The consulting pyramid was never a business model. It was a solution to a coordination problem. AI just solved the problem cheaper.
For seventy years, management consulting sold information processing at scale: hire smart graduates, train them in a methodology, deploy them in hierarchies that could synthesize information faster than any client's internal team. The pyramid shape was structural — you needed many juniors doing research and synthesis to support a few seniors doing judgment and client management. The ratio made sense because cognition at scale was expensive. That cost just hit a floor of approximately zero.
McKinsey is celebrating its 100th anniversary by cutting up to 10% of non-client-facing staff. Accenture eliminated 11,000 positions in three months, spent $865 million on severance, and generated $2.6 billion in AI consulting revenue in the same period — revenue up, headcount down. Deloitte, KPMG, PwC, Bain, EY: every major firm cut in 2025. The industry lost more jobs in one year than it created in the previous five.
This isn't a McKinsey story. McKinsey is just the most visible evidence of a structural shift that applies to every firm built on the same coordination model.
The Talent Arbitrage Is Over
The traditional consulting business model is elegant in its simplicity: hire the smartest graduates from the best universities, train them in a methodology, bill them to clients at three to five times their salary, and pocket the spread. The pyramid shape is structural — you need many juniors doing research and synthesis to support a few seniors doing judgment and client management. The partner's value is directing the pyramid. The junior's value is being the pyramid.
McKinsey hired from 17,000 to 45,000 employees in a decade. The machine needed bodies because bodies were the product. Every slide deck, market analysis, and strategic recommendation required weeks of human research and synthesis. Clients paid for that processing capacity because they didn't have it themselves.
AI doesn't augment this model. It vaporizes the base. McKinsey is now running thousands of internal AI agents that summarize documents, build slide decks, analyze data, check logic, and generate first drafts — the exact work that justified hiring those 28,000 additional people. When a synthesis engine can access 100,000 internal documents and produce a first-pass analysis in minutes, the question every partner is asking becomes unavoidable: why do I need ten analysts when three analysts and an AI produce the same output faster?
The answer is: you don't. And everyone knows it. McKinsey's own 2025 State of AI survey found that 32% of companies expect AI to reduce their workforce by at least 3% within the next year. They published the data. Then they became the data.
Accenture's version of the same story is even more revealing. The firm generated $2.6 billion in AI consulting revenue in six months while simultaneously cutting 11,000 jobs. Revenue up, headcount down. That's not a contradiction — it's the new model revealing itself. Fewer humans, more AI, higher margins. The CEO said it explicitly: the restructuring addresses "rapid AI adoption and declining demand for conventional consulting."
The demand isn't declining for insight. It's declining for the human-hours packaging that insight used to require.
The Obelisk Emerges
One analyst described what's replacing the pyramid as the "obelisk company" — a structure where AI absorbs the execution layers, leaving something slimmer, flatter, and almost entirely dependent on senior judgment and technical specialization. No wide base of juniors. No middle management translating between strategy and execution. Just a narrow column of high-conviction people operating with agents that handle the synthesis, research, and coordination that used to require teams of twenty.
Every major consulting firm is building toward this shape whether they admit it or not. McKinsey has Lilli. Bain has Sage. PwC has ChatPwC. KPMG has KymChat. Accenture merged its five service lines into one unit called "Reinvention Services" because the old model of selling strategy, then handing off to consulting, then handing off to technology, then handing off to operations is too slow and too expensive when execution cost has collapsed.
They're all converging on the same architecture: senior judgment at the top, AI-powered execution in the middle, and the base of the pyramid — gone.
But here's what none of them have solved: how does the obelisk actually coordinate? The pyramid had hierarchy for that. Partner told the engagement manager, engagement manager told the associates, associates did the work. The chain of command was the coordination mechanism. Remove the base and the middle, and the coordination problem doesn't disappear — it gets harder. Now you have partners who need to direct agents instead of people, and none of them have an operating model for that.
Deloitte's own research makes this visible. They found that while 30% of organizations are exploring agentic options and 38% are piloting solutions, only 11% are actually using agentic systems in production. 42% are still developing their strategy, and 35% have no strategy at all. Gartner predicts that over 40% of agentic AI projects will fail by 2027 because legacy systems can't support them.
The firms can see the obelisk. They can't build it. Because the missing piece isn't the AI — it's the operating model that tells humans and agents how to coordinate.
What Replaces Coordination by Hierarchy
The consulting pyramid solved two problems simultaneously. The obvious one: it scaled labor. The subtle one: it coordinated knowledge. The hierarchy wasn't just a management structure — it was a coordination mechanism. Information flowed up, decisions flowed down, and everyone knew their role because the org chart told them.
When you remove the pyramid, you lose both functions. AI can replace the labor. Nothing in the current playbook replaces the coordination.
This is the same problem that every alternative org model has tried to solve for seventy years. Mondragon, the Basque cooperative federation, built a three-in-one system — enterprise, bank, university — to create coordination without hierarchy. It worked at scale for decades, but participation decayed as the org grew because the governance overhead was too high. Valve eliminated all management and let employees choose their own projects. Former employees compared it to Lord of the Flies — flat without governance isn't flat, it's chaos. Holacracy replaced hierarchy with a formal constitution of roles and circles. Zappos adopted it, lost 30% of their employees, and discovered that process bureaucracy is just as expensive as management bureaucracy.
Every one of these models tried to solve coordination without hierarchy. Every one either created invisible hierarchy, unsustainable founder-dependency, or replaced management overhead with process overhead. The pattern is consistent: humans coordinating with humans hits a ceiling regardless of the governance model, because human coordination has irreducible cost. Meetings take time. Context transfers lose information. Alignment requires synchronous attention.
Agents change this equation. Not because they're smarter than humans at coordination — they're not. Because they can serve as the coordination substrate itself. An agent that reads the operating manual at runtime knows: what decisions it can make autonomously, what requires human approval, how to escalate, what the org's priorities are this week. It doesn't need a standup. It doesn't need a status update. It doesn't lose context between Monday and Wednesday.
The operating manual replaces the hierarchy as the coordination mechanism. It's a document — machine-readable, human-legible — that governs how every actor in the system (human or agent) coordinates. GitLab did a version of this for remote-first orgs: 2,000 pages of handbook that IS how the company works, not documentation about how it works. The agentic version goes further. The manual isn't just read by humans. Agents consume it at runtime and align their behavior to it. Update the manual, and every agent in the system updates its behavior on the next run.
This is what I've started calling SOUL.md for organizations — borrowing from the pattern in agentic systems where a governance document defines an agent's identity, values, and decision boundaries. Scale that from a single agent to an entire org, and you have the missing piece that none of the consulting firms have built: the coordination layer for the obelisk.
The Four Decision Classes
The operating manual's load-bearing structure is a classification system for decisions. Every decision in an agentic org falls into one of four classes:
Class A — Agent-Autonomous. The agent decides and executes. The human is notified after the fact. Scheduling, formatting, routine data lookups, status updates. Low stakes, high frequency, zero reason for a human to be in the loop.
Class B — Agent-Proposes, Human-Confirms. The agent does the work, presents a recommendation, and waits. Client communication drafts, budget allocations under a threshold, architectural decisions, anything published under someone's name. The agent does the thinking. The human does the judgment.
Class C — Human-Decides, Agent-Supports. The human makes the call. The agent provides context, options, and analysis but does not recommend. Strategic direction, pricing, legal matters, reputation risk. The agent's job is to surface the best available information, not to decide.
Class D — Human-Only. No agent involvement. Sensitive conversations, crisis communication, anything requiring judgment about another person's emotional state. Some decisions are human because they need to be human.
This is the concrete answer to "humans in the loop" — a phrase that has become one of the most hand-waved in the industry. Every company says they keep humans in the loop. Almost none can tell you which humans, for which decisions, with what governance, at what speed. The decision classes make it operational. Not a principle. A protocol.
When McKinsey's partners direct AI agents instead of associates, they need exactly this framework. Which outputs does the agent ship without review? Which require partner sign-off? How fast? What's the escalation path when the agent encounters ambiguity? Right now, every McKinsey team is figuring this out ad hoc, per project, per partner. The operating manual makes it systematic.
65% Say the Old Model Is Dead
An HFS Research report found that 65% of enterprises say traditional consulting models no longer provide enough value. Two-thirds of the market is already looking for something different. Not incrementally different — structurally different. They don't want a cheaper pyramid. They want a different shape entirely.
The firms that capture this shift won't be smaller versions of McKinsey. They'll be practitioners with operating manuals, agentic infrastructure, and the methodology to deploy both. The economics are stark: a McKinsey engagement runs $500,000 to $5 million. Smaller firms with agents and documented methodology are beginning to offer comparable rigor at a fraction of the cost and twice the speed — because the coordination overhead that inflates McKinsey's price to $500K doesn't exist in the obelisk.
This is already happening. Smaller, AI-driven consulting firms — often started by ex-consultants — are beginning to challenge the traditional dominance. But the ones that will win aren't just "McKinsey but cheaper." They're structurally different organizations that operate from an agentic operating manual rather than a hierarchical org chart.
Deloitte published a prediction that 15% of day-to-day work decisions will be made autonomously by agentic AI by 2028. That's conservative. In an org designed around the operating manual — with decision classes defined, coordination automated, and agents consuming governance documents at runtime — 15% is where you start, not where you end.
The Manual Is the Moat
The consulting industry spent seventy years building a moat out of information asymmetry. Clients didn't know what they didn't know, and firms charged handsomely to fill the gap. AI collapsed the information asymmetry. The new moat isn't knowledge — it's coordination methodology.
The firm that publishes its operating manual doesn't give away competitive advantage. It demonstrates it. GitLab's 2,000-page handbook didn't help competitors replicate GitLab. It proved that GitLab had solved remote coordination at a depth no one else had reached. The agentic operating manual works the same way. Publishing it says: we've solved the coordination problem for human-agent organizations, and here's the proof.
The pyramid is collapsing. What replaces it isn't a leaner pyramid or a flatter hierarchy or a fancier AI chatbot bolted onto the same business model. It's a fundamentally different organizational shape — the obelisk — governed by a fundamentally different coordination mechanism: the operating manual.
McKinsey is spending its 100th anniversary learning this lesson from the inside. The rest of the industry has a choice: restructure around the new model now, or spend the next decade managing decline. The SaaSpocalypse is the same dynamic running through the software layer simultaneously — AI agents replacing workflow-by-workflow what subscription software charged per seat for.
The operating manual is the first artifact. Everything else follows from it.