McKinsey now operates 25,000 AI agents alongside 40,000 humans. Eighteen months ago, that number was 3,000. CEO Bob Sternfels originally expected to reach one agent per human by 2030, and he now thinks it will happen within 18 months.
The strategy behind this shift is what Sternfels calls “25 squared”: grow client-facing roles by 25%, cut non-client-facing roles by 25%. The agents own research and synthesis, while humans own judgment and relationships. The firm saved 1.5 million hours in 2025 on tasks that agents now handle, and non-client output is growing 10% despite fewer people doing it.
But before you treat McKinsey’s numbers as a roadmap, it is worth understanding why consulting was the easy case.
Why Consulting Was the Easy Case
Consulting has a structural advantage that most industries do not share. The split between client-facing work and back-office synthesis is unusually clean. Research, data assembly, report drafting, and pattern matching across engagements are all tasks that large language models handle well. The roles that require human presence, reading a boardroom, navigating politics, building trust with a CFO, are clearly separable from the roles that agents can absorb.
That clean division is what makes 25 squared work at McKinsey. In manufacturing, healthcare, logistics, or financial services, the boundary between judgment and execution is rarely so obvious. A warehouse manager’s decision-making is embedded in the physical workflow. A clinician’s synthesis happens at the point of care. The work that agents could theoretically do is tangled with the work that humans must do, and separating them requires redesigning processes, not just reassigning tasks.
McKinsey’s hiring shift reinforces this point. The firm now tests candidates on their ability to challenge AI output, not just produce analysis. Their internal tool Lilli is part of final-round interviews, and what interviewers watch for is curiosity and judgment. They are hiring for the skill of working alongside agents, which is a luxury that only makes sense when the agent’s role is already well-defined.
The Structural Question Most Organizations Have Not Asked
Rivals like EY and PwC argue that counting agents is a vanity metric. What matters, they say, is output quality and cost impact. They have a point, but they are also responding to the surface of what McKinsey did while missing the structural shift underneath.
The real signal is not the agent count. It is that McKinsey looked at every role in the firm and asked two questions: does this role require human judgment, and does this role require human presence? Roles that required neither became candidates for agent augmentation. That exercise, not the technology deployment, is what most organizations have not started.
Gartner projects that more than 40% of agentic AI projects will fail by 2027 due to legacy system incompatibility. Deloitte reports that only 11% of organizations are actively using agentic AI in production. The gap between McKinsey’s pace and everyone else’s is not about willingness to invest. The infrastructure bets are already being placed. The gap is about willingness to ask the structural question and act on the answer.
Running Your Own Version
The 25-squared math will not translate directly to your organization. The ratios will differ, the timeline will differ, and the roles that agents can absorb will depend on how separable judgment is from execution in your specific workflows.
But the exercise itself translates everywhere. Meta’s CEO is taking this further, building an AI agent that questions whether the information path itself needs to exist. Map every role against the two questions McKinsey asked. Identify where the boundary between human judgment and machine execution actually falls in your operations, not where you assume it falls. Then test whether your systems, your data infrastructure, and your team’s willingness to work alongside agents can support the shift.
The companies that do this now will spend the next two years redesigning deliberately. The ones that wait will eventually face the same restructuring, but on someone else’s timeline, and with far less room to shape the outcome.
Questions this article gets
What is McKinsey's 25-squared AI strategy?
25 squared is McKinsey CEO Bob Sternfels' framework for restructuring the firm around AI. It means growing client-facing roles by 25% while cutting non-client-facing roles by 25%. The agents handle research and synthesis, while humans focus on judgment and relationships.
How many AI agents does McKinsey use?
McKinsey currently operates approximately 25,000 AI agents alongside 40,000 human employees. That number grew from 3,000 just eighteen months ago, and Sternfels expects to reach one agent per human within the next 18 months.
Why is consulting uniquely suited to the 25-squared model?
Consulting has a clean structural split between client-facing judgment work and back-office synthesis work. The roles that agents can absorb, such as research, data assembly, and report drafting, are clearly separable from the roles that require human presence. Most industries do not have such a clear dividing line.
How can organizations run their own 25-squared exercise?
Start by mapping every role in your organization against two questions: Does this role require human judgment, and does this role require human presence? Roles that require neither are candidates for agent augmentation. Then test whether your systems can actually support that shift, since legacy infrastructure is the most common reason agentic AI projects fail.