens, arguments are assembled, and logic is sharpened, has moved onto AI-native tools.
Creating slides was once treated, in some quarters, as the core craft of consulting work. Building a deck while thinking something through can genuinely train analytical discipline. But the inverse is also true: time spent on layout, formatting, and visual consistency has always pulled attention away from the thinking itself. That equilibrium is now beginning to break down.
04One Consultant Replaced Slide Decks with a Website
One example stays with me. A consultant working with a North American cable company built, with AI assistance, what became known as a "client visualization hub": a collection of dozens of HTML files, each containing analysis, charts, tables, and narrative text, assembled into a single shareable web environment. Instead of a proliferating stack of PowerPoint decks, the team delivered a site that could be updated and shared in real time.
This is not a cosmetic change. A PowerPoint file freezes the moment it is sent. The version that arrived in the client's inbox and the version revised the night before the next meeting silently diverge. The cost of managing that drift has never been honestly accounted for. The visualization hub addresses that problem at its root: the consultant and the client are looking at the same information, updated together.
05Bringing It Back to Pharmaceutical Practice
The reason this story resonates so strongly for me is the closeness of the parallel to pharmaceutical and medical work. Medical affairs materials, regulatory submissions, internal approval workflows. All of these run on a large volume of slides and Word documents, and much of the labor involved is data formatting, table construction, and consistency checking.
If the same shift were to arrive in pharmaceutical settings, the work most immediately affected would likely be the MSL pulling data and layering in interpretation, the medical writer summarizing literature and placing it in a label or off-label context, or the version management of safety documents. The move from "someone who produces" to "someone who poses the question" is pointing in the same direction here.
06The Weight of Verification and Record in Medical Contexts
Pharmaceutical and medical practice, however, carries constraints that general consulting does not. Regulatory documents require version control, electronic signatures, and audit trails. If an AI-generated summary ultimately circulates as part of official drug information, the process that produced it must be verifiable. "AI organized it, so it must be correct" is not a statement that survives regulatory scrutiny.
Looked at from the other direction, there is a possibility that with AI handling groundwork, human attention can focus more deeply on the accuracy of the content itself. If time previously consumed by formatting is redirected to scrutinizing scientific evidence, that is a benefit that ultimately reaches patients. The challenge is how to design that transition, and precisely where to draw the line between what is delegated to AI and what is not.
What a pharmaceutical audit asks is: who made this decision? That question does not change regardless of how many AI agents exist in an organization. Human accountability must remain a non-negotiable premise in the medical context.
07What to Do with Recovered Time
The deepest question in McKinsey's transformation, as I see it, is not what AI can do but what humans will do with the time returned to them. If the hour-one answer is now available, what fills the remaining hours? Does one pursue harder questions? Serve more clients? Or simply accept producing the same output in less time and call it progress?
The same three-way choice will arrive in medical practice. When AI delivers a first draft of a materials document, does the recovered time go toward understanding the patient's context more carefully, or only toward accelerating the approval cycle? The same technology carries entirely different meaning depending on how the question is framed.
08Holding onto the Question, Day by Day
McKinsey operates at a scale and in a context far removed from my own workplace. But as a concrete prompt for reconsidering what the time spent making slides has actually been worth, few examples offer this degree of specificity. The figure moving from 3,000 to 20,000 is not a marketing claim. It is a record of a real organization having actually changed the way it works.
Whether one reads this change as a threat or as an occasion for rethinking is a choice that belongs to the reader. I would rather choose the latter. To keep thinking, to keep questioning, not to relinquish that capacity regardless of what tools are available. That, I believe, is what the phrase コラム — each day a good day — means when it is not a platitude but a practice.