McKinsey Now Has 25,000 AI Agents. Here’s What That Means for the Rest of Us.

Eighteen months ago, McKinsey had 3,000 AI agents and 40,000 human consultants. The humans outnumbered the agents more than 13 to 1. Today, McKinsey CEO Bob Sternfels says the firm has 25,000 agents, and the stated goal is to reach parity with the human headcount before 2026 is out. [1]

That’s not incremental. That’s a fundamental reorganization of how one of the world’s most influential professional services firms does its work. And if you think this is just a McKinsey story, you’re looking at it wrong.

The numbers that should get your attention

Let’s be concrete about what “25,000 AI agents” actually means. These aren’t chatbots sitting in a corner waiting to be asked a question. McKinsey’s agents, powered largely by an internal platform called Lilli, handle research, document synthesis, comparative analysis, and first drafts of client reports. [2,3] They run continuously. They pull from proprietary databases. They produce structured outputs that human consultants then review, refine, and present.

The growth rate is what’s striking. McKinsey went from 3,000 agents 18 months ago to 20,000 agents by January 2026, then to 25,000 by the time of Sternfels’ most recent Harvard Business Review interview. [1,4] That’s more than 500% growth in a year and a half, at one of the most cautious, reputation-sensitive firms on the planet.

What Lilli actually does day to day

Lilli is McKinsey’s internal AI platform. Think of it as what happens when you give every consultant a brilliant, always-available research assistant that has read every client report, case study, and industry analysis the firm has ever produced, plus live access to external research databases.

In practice, Lilli handles the front-end grunt work of knowledge consulting: reading and synthesizing large volumes of documents, pulling relevant precedents from past client work, running comparative analyses across industries, and producing structured first drafts. [2] The human consultant’s role then shifts toward judgment: what questions to ask, which findings matter, how to frame a recommendation for a specific client’s context.

This is not magic. It’s a very sophisticated version of what you can do right now with ChatGPT’s Agent Mode or Claude’s Projects feature. The difference is scale, proprietary data, and institutional tuning. The underlying principle is identical: use AI for the information-intensive first pass, use humans for the contextual, relational, judgment-intensive layers.

The new AI job interview stage

This is the part that caught a lot of people off guard. McKinsey has added a stage to its graduate recruitment process that specifically tests candidates’ ability to work with Lilli. Candidates are given exercises that resemble real client scenarios and asked to complete them using the AI tool. [3]

What McKinsey says it’s measuring: reasoning and communication, not specialist AI knowledge. Fortune reported that this shift has made liberal arts majors suddenly relevant again, because the skill that matters most isn’t coding or prompt engineering. It’s the ability to think clearly, ask the right questions, direct AI toward a useful output, and critically evaluate what comes back. [5]

That’s a significant signal for anyone hiring or being hired in the next two to three years. The firms that matter are no longer testing for information retrieval or raw analytical horsepower. They’re testing for the ability to collaborate with AI effectively.

What this means for your team (not McKinsey)

Here’s the translation. McKinsey has 40,000 people and a proprietary AI platform with 18 months of development behind it. You probably don’t. But the underlying logic scales to any size:

1. Identify your repetitive knowledge tasks. What does your team spend time on that is information-heavy but not judgment-heavy? Research synthesis, first-draft reports, data compilation, meeting prep, email drafting. These are the agent-addressable tasks.
2. Deploy AI for first-pass work, not final output. McKinsey doesn’t publish Lilli’s drafts directly to clients. Humans review, refine, and apply contextual judgment. The same should be true for your team. AI handles the time-consuming starting point. Humans handle the parts that require relationships, nuance, and accountability.
3. Measure the shift in where your team’s time goes. If AI is working, your team should be spending less time on information gathering and more time on interpretation and decision-making. If the ratio hasn’t changed, the integration isn’t working yet.

A 10-person team can run this playbook with ChatGPT Business or Claude for Work, a shared project space, and consistent prompt templates. It doesn’t require a custom platform. It requires a change in how work gets started.

The business model shift no one talks about enough

McKinsey’s CEO has also confirmed that the firm is migrating away from pure fee-for-service consulting toward an outcomes-based model. Fees are increasingly tied to the measurable impact of their work, not the hours delivered. [4]

This matters because it reflects a broader shift that AI is accelerating: when AI can handle volume and speed at near-zero marginal cost, the value is no longer in the delivery of information or analysis. It’s in the judgment, the relationship, and the accountability for results. Firms that charge by the hour for knowledge work are already facing pressure. The ones that charge for outcomes have a more durable model.

If you’re in consulting, professional services, or any knowledge-intensive business, this is worth thinking about now. Not because you need to match McKinsey’s AI infrastructure, but because the question of where you create genuine value (versus where AI can replicate it) is one that clients will start asking more directly.

Three questions to answer this week

Rather than telling you to “embrace AI” in a vague and useless way, here are three specific questions worth sitting with:

1. Which 3 tasks in your team’s week are information-intensive but not judgment-intensive? Research compilation, data entry, first-draft documents, meeting summaries. These are your starting points for AI delegation.
2. Are you testing AI ability in your hiring? You don’t need to run a Lilli-style assessment. But asking candidates to complete a work sample using an AI tool (with full transparency that this is the exercise) tells you something important about how they’ll actually function in an AI-augmented environment.
3. What do your clients pay you for that AI can’t replicate? Your judgment, your relationships, your accountability, your network. These are the parts of your value proposition to protect and emphasize. Everything else is worth automating.

McKinsey’s 25,000 agents aren’t the point. The point is that one of the world’s most conservative professional services firms moved that fast, that deliberately. The question isn’t whether this is coming for your industry. It already has.

Keep Reading

Most Companies Are Getting Enterprise AI Wrong: What the 2026 Research Reveals

ChatGPT Agent Mode: What It Can Actually Do for Business Professionals

AI Is Creating More Jobs Than It Eliminates: What the Latest Research Shows

Sana Mian – Co-Founder, Future Factors AI

Sana is an AI educator and learning designer specialising in making complex ideas stick for non-technical professionals. She has trained 2,000+ learners across corporate teams, bootcamps, and keynote stages. Future Factors offers AI Bootcamps, Corporate Workshops, and Speaking & Consulting for businesses ready to adopt AI without the overwhelm.

More about Sana →