Why AI just made consulting more expensive
What you actually buy when you hire McKinsey
You did not hire McKinsey for the deck. You can prove this to yourself in an afternoon: feed your strategy problem to Claude, let it generate seventy-five slides, and notice that you would never send the result to your board. The slides look right. The logic tree is clean. The pyramid principle is intact. And you would still pay a partner $1,193 an hour to throw it away and start over.
That instinct is worth examining, because most people are drawing the wrong lesson from it. The common read is that AI decks are still bad and will get better, so consulting has a few years left. The actual lesson is that the deck was never the product. When a generative model can produce the visible artifact of a knowledge job at near-zero cost, it exposes a split that was always there and that almost nobody priced correctly: the gap between the task and the job.
The Shift
Benedict Evans framed it on Lenny’s podcast in May, and the framing is sharper than the usual “AI will take jobs” discourse because it refuses to treat a job as one thing.
“Are you hiring McKinsey to get a seventy-five-slide deck? Claude Code will make a really, really crappy version of that. Even if it was good, that’s not what you paid them for. What you actually pay Bain to do is go walk all over your enterprise and work out, yes, but why didn’t you do that, and how do the politics work.”
The artifact is the task. The walking-around, the politics, the figuring out what the client will actually execute: that is the job.
Watch where the cuts are landing and the split becomes visible in the org chart. McKinsey is reducing headcount by roughly 10%, somewhere between three and four thousand people, across 2025 and 2026 (Inc / Fast Company, 2026). The cuts are concentrated in back-office functions, junior research, and synthesis roles, the parts of the firm whose output is a draft, a chart, a literature scan. The firm’s own Quantum Black researchers put the productivity gain on research and synthesis at 30% or more per engagement. And revenue over the same five years sat flat at $15 to 16 billion (Inc, 2026).
Read those three numbers together. The task layer is collapsing: fewer analysts, faster synthesis, compressed draft cycles. Revenue is holding. A senior partner still bills $1,193 an hour; an analyst bills $327 to 498 (federal GSA rates via RoadToOffer, 2026). The firm is shedding the layer AI cheapened and keeping the layer it didn’t touch, at the same price. That is the task and the job, separating in real time on a P&L.
Mechanism
Slow down on this part, because the surface story (”AI is automating consulting”) gets the direction exactly backwards.
Every knowledge job is a bundle. On one side sits the task: the artifact you can point to, hand over, and check. The deck. The code. The fetched SKU. The financial model. The contract draft. The customer-service reply. On the other side sits the job: the judgment that decided which artifact was worth making, the context that made it fit this client and not a generic one, the politics that determine whether anyone executes it, and the accountability that means a human’s name is on the outcome.
For most of the history of professional work, these two were welded together because you could not buy one without the other. The only way to get a partner’s judgment was to also pay for the team that produced the deck. The judgment was real, but it was never invoiced as a separate line item. It rode along, bundled into the artifact, priced as if the artifact were the thing.
A generative model is a machine for unbundling exactly this seam. It drives the marginal cost of the artifact toward zero and does nothing to the judgment that should govern the artifact.
Evans is precise about which side is which. The task is getting the SKU, writing the code, building the spreadsheet, writing the spec. The job is the implicit knowledge, the opinion, the taste, the ideas, the part you cannot write down as a rule. That last clause is the load-bearing one, and it has a history. The reason you cannot write the job down as a rule is the same reason expert systems failed in the 1980s. You cannot enumerate the steps a senior partner takes to know that a client’s real problem is a feud between two division heads, any more than you could enumerate the steps to recognize a cat. The judgment is non-decomposable. It resists being turned into the kind of explicit, checkable artifact that a model can now generate on demand.
So the model eats the decomposable half of the bundle and leaves the non-decomposable half standing.
Now follow the second-order effect, because this is where the apocalypse narrative inverts. When artifacts were expensive, judgment was hidden inside their price and partly subsidized by it. When artifacts go to zero, two things happen at once. The judgment loses its subsidy and has to stand on its own, which exposes everyone whose “judgment” was actually just the ability to produce the artifact. And the judgment that is real becomes more valuable, not less, because the artifact it governs is now cheap and undifferentiated. When anyone can generate seventy-five plausible slides, the scarce thing is knowing which slide is correct and which recommendation the org will actually run.
Cheap artifacts raise the price of the judgment that selects among them.
This is why McKinsey can cut a tenth of its people and hold revenue flat. It is monetizing the productivity gain by shedding the subsidized task layer and repricing the judgment layer at full freight. The same dynamic shows up in the billing structure: clients increasingly refuse to pay for analyst-hours and instead pay for outcomes tied to operational results (TheStreet, 2026). An hour was a unit of task; an outcome is a unit of job, and the buyer has started paying for the second one directly now that the first one is cheap.
It helps to see what this does to the underlying business model, because consulting’s economics were built entirely on the bundle. The classic firm runs a pyramid: a few expensive partners at the top, a wide base of cheap analysts underneath, and the margin comes from billing the analysts’ time at a steep markup over their salary. The pyramid only works because the artifact requires labor. Someone has to build the model, scan the literature, format the deck, and that wide base of billable analyst-hours is what funds the partner.
Generative models knock out the base of the pyramid. When the artifact no longer needs a floor of analysts, the markup that funded the partner disappears, and the firm has to justify the partner on judgment alone. Some firms will discover that their partners were senior analysts in disguise, fast at artifacts, thin on judgment, and those firms will compress toward the elevator pole. The firms whose partners hold real, non-decomposable judgment will keep the rents and shed the floor, which is precisely the shape of the McKinsey cut: the base thins, the apex holds price. The org chart is the decomposition, drawn in headcount.
The cleanest proof that the two are separable is what happens when someone tries to sell the task alone, priced as if it were the job. Klarna ran the experiment for everyone. Between 2022 and 2024 it cut roughly 700 customer-service roles and routed two-thirds to three-quarters of interactions through an OpenAI-built assistant (Entrepreneur, 2026). The AI handled the task, the routine, high-volume, decomposable query, competently. Then complaints rose, satisfaction fell, and by 2025 the company was rehiring humans for escalations, complex cases, and high-value relationships. “We focused too much on efficiency and cost,” the CEO admitted. “The result was lower quality.” Klarna had automated the task and assumed it had automated the job. The job, judgment under ambiguity, accountability to an angry customer, was a different thing wearing the same job title.
There is a power dimension here that decides who captures the repriced judgment, and it is moving fast. If judgment is the scarce asset, the people who own access to enterprise judgment own the rents. The AI labs have noticed. In May 2026, OpenAI stood up “The Deployment Company,” a joint venture it majority-controls, capitalized at over $4 billion, and used it to acquire Tomoro, an applied-AI consultancy bringing roughly 150 engineers with deployment scars from Tesco, Virgin Atlantic, and Supercell (TechCrunch, 2026; aibusiness, 2026). Days earlier, Anthropic launched a $1.5 billion AI-native services venture with Blackstone, Hellman & Friedman, and Goldman Sachs (Fortune, 2026). The companies that build the artifact-generation machines are spending billions to own the judgment-and-deployment layer on top of it. They understand their own product better than the market does: the model commoditizes the task, which moves the margin into the job, so they are buying the job before the rest of the market reprices it.
The Model: Task/Job Decomposition
Name the tool so you can reuse it. Task/Job Decomposition is a two-step operation you run on any role, product, or company.
Step one: split the bundle. For the role in front of you, separate the task, the artifact a model can produce and a human can check, from the job, the judgment, context, politics, and accountability that decided the artifact was the right one. Write both columns explicitly. The task column is what gets automated. The job column is what gets repriced.
Step two: locate the role on the spectrum. Roles live on a line between two poles.
At one pole, the task is the job. Evans’s example is the elevator attendant: the entire value was operating the lever, so when the button arrived, the role vanished. Pushing a button is the whole job, with no residual layer of judgment underneath it. When the task is the job, automation is subtraction, and the role disappears cleanly.
At the other pole, the task is the tip of a judgment iceberg. The visible artifact, the deck, the diagnosis, the architecture, is a thin output sitting on a deep mass of non-decomposable judgment. Here automation commoditizes the tip and raises the price of the iceberg, because the scarce skill shifts from producing the artifact to knowing which artifact is correct.
Most knowledge work that people are panicking about sits closer to the iceberg pole than the elevator pole, and the panic comes from measuring the tip. The genre of “X% of this job is automatable” studies makes exactly this error: it scores the task column, finds it large, and reports the job as endangered. It is counting the part of the iceberg above the water and calling it the iceberg.
The decomposition cuts both ways, and the cut is the point. Run it honestly on your own role and you may find your task column is the whole column, that what you call judgment is really just fluency at producing the artifact. That is the elevator-attendant finding, and it is better to learn it from the model than from a layoff.
Implications
For founders: Stop building tools that produce the artifact and start building tools that own the judgment around it. An AI that generates a contract is a feature; an AI that knows which contract this company should sign, flags the clause that will detonate in eighteen months, and carries the accountability when it is wrong is a business. The artifact layer is racing to zero and you cannot win a race to zero. Build for the job column. Concretely: sell outcomes, not outputs. If your pricing page charges per generated artifact, you have priced yourself on the side of the bundle that is deflating.
For big companies: Audit every function for its task/job ratio before you cut. The McKinsey move is the right one done deliberately: shed the subsidized task layer, reprice the judgment layer, hold revenue. The Klarna move is the wrong one done blindly: automate the task, assume you got the job, eat the quality collapse, rehire at a loss. The difference between them is whether you decomposed first. A layoff that removes the task layer and keeps the judgment layer is restructuring. A layoff that removes both because you could not tell them apart is self-harm with an AI alibi.
For careers: The defensive move and the offensive move are the same move: get off the task layer and onto the job layer. If your day is producing artifacts to spec, the deck, the model, the ticket, the standard reply, your column is the one being priced to zero, and your seniority will not protect you, because seniority that only buys faster artifact production is task-shaped. The durable position is the one that owns problem definition, context, the politics of execution, and accountability for the outcome. Those who can only operate the lever should find a different role inside the same building before the button ships.
For product design. The new default is that the artifact is free and the judgment is the interface. Design products that assume generation is solved and compete on selection, context, and trust. The valuable surface stops being a tool that makes the thing and becomes a system that knows which thing you should make, in your specific situation, and stands behind the call. When generation is a commodity, the product is the judgment wrapped around it: the defaults, the guardrails, the place where a name gets attached to the recommendation.
If I Were Building Today
I would build the judgment layer for one vertical I understand deeply, and I would let the AI labs commoditize the artifact underneath me for free.
Concretely: pick a knowledge profession with a fat judgment iceberg and a cheap artifact tip, commercial real estate underwriting, clinical trial design, M&A diligence, industrial procurement. Use the frontier models to drive the artifact to zero, which the labs are subsidizing anyway. Then build everything the model cannot: the proprietary context about how this industry actually makes decisions, the workflow that captures a senior expert’s judgment as they exercise it, and the accountability layer where the system commits to a recommendation and someone can be held to it. Price for outcomes, never for artifacts, because artifact pricing puts you on the deflating side of the bundle by construction.
I would hire for the job column and rent the task column. My expensive people would be the ones with non-decomposable judgment in the domain, the partner who knows why the client didn’t already do the obvious thing. The artifact production I would treat as infrastructure: cheap, abundant, and someone else’s capex problem.
And I would run Task/Job Decomposition on my own company every quarter, because the seam moves. Each model release shaves another inch off the task tip and exposes another sliver of judgment that turns out to have been artifact-fluency in disguise. The teams that survive are the ones holding the iceberg, not the ones polishing the tip, and the only way to know which you are holding is to keep splitting the bundle and looking at the columns honestly.
The deck was never the product. Build the thing the deck was always standing on.
The task/job decomposition is the clearest framing of this I've read. But the job column carries a hidden load: the judgment that selects the right artifact is still acting on a premise about what 'right' means for this client, this situation, this moment. That premise isn't always verified before the judgment runs. The McKinsey partner who knows why the client didn't do the obvious thing is still operating from a read of the situation that could be wrong — and when it is, the artifact was the least of the problems. The article draws the seam between task and job.
The question one level up is: what verifies the definition the job column is working from, before the judgment commits?
Excellent piece. What strengthens the argument further is that the weld you’re describing was recent.
Before the platform era you paid advisory for the job column directly, the partner who'd walked your halls and would put their name on the call. The 00s shift to multi-year lock-in builds is what fused judgment to a floor of billable artifact-hours and let firms price the bundle as if the deck were the thing.
So what we’re watching isn't consulting being disrupted. It's consulting being handed back the one thing it could never decompose, and finding out how few people were actually still selling it.