We didn't get a technology announcement last week. We got a Rorschach test. McKinsey, the consulting firm that has spent a century selling other people's certainty, reportedly told the world it now has "up to 25,000 AI agents" deployed across a 60,000-person workforce. Four data points. No named executive. No statistical methodology. No date. No definition of what an "agent" actually is. And yet the number traveled โ from trade press to Crypto Briefing to every LinkedIn feed that has ever used the word "transformation" unironically.
I have been auditing enterprise AI deployments and DAO governance frameworks for most of my career, and I can tell you with confidence: when a professional services firm releases a number this round and this undefined, the number is not the message. The message is that they felt compelled to release a number at all. That compulsion is the real signal, and it points at something more structural than any agent count: the billable hour is dying, and the firms that built empires on it know it.
Context: Why a Consulting Giant Suddenly Cares About Agent Counts
McKinsey is a partnership. No shareholders, no quarterly earnings call, no investor relations department under pressure to spin. When it chooses to leak a statistic, the audience is not Wall Street. It is clients and talent โ the two constituencies that determine whether a knowledge-leverage business survives the next decade.
The firm has had an internal generative AI platform called Lilli for a few years now. Based on my own experience building and auditing retrieval-augmented systems for governance clients, Lilli sits squarely in what I'd call the L1โL2 maturity band: prompt templates, knowledge retrieval, draft generation. Useful. Reliable enough. Nowhere near an autonomous digital employee.
So what does "25,000 agents" mean? In current engineering reality, an enterprise "agent" is almost always a bounded workflow โ an LLM wrapped around a retrieval layer wrapped around a few tool calls. The autonomy ladder runs from L1 (template assistants) through L2 (single-tool workflows) to L3 (multi-step planning) and L4 (long-horizon autonomous execution). Public deployments cluster hard at L1 and L2. L3 and above degrade fast because error compounds exponentially with each step in the chain. A 95% per-step reliability rate collapses to roughly 60% over ten steps. Consulting deliverables carry professional liability. You do not put a hallucinating agent in the signing line.
Which is why the count almost certainly includes every custom assistant a consultant ever spun up, every orchestration node, every dormant experiment. Count like that and you can hit 25,000 without a single agent that runs unsupervised. The word "agent" was chosen over "Copilot" or "assistant" deliberately โ it carries an implication of autonomy that the engineering does not support. This is marketing language wearing an engineering costume.
Core: The Billable Hour Paradox Eating Professional Services Alive
Here is the mechanical problem nobody at McKinsey wants to say out loud. Their revenue is roughly hours times rate. If AI compresses the hours required for a given deliverable by 30 to 50 percent โ and in document-heavy, research-heavy consulting work, that compression is real โ then revenue shrinks by the same proportion at constant rates. Efficiency improvement is revenue destruction. The industry's core product is its own disruption.
That is why the only strategically meaningful phrase in the entire disclosure was "outcome-based engagements." Everything else is decoration. Outcome pricing is the structural response to a model that is eating itself, and it is brutally hard to execute for three reasons I have watched play out in DAO treasury governance, which faces the identical attribution problem.
First, attribution. When a client's performance improves, how much was the consulting? Clients have every incentive to claim none of it. In decentralized governance I have seen the same fight over treasury yield: multiple agents, one outcome, endless argument about who earned the fee.
Second, risk transfer. If you price on outcomes, you absorb delivery risk and must charge a premium for it. Clients resist that premium. The spread between what the risk is worth and what the market will pay is thin and getting thinner.
Third, cash flow. Hourly billing produces predictable recurring revenue. Outcome billing produces lumpy event-driven revenue. For a partnership that distributes profits annually, that volatility is an existential governance problem, not an accounting footnote.
And here is the part the press release omitted entirely: in a competitive market, most of the efficiency gain flows to the buyer. Consulting firms compete on price at the margin. If everyone's AI cuts delivery cost 40 percent, the savings show up as lower bids, not higher margins. We watched this exact dynamic with offshore delivery centers two decades ago. The arbitrage got competed away. The firms that survive outcome pricing will not be the ones with the most agents โ they will be the ones who lock in standardized outcome metrics and multi-year contracts before the price war starts.
Now the reverse evidence, which I find far more interesting than the headline number. If the 60,000 headcount figure is accurate โ and note it diverges from the roughly 45,000 figure in public reporting, so treat it as unverified โ then McKinsey grew its human workforce while deploying AI. That simultaneously falsifies the lazy "AI replaces consultants" narrative and confirms the more uncomfortable one. The pyramid is not collapsing at the top. It is narrowing at the bottom. Junior analysts do data collection, model building, slide assembly โ precisely the L1โL2 work agents handle well. The partner layer, where judgment and client trust live, is untouched. The entry point is what's disappearing.
I have seen this movie before in crypto. When automated market makers replaced order-book market making, the jobs that vanished were not the senior quants. They were the junior desk analysts who used to learn the craft by doing the repetitive work. The judgment layer survived. The training pipeline did not. Consulting is about to run the same experiment on its own talent ladder, and nobody has designed a replacement for the apprenticeship that junior grunt work provided.
Let me put a rough number on the infrastructure question, because it reframes where the real constraint sits. Assume all 25,000 agents run heavy โ say 100 calls per day each, 6,000 tokens per call. That's roughly 15 billion tokens daily. At enterprise blended pricing of fifty cents to three dollars per million tokens, you're looking at somewhere between $2.7 million and $16.5 million annually. Against a firm with revenue in the neighborhood of $16 billion, that is 0.02 to 0.1 percent. Noise. Compute cost is not the bottleneck. Reliability, compliance, and human reallocation are. Any firm treating inference spend as its AI constraint is solving the wrong equation.
Contrarian: The Number Is a Distraction From the Only Metrics That Matter
Here is where I'll push against my own instinct to dismiss the disclosure. The contrarian read is not that McKinsey is lying. It is that the agent count, however inflated, is a leading indicator of a measurement regime that is about to infect the entire professional services industry โ and that regime is broken.
Agent count is a vanity metric. It counts inputs, not outputs. It tells you nothing about task completion rates, human takeover frequency, or full-time-equivalent hours displaced. If the industry standardizes on instance counts the way crypto standardized on total value locked โ a number that was famously easy to inflate and famously disconnected from actual economic activity โ then every board evaluating AI investment will be flying on instruments that read high by an order of magnitude. I watched TVL become a marketing arms race in 2020 and 2021. Protocols double-counted the same liquidity across chains. The metric detached from reality so thoroughly that sophisticated allocators stopped trusting any of it. Agent counts are on the same trajectory.
The honest leading indicators are unglamorous: revenue per employee, junior hiring volume, and outcome-based contract share. If revenue per employee is not rising, the AI investment is not converting to economic value regardless of how many agents exist. If junior hiring is flat while agent counts triple, the augmentation story is a press release, not a strategy. And if outcome contracts stay under ten percent of revenue, the pricing transformation is theater.
There is also a data governance vacuum hiding behind the announcement. Consulting engagements involve a client's most sensitive strategic and financial data. Routing that through third-party model APIs means the data leaves the client's perimeter. Most enterprise contracts restrict this explicitly. Whether McKinsey runs self-hosted models or zero-retention APIs determines whether the whole deployment is even contractually legal โ and the disclosure says nothing. Add the liability question: if an agent hallucination drives a bad strategic call, who is responsible? The signing partner? The firm? The model vendor? Professional liability insurance has no settled answer, and the disclosure does not pretend to offer one.
Takeaway: Watch the Pyramid, Not the Counter
The 25,000-agent number will be forgotten within a quarter. What will not be forgotten is the admission embedded beside it: a firm that has sold time for a hundred years is now selling outcomes, because time is no longer scarce. We didn't need the exact count to know the billable hour was under threat. We needed a century-old institution to blink first, and it did.
The question worth carrying forward is not how many agents McKinsey has. It is whether the apprenticeship that turned sharp twenty-five-year-olds into trusted advisors can survive the automation of the work that used to train them. If it cannot, the firms will keep their margins and lose their future partners โ and the number they announce next year will be higher, and mean even less. Freedom isn't the absence of constraint. It's the presence of consent โ and the next generation of professionals never consented to an apprenticeship that no longer exists.