McKinsey has up to 25,000 AI agents. McKinsey has 60,000 employees. The ratio is 1:2.4. The headline writes itself, and that is precisely the problem.
I have spent enough time in on-chain forensics to recognize a specific pathology when I see it: a number that travels faster than its methodology. In 2017, I watched token projects publish "community size" figures that counted wallet creations rather than active users, and the market priced the number anyway. The McKinsey agent count is the same instrument, retuned for enterprise AI. It is a metric without a denominator, a claim without a ledger.
What follows is not a takedown. It is an audit. And the audit's first finding is that the source material is nearly empty: four information points, three of which restate each other; a core figure with no definition, no sampling method, and no timestamp; a title attributing the claim to McKinsey with no executive name, no interview citation, and no publication date. The employee base of 60,000 diverges materially from publicly reported headcount in the 45,000 range. Zero technical detail. One commercial phrase: outcome-based engagements.
s silence.
But an empty ledger still tells you who kept it and why. Three things are worth analyzing here, and none of them is the number itself: the actual deployment stage of AI inside professional services, the collapse of "agent" as a countable unit, and the diffusion of AI narrative into non-technical vertical media. Let me work backward from the conclusion, the way any forensic accountant would.
The Counting Problem Nobody Wants to Solve
Start with the definitional vacuum. There is no industry standard for what constitutes an "agent." The same number could refer to custom GPT instances, Copilot seats, workflow orchestration nodes, or API keys. Between these categories the variance spans one to two orders of magnitude. A firm that counts every experimental assistant a consultant spun up on a Tuesday afternoon and a firm that counts autonomous multi-step processes running unattended are both entitled to say "we have X agents." They are not describing the same universe.
In my audit work I use a four-level maturity model. L1 is prompt templates and retrieval-augmented Q&A. L2 is single-tool workflow invocation. L3 is multi-step planning across multiple tools. L4 is long-horizon autonomous execution. Public enterprise deployments cluster hard at L1 and L2. L3 and above collapse under error accumulation, because long-chain success probability decays exponentially with step count. A workflow with 95% per-step reliability survives ten steps at roughly 60%. It survives twenty steps at roughly 36%. Professional services delivery tolerates neither figure.
McKinsey already operates a public generative AI platform called Lilli. Its documented function is knowledge retrieval and draft generation. That is textbook L1-L2. It is not an autonomous digital workforce, and nothing in the reporting suggests otherwise.
The number 25,000 almost certainly counts deployed assistant and workflow instances. If a single consultant can create multiple assistants, the count inflates mechanically. Low-frequency, experimental, and abandoned instances all get counted. This is not fraud. It is the natural upward bias of a metric that counts creations rather than outcomes.
The word choice matters too. "Agent" carries an autonomy connotation that "Copilot" or "assistant" does not. That is marketing language, not engineering language. And the article omits the three figures that would actually measure deployment depth: utilization rate, task completion rate, and human review rate.
Compare this to on-chain metrics I use daily. Active addresses are a vanity metric; unique signers weighted by gas expenditure and retention across epochs are a real one. The agent count is active addresses. Nobody in this story is publishing the equivalent of realized value.
The Business Model That AI Is Eating From Inside
Here is the signal that actually matters, buried in four words: outcome-based engagements.
Consulting revenue approximates hours invested multiplied by billing rate. If AI reduces the hours required for a given deliverable by 30 to 50 percent, revenue contracts proportionally at constant rates. This is the industry's self-cannibalization paradox, and it is the reason every professional services firm must eventually change its pricing anchor. The shift to outcome pricing is not strategic ambition. It is structural survival. The article treats it as a forward-looking initiative when it is closer to a controlled demolition.
The obstacles are three, and none are trivial.
Attribution is the first. Enterprise performance improvement is multi-causal. Clients have a natural incentive to deny consulting contribution, particularly when the invoice is tied to a measured outcome they can dispute. I have watched DeFi protocols attempt the same structure with performance fees, and the disputes always converge on the same question: what would have happened anyway? Defining the counterfactual is where these contracts die.
Risk transfer pricing is the second. When the consultant bears delivery risk, the quote must include a risk premium. Clients systematically undervalue that premium until the first disputed engagement. The pricing conversation becomes adversarial before the work begins.
Cash flow volatility is the third. Hourly billing produces periodic, predictable revenue. Outcome billing produces event-driven, lumpy revenue. For a partnership that distributes profits annually, that shift creates genuine internal friction. Partners plan distributions on predictable cash. Structural unpredictability is not a neutral change to a partnership's operating model.
The industry has already validated the direction. Accenture, BCG, and Bain are all pushing outcome, subscription, and managed-service contracts. This is a sector-wide migration, not a McKinsey idiosyncrasy. Which means the agent count is not a competitive differentiator. Everyone building the same thing means nobody is differentiated by building it.
What the reporting omits is where the model succeeds or fails: the outcome metric definition, the baseline methodology, the dispute resolution mechanism, and how AI-derived savings are split between firm and client. Those four items determine whether outcome pricing is a business model or a press release.
My prior from watching efficiency gains in offshoring is that the surplus accrues to the buyer. In a competitive market, productivity gains get competed away into price. The firm's net gain equals the efficiency improvement minus the price decline, and history suggests that spread compresses toward zero. The real purpose of outcome pricing is client lock-in and renewal rate, not unit margin expansion.
And note who benefits from disclosing the agent figure at all. McKinsey has no external shareholders and no quarterly earnings pressure. The audience for this disclosure is clients and talent, not capital. The number is a sales asset. It says: we have already run this transformation internally, so buy our version of it. That explains the timing and the framing. It does not explain the methodology.
The Pyramid Narrowing, Not the Headcount Falling
The most underreported fact in this story is the one that contradicts the obvious narrative. If the 60,000 figure is accurate, McKinsey grew headcount while deploying AI at scale. Agent growth and employee growth are happening simultaneously.
That single observation falsifies the short-term "AI replaces consultants" thesis. What is actually occurring is augmentation, with a structural squeeze concentrated at one specific layer.
Consulting economics are built on leverage: a small number of partners multiplying the output of a large number of junior analysts who do data collection, model construction, and slide production. AI agent capability maps almost perfectly onto that layer's core tasks. This is the point of maximum substitution pressure, and it is not the senior judgment layer.
The observable signals sit on the recruiting side. Top business school MBA placement into consulting has declined in recent years. Headline firms have slowed campus hiring and reduced junior class sizes while expanding AI, data, and implementation roles. That is structural substitution, visible in the pipeline before it is visible in the totals.
If headcount and AI are growing together, the firm is using AI to expand the scope of work it can accept and the volume it can deliver, rather than compressing the establishment. That is a Jevons-style path: efficiency gains expand demand. It also makes revenue per employee the single most important observable metric. If revenue per employee does not rise, the AI investment has not converted into economic value, and the agent count is decoration.
There is a second-order effect the article ignores entirely. A consulting firm's internal practice gets packaged and sold. McKinsey is simultaneously an AI adopter and an AI diffusion amplifier into financial services, manufacturing, and consumer sectors. The firm's internal transformation is a product demo that gets deployed inside client organizations, which accelerates substitution pressure far beyond consulting itself.
| Affected Group | Substitution Rate | Augmentation Rate | Window | Key Judgment | |---|---|---|---|---| | Consulting junior analysts | Medium-high (40-60%) | High | 12-24 months | Entry points narrow, pyramid base contracts | | Project managers and partners | Low (<15%) | High | 24-36 months | Judgment and client relationships remain human assets | | Corporate strategy and analytics | Medium (20-40%) | High | 12-30 months | Consulting deliverables internalized, external spend may fall | | Legal and audit services | Medium | High | 18-36 months | Isomorphic logic: leverage model plus document density | | Content and research outsourcing | High (>50%) | Medium | 6-18 months | Most directly squeezed adjacent layer |
Logic is the only audit that never expires. And the logic here says the pyramid base narrows while the apex holds.
The hidden cost is not layoffs. It is the broken training path. If junior analysts no longer perform the foundational work, the profession loses its mechanism for producing judgment-capable senior partners. Ten years out, the supply of experienced principals degrades. This is a structural risk that is invisible in the short term and irreversible in the long term. I have watched protocols make the same mistake with liquidity mining: extract the bootstrapping phase for efficiency, then discover you have removed the mechanism that created competence.
The Compliance Vacuum and the Vanity Metric
Consulting engagements handle a client's most sensitive strategic and financial data. Agent invocation against a third-party model API means that data leaves the perimeter. Most client contracts restrict this explicitly. Whether McKinsey runs self-hosted models or zero-retention endpoints determines legal feasibility, and the reporting says nothing about it.
Responsibility attribution is the second gap. If an agent hallucination drives a client's strategic misjudgment, who is liable? The signing partner, the firm, or the model vendor? Does professional liability insurance cover AI-assisted delivery? These are unresolved legal gray zones, and every firm deploying at scale is operating inside them.
The disclosure obligation is the third. Do clients have the right to know that a deliverable was AI-generated, and in what proportion? There is no mandatory disclosure requirement today, but buyer pressure is rising. Watch for standardized contract clauses requiring AI participation disclosure. That clause, when it appears, will be the real marker of institutional adoption.
On infrastructure, a rough magnitude check is worth running because it settles an argument. Assume 25,000 agents at 100 calls per day each: 2.5 million daily calls. At roughly 6,000 tokens per call (5,500 input, 500 output), that is about 15 billion tokens daily. At blended enterprise pricing of $0.5 to $3 per million tokens, that is $7,500 to $45,000 per day, or roughly $2.7 million to $16.5 million annualized. Against a firm with revenue in the $16 billion range, that is 0.02% to 0.1%. Noise.
The implication is decisive. Inference cost is not the binding constraint on enterprise AI deployment. Reliability, compliance, and human reallocation are. Any analysis that frames enterprise agent adoption as a compute problem has misidentified the bottleneck. This is structurally different from consumer AI products, where inference cost genuinely drives unit economics.
Which brings me back to the vanity metric. If the cost is noise and the capability is L1-L2, then the agent count measures neither cost pressure nor value creation. It measures activity. Activity metrics are the oldest trick in the ledger: they always go up, they never require a denominator, and they never survive contact with revenue per employee.
The risk is imitation. If "agent count" becomes the industry's standard disclosure, enterprise AI evaluation systems will be built on a unit that has no economic meaning, and capital will flow toward firms that create the most instances rather than the most value. The correction, when it comes, will be expensive.
What the Ledger Will Show
Follow the money, not the narrative. The money here is not in the 25,000 figure. It is in the four words the article treats as a footnote. Outcome-based engagements are the confession that time-and-materials billing cannot survive the efficiency it just created. That is the structural event. Everything else is packaging.
So stop tracking agent counts. Track three numbers that cannot be inflated by counting creations: revenue per employee, junior hiring volume, and the share of revenue under outcome-priced contracts. If revenue per employee rises while junior hiring falls and outcome contract share climbs, the transformation is real. If revenue per employee stays flat while agent counts multiply, you are watching a firm measure its own activity while its economics stand still.
The data will not disappear. It will simply wait to be found, and it will be found in the hiring pipeline long before it appears in a press release.
One question for the next six months: when the first client contract clause requiring AI participation disclosure gets signed, will it be the buyer or the seller who insisted on it? The answer will tell you who actually understands what is being sold.