McKinsey Deploys 20,000 AI Agents as Digital Employees Inside Lilli Platform

McKinsey & Company has crossed a threshold that most enterprises are still debating in PowerPoint: the firm has deployed approximately 20,000 AI agents inside its internal platform, Lilli, effectively giving every consultant a digital counterpart that handles research, synthesis, and analysis tasks around the clock.

These aren't chatbots with guardrails. McKinsey's agents function as digital employees — autonomous workers that take on discrete tasks, hand off outputs, and operate across the firm's knowledge base without waiting for a human to prompt each step.

What Lilli Actually Does

Lilli is McKinsey's proprietary AI knowledge platform, built on top of large language models and trained on the firm's internal research, methodologies, and client-facing frameworks. At its core, it gives consultants a way to query decades of institutional knowledge without manually digging through databases or asking a senior partner.

The agent layer takes it further. Rather than a consultant running a single query, AI agents can:

• Draft research briefs by pulling from internal and external sources autonomously
• Build slide structures based on engagement context and past deliverable patterns
• Run scenario analyses across financial models without human hand-holding at each step
• Monitor and summarize industry developments relevant to active engagements

The result is a firm where a significant portion of the analytical groundwork — the kind of work that used to consume a first-year analyst's 80-hour week — is handled before a human reviews the output.

20,000 Agents: What That Number Means

Scale matters here. McKinsey employs roughly 45,000 people globally. Deploying 20,000 AI agents means the firm has effectively doubled its working capacity on the knowledge work side — without doubling headcount.

Each agent isn't a standalone system. They operate within workflows, triggered by consultant needs or automated pipelines, completing tasks and feeding outputs back into Lilli's central knowledge layer. The more agents work, the richer the platform's context becomes.

This is the compounding effect that enterprises building agent infrastructure are betting on: agents that generate outputs that train better agents.

The Broader AI Agent Moment

McKinsey's deployment isn't happening in isolation. Across the enterprise landscape, 2025 and early 2026 have seen a decisive shift from AI assistants to AI agents — systems that don't just respond but act, plan, and execute over multi-step workflows.

Key developments driving this shift:

• Anthropic, OpenAI, and Google have all released agent frameworks explicitly designed for enterprise multi-agent orchestration
• Agentic workflows have moved from research prototypes to production deployments at companies including Salesforce, ServiceNow, and dozens of financial institutions
• The "agent workforce" framing — where AI agents are treated as digital employees with defined roles, access scopes, and performance metrics — is becoming standard language in enterprise AI strategy

McKinsey's 20,000-agent deployment is the clearest signal yet that this framing has moved from metaphor to operational reality.

What Changes When Agents Are the Default

The implications for how consulting work gets done are significant.

Junior analysts at McKinsey have historically been the engine of the firm — the people who run the models, build the slides, and synthesize the research. As agents absorb more of that work, the role of early-career consultants shifts toward task definition, output quality control, and client-facing interpretation. The work doesn't disappear; it moves up the stack.

For clients, the experience may change too. Engagements that previously required weeks of research setup can potentially begin with richer initial analysis, faster iteration cycles, and more scenario coverage than a human team alone could produce in the same timeframe.

The question McKinsey — and every enterprise deploying agents at scale — has to answer: when the agents handle the groundwork, what does the human judgment layer actually look like, and how do you keep it sharp?

The Competitive Pressure This Creates

McKinsey publishing these numbers — or allowing them to circulate — is not accidental. Enterprise AI deployment is increasingly a competitive signal. Clients choosing between consulting firms want to know which firm's infrastructure gives them faster, more rigorous output.

Lilli and its 20,000 agents are, at this point, part of the pitch.

Other major consultancies are building comparable systems. BCG has invested heavily in its own AI platform infrastructure. Deloitte and Accenture have both announced significant agent deployments. But McKinsey's scale and the specificity of the 20,000-agent figure makes it the current benchmark.

The race to the agentic enterprise is underway. McKinsey just posted its lap time.