The world’s most famous consultancy just revealed its workforce is 40% AI. Here is what the model actually looks like, and what every business leader should take from it.
McKinsey’s CEO announced the firm now has 25,000 AI agents working alongside 40,000 humans. The agents saved 1.5 million hours of work last year. Client-facing roles grew 25%, back-office roles shrank 25%, but total output from the back-office actually went up 10%. McKinsey is moving from fee-for-service consulting to an outcomes-based model. This is not a story about AI replacing jobs: it is a story about what AI-augmented organisations actually look like at scale.
Bob Sternfels, McKinsey’s global managing partner, made headlines in early 2026 when he described the firm as having a “workforce” of 60,000: 40,000 humans and 20,000 AI agents, with that agent number since grown to 25,000. [1] The framing was deliberate. McKinsey is not describing AI as a tool or software. It is describing these agents as employees, with tasks, outputs, and roles within project teams.
That is a meaningful distinction. “AI tool” implies something you open and close. “AI workforce member” implies something that holds responsibility for a defined scope of work and is held accountable for output quality.
McKinsey’s AI agents are built primarily around its internal AI platform, called Lilli. Lilli is McKinsey’s proprietary knowledge retrieval and generation system, trained on decades of the firm’s research, frameworks, and client deliverables. The agents built on Lilli can retrieve relevant case studies, synthesize research from multiple sources, draft structured analyses, and surface precedents from previous projects. [2]
This matters for understanding what McKinsey actually announced: it is not “we installed ChatGPT.” It is “we built a custom AI infrastructure and taught it everything we know, then gave it tasks.”
The word “agent” is used loosely in AI conversations, so let me be specific about what McKinsey’s agents are actually doing day to day.
An AI agent is different from a standard AI chatbot in one key way: it can take actions autonomously over time. You give an agent a goal. It plans a series of steps, executes them, evaluates the results, and continues toward the goal without needing you to prompt it at each step. It can use tools (search, databases, document editors) and make decisions.
At McKinsey, the agents are primarily handling: research synthesis (pulling relevant studies and precedents on a client problem), competitive analysis (scanning industry data and structuring comparative findings), document generation (producing first drafts of slide frameworks and summary memos), and meeting preparation (generating briefing documents with key questions and context). [3]
The firm says these agents saved 1.5 million hours on search and synthesis work alone in one year. [4] That is the equivalent of roughly 750 full-time employees working a full year, focused entirely on finding and organizing information. And notably, the total output of the teams involved actually went up 10% even as headcount in those roles went down.
What this means: The agents are not replacing the thinking. They are replacing the finding, organizing, and first-draft stages. The consultants are still making the judgments. The difference is they are making those judgments with the background work already done.
The most concrete data point from McKinsey’s announcement is the dual 25% figure: client-facing roles grew by 25%, while back-office and analytical support roles shrank by 25%. Both numbers matter. [4]
The roles that grew are the ones that require direct client interaction, strategic judgment, and relationship management. The roles that shrank are the ones primarily involving information gathering, structured analysis, and document production. Research analysts, data processors, and certain categories of junior analytical work saw the biggest reductions.
But here is the number that gets underreported: the teams in those shrinking categories still produced 10% more output. Same total team, less headcount, more deliverables. That is the actual proof point for the AI agent model.
For junior consultants, the shift is significant. Their roles are moving from doing analytical work to supervising it. Instead of building the slide with the competitive landscape, they are now reviewing the slide the agent built, correcting errors, adding strategic framing, and integrating it into the broader narrative. That is a different skill set, and it requires understanding what the agent is likely to get wrong. [5]
McKinsey is responding to this by changing what it hires for. It is actively recruiting more liberal arts graduates now, prioritising creativity, judgment, and communication over analytical horsepower. The firm has also added an AI collaboration test to its graduate interviews, requiring candidates to work with Lilli during the assessment process. [6]
Understanding the workforce announcement requires understanding why McKinsey is making it. There is a strategic business reason this is being shared publicly, and it has to do with how consulting firms get paid.
Traditional consulting fees are based on hours of senior consultant time. You pay for a team of six consultants for twelve weeks. The value is in their expertise and the time they invest. But if AI agents can do in two hours what used to take three weeks, the hours model starts to break down. Clients will start asking: why am I paying for fifty analyst hours when you could have a machine do it in an afternoon?
McKinsey is getting ahead of this by explicitly shifting to an outcomes-based model. Sternfels has stated publicly that the firm is moving toward arrangements where fees are tied to the impact delivered, not the hours logged. McKinsey identifies a joint business case with a client, defines measurable outcomes, and underwrites the result. [1]
This is a significant structural change. It means McKinsey can justify higher fees for better outcomes, not larger teams. And it means the AI agents are not eating into revenue: they are enabling McKinsey to deliver outcomes faster, which makes the outcomes-based model work in their favour.
If your organisation uses consulting services or professional services of any kind, this is worth watching. The firms that figure out the outcomes model first will have a structural advantage over those still charging by the hour.
The McKinsey model makes three skill categories clear winners in the AI-augmented workplace.
The ability to read AI-generated analysis and know when it is right, when it is plausible but wrong, and when it is confidently fabricated. This is not a technical skill. It is a domain expertise skill. The better you know your field, the better you can evaluate what an AI produces in it. This is one reason McKinsey’s senior partners are irreplaceable while junior analysts are not: judgment cannot be automated.
The ability to break a complex task into components, identify which parts AI can handle autonomously, and sequence them so the human contributions happen where they add the most value. This is project management applied to human-AI collaboration. It is learnable, and most professionals have never been taught to think about their work this way.
AI can generate analysis. What it cannot do is make the strategic case for acting on it in a way that moves a specific room of specific people. Writing, presenting, and facilitating decisions based on AI-generated insights is a growth area. The professionals who can translate AI output into compelling, actionable recommendations have a genuine edge.
None of these are exotic capabilities. But they are skills that need deliberate development. Most organisations are not training for them yet. The ones that start now will have a head start.
You do not need Lilli or a custom AI infrastructure to start testing the agent model. You can build a basic version of what McKinsey is doing using tools available today.
List the tasks in your team that primarily involve gathering information, organizing it, and producing a first draft. Competitor monitoring. Meeting preparation. Weekly report assembly. These are your candidate tasks for agent-style automation.
Take one task and write a structured prompt that produces a reliable first draft. Use ChatGPT or Claude. Test it with real data. Refine until it consistently produces something your team would normally spend two hours building in thirty minutes.
Search for news about [competitor name] from the past 7 days. Identify: (1) any product announcements, (2) any pricing changes, (3) any executive statements or strategic signals, (4) any customer feedback or reviews. Format as a structured briefing with one paragraph per category. Flag any items that require immediate attention.
The key shift is moving your team from doing the first draft to reviewing it. This requires a culture change as much as a process change. People need to trust that reviewing AI output is real work, not cutting corners. Frame it clearly: their judgment on the output is where the value lives.
McKinsey tracks hours saved as a metric. You should too. Even a rough estimate of time reclaimed per week gives you data to show leadership and builds the business case for investing more seriously in the approach.
Start small. One task, one workflow, two weeks of testing. The goal is not to build Lilli. The goal is to understand what the agent model feels like at a human scale, so you can make smart decisions about where to go next. You can also read our guide to ChatGPT Agent Mode for a more detailed walkthrough of how to set up multi-step tasks in ChatGPT itself.
McKinsey’s announcement is framed as a story about efficiency, transformation, and the future of knowledge work. And all of that is true. But there is a question underneath it that most business leaders are not asking out loud: if the agents can do the analytical work and output goes up 10% with 25% less headcount, was the headcount always unnecessary?
That is an uncomfortable question. The answer is probably: no, but the work was structured in a way that required more human time than the underlying value creation demanded. The 32% of companies expecting workforce reductions of 3% or more in the next year are not all being malicious. [4] They are recognizing that some of what people were paid to do was, in retrospect, something that could have been more efficient.
The leadership challenge is not just “how do we adopt AI.” It is: how do we restructure work so that the humans in our organization are doing the things that genuinely require humans, and not spending their best hours on tasks that a well-prompted AI could complete at 80% quality in five minutes?
McKinsey’s 25,000 agents are a provocation. The firms that treat it as a headline will fall behind the ones that treat it as a blueprint.
Workforce AI stories are often covered either as pure tech news or as anxiety-inducing “robots are coming” narratives. This article is neither. It is a practical look at what one of the world’s largest knowledge-work organisations is actually doing, so you can decide what it means for yours.
Sources
Weekly tips. Real examples. Practical help for busy professionals.
We care about your data, check out our privacy policy.
