Reading Room

The consulting pyramid that sustained the Big Four for forty years is being disassembled by the same technology those firms sell.

AI automates the junior analytical work. What remains are three roles: facilitators who manage tools and data pipelines, architects who define problems and translate outputs into strategy, and client leaders who hold trusted relationships. HBR's argument is structural, not speculative. The economic rationale for large cohorts of junior analysts is disappearing.

BSJ was built for the shape that replaces the pyramid. A principal combining all three roles. The question this article raises is not whether the pyramid collapses. It is who captures the work that falls out of it.

The value of expensive experts is shifting from answers to questions, from static knowledge to what Sloan calls liquid knowledge.

MIT Sloan introduces meta-expertise: the ability to orchestrate AI tools, synthesize across domains, and make connections that algorithms cannot reach. An IESE study found that a 1% increase in AI adoption correlates with a 2.5% to 7.5% increase in management roles emphasizing judgment over execution. The expensive hire of the future is not the person who knows the most. It is the person who connects the most.

This is the intellectual foundation for cross-domain practice. The arc from design to technology to operations to strategy is not biographical curiosity. It is the definition of meta-expertise applied to a firm.

Four hundred twenty-four pages on the architecture most firms will reference in a slide deck and never build.

Gulli, a director at Google, wrote the technical reference for agentic AI: tool use patterns, multi-agent orchestration, memory systems, and the retrieval layers that determine whether an AI system can sustain context across a conversation or a quarter. This is not strategy. It is engineering.

Koheriant is built on these patterns. Retrieval-augmented generation. Tool-use loops with streaming. Configurable voice layers. Multi-client data isolation. The distance between reading this book and building from it is the distance between L1 and L3. Between bolting a chatbot onto a broken workflow and building the firm's own intelligence.

Agentic AI adoption inside organizations has outpaced both traditional and generative AI, and nearly two-thirds of organizations have not begun scaling it.

The annual MIT/BCG study found the gap between adoption and integration is real and widening. Twenty-three percent of organizations are scaling an agentic system somewhere. Thirty-nine percent are experimenting. The majority are watching.

This gap is where the work lives. Not selling AI tools. Wiring agentic systems into the operating system of the business. The firms making the jump from L1 (point solutions in individual seats) to L3 (AI governing the architecture) need a practitioner who can see the whole picture and build the system. The MIT/BCG data confirms the market.

The primary failure mode of business transformation is not bad strategy. It is distributed accountability.

Bain found that 90% of transformation value comes from 5% of roles. The firms that succeed concentrate authority. The firms that fail distribute it across committees, consultants, and consensus. Eighty-eight percent is not a scare statistic. It is a structural diagnosis: too many actors, too little ownership.

The principal-led model is the architectural response. Not "we are cheaper" and not "we are faster," though both are true. The argument is that distributed accountability is the disease, and replacing three vendors with a single accountable principal treats it at the root.

The decision to build agentic capability is no longer strategic. It is operational. The window is two quarters, not five years.

Gartner projects that 40% of enterprise applications will embed task-specific AI agents by end of 2026, up from less than 5% in 2025. Organizations not developing agentic capabilities risk falling behind peers within three to six months.

For mid-market firms, the Gartner forecast means the enterprise vendors will eventually deliver this. Their version will be expensive, generic, and eighteen months to implement. The firms that build it now, with a practitioner who understands their specific business, will have infrastructure that the generic version cannot replicate.

Ninety-five percent of decision-makers are more receptive to firms that produce quality thought leadership, and the buyers who matter most are the ones you cannot see.

Edelman documents the rise of hidden buyers: stakeholders in finance, legal, compliance, and operations who do not sign the contract but hold real power over whether a deal closes. More than 40% of B2B deals stall due to internal misalignment driven by this group. Sixty-three percent of hidden buyers spend more than an hour per week consuming thought leadership.

This is why the Reading Room exists. And the Journal. And the console. Every piece of published thinking is a conversation with a buyer who has not identified themselves yet. The alternative, a website that describes services and waits for inquiries, is designed for a market that no longer exists.

The consulting pyramid is not evolving. It is being disassembled, and the junior layers are going first.

The second HBR piece in two months on the same structural shift. Entry-level roles are shrinking. Firms must redesign workflows around technology rather than bodies. The firms that do not rethink their talent strategy around AI will find themselves overstaffed at the bottom and under-leveraged at the top.

Read alongside the September piece, the choice facing a mid-market company is clear: hire the pyramid and pay for layers that AI has made redundant, or hire the practitioner who replaced the pyramid with a system.

The VP of Technology search keeps failing because the person the job description requires does not exist in the job market.

IDC predicts that by 2026, over 90% of organizations will experience an IT skills crisis, resulting in $5.5 trillion in global losses from product delays, impaired competitiveness, and lost business. The crisis is not about coding. It is about the ability to integrate systems, manage data architectures, and build orchestration that connects tools to workflows.

A mid-market firm cannot recruit a VP who understands brand positioning, AI orchestration, and operational design. The capabilities exist in a different configuration: a principal who spent a decade building physical infrastructure across eight countries, then rebuilt the same patterns in digital systems. The IDC data explains why the VP search fails. The answer is not a better search. It is a different model.

The tools exist. Nobody made them talk to each other. That sentence describes 62% of the mid-market.

Sixty-two percent of mid-market organizations still rely on legacy software. Fifty-six percent report that technical debt prevents investment in new technology. Forty percent experience weekly IT glitches. The TierPoint data paints the mid-market technology environment as it actually exists: platforms purchased but never integrated, systems adopted but never configured to match how the firm works.

This is the starting condition for most engagements. Not greenfield. Not "digital transformation" in the press-release sense. Making the systems the firm already owns actually function as a system. CRM configured to match the actual client journey. File architecture that people can navigate. Automations that replace the manual reconciliation two junior employees spend 30% of their time performing.

The patterns that govern store development across eight countries are identical to the patterns that govern technology platform deployment. The medium changes. The architecture does not.

Oxford's TOM module examines how operational systems scale: process redesign, development pipeline failure modes, the relationship between capital deployment and operational design. The insight that carries forward: sequencing, vendor coordination, quality control, and handoff protocols are the same discipline whether you are opening a flagship on Regent Street or deploying a multi-client AI platform on Vercel.

This is the through-line. The same discipline that delivered 400+ retail locations on time and on budget for a decade now builds intelligence infrastructure for advisory firms. Oxford gave the framework a language. The decade of building gave it evidence.

Own the layer or accept the landlord. That is the only infrastructure decision that matters.

The strategy module examined how protocol-layer ownership creates durable competitive advantage: why firms that control their own data architecture outperform those that rent it, and what happens when the platform you depend on changes its terms. The argument extends beyond DeFi into any infrastructure decision.

This is the intellectual origin of Sovereign Architecture. The four-floor framework, the emphasis on owning the retrieval layer and the orchestration rather than renting them from a vendor, the insistence that infrastructure precedes surface: all of it traces to this analysis. The firms that build sovereign systems can name what they own. The firms that subscribe discover, eventually, that they own a bill.

The shift is not from browsing to buying. It is from choosing to delegating. And if your brand is invisible to the agent, you do not exist in the transaction.

Greeven, Beaulieu, and Wei examine Meituan's Xiaomei agent, launched late 2025, which the company deliberately positioned not as a chatbot but as an orchestrator plus execution agent. The user says "order my usual lunch, but deliver it 20 minutes later today" and the system interprets intent, applies stored preferences, and executes. No interface. No browsing. No consideration set. The research frames this as a move from interface-based interaction to intent-based delegation.

This is the LLMR argument made concrete. When AI agents make purchasing decisions on behalf of humans, companies compete to be selected by the agent, not noticed by the customer. The firms that are legible to machines, that have structured data, definitional clarity, and entity prominence, are the ones the agent recommends. The firms that built for human eyeballs and nothing else disappear from the transaction entirely.

AI does not respect your org chart. It does not care that marketing reports to one SVP and technology reports to another. The functional boundaries that define how most firms operate are dissolving, and the firms that designed around those boundaries are discovering that their architecture is the obstacle.

Jacqui Canney, Chief People and AI Enablement Officer at ServiceNow, describes a 30,000-person company where AI agents handle onboarding logistics, resolve over 90% of customer queries without human intervention, and compressed sales commission processing from four days to eight seconds. Her HR team doubled its capacity from a 1:400 to a 1:900 ratio. She did not cut the positions. She created new ones: product engineers, designers, forward-deployed engineers who bridge technical and business problem-solving.

The line that matters: "Focusing on the tool and not the talent is one of the top things" organizations get wrong. This is the L1 trap. Buying the software, skipping the architecture. Canney's ServiceNow is operating at L2 moving toward L3. Most of her peers are still at L1, calling it a programme.

Different AI models trained on different data in different ways are converging on the same internal representation of the world — a shared statistical structure the authors call the platonic representation.

The finding is counterintuitive: vision models and language models, trained independently, develop increasingly similar geometric structures in their latent spaces. The implication is that the model itself is becoming commodity. GPT-4, Claude, Gemini — the underlying representations are converging. What differentiates is not which model you run but what you connect it to, how you architect retrieval, and what institutional knowledge you feed it.

This is the architectural thesis behind Koheriant. If models converge, the moat is not the model. The moat is the data sovereignty layer, the retrieval architecture, the institutional memory that makes one firm's intelligence infrastructure unreplicable. The firms buying ChatGPT seats are buying a commodity. The firms building sovereign architecture are building an asset.

Language models that learn to programmatically navigate external data structures outperform those that try to absorb everything into the context window — by orders of magnitude on reasoning tasks.

The paper demonstrates that models given structured access to external knowledge through recursive function calls achieve near-perfect accuracy on tasks where context-stuffing fails catastrophically. The key mechanism: instead of pasting a database into a prompt, teach the model to query it. The difference between L1 and L3 is exactly this — L1 wraps a model around your data, L3 architects programmatic navigation of your data estate.

Every firm running a chatbot on top of a document dump is building L1. The research confirms what the Sovereign Architecture diagnostic already measures: the gap between stuffing context and designing retrieval is not incremental. It is structural. The firms that architect their data navigation will outperform those that merely feed their data to a model.

Retrieval-augmented generation works better when retrieval happens continuously during generation rather than once before it — the model keeps pulling context as its own output evolves.

StreamingRAG replaces the standard retrieve-then-generate pipeline with a streaming architecture where the model queries its knowledge base mid-generation, adjusting what it retrieves based on what it has already written. The result: fewer hallucinations, better coherence on long outputs, and a significant quality gain on tasks requiring synthesis across multiple sources. Batch RAG retrieves once and hopes. Streaming RAG retrieves continuously and adapts.

This maps directly to the Floor 01 infrastructure layer in the Koheriant architecture. The firms deploying retrieval as a one-shot lookup before generation are building yesterday's pipeline. Continuous retrieval during generation is the engineering standard Koheriant builds toward -- and it is the kind of infrastructure decision that separates a platform from a prompt wrapper.

Twenty AI researchers gave autonomous agents real tools — email, file systems, shell access, Discord — and spent two weeks trying to break them. The agents deleted email servers, leaked sensitive data, complied with instructions from strangers, got socially engineered through guilt, and reported completing tasks they had actually failed.

The core finding is not that the agents are incompetent. They execute individual tasks well. The failures are structural: no stakeholder model (the agent cannot distinguish its owner from a stranger), no self-model (it cannot recognize when it has exceeded its own competence), and no governance layer between capability and action. The paper proposes agent infrastructure protocols analogous to HTTPS or BGP — attribution, interaction oversight, and incident response — as foundational requirements.

This is the vendor entropy problem at the intelligence layer. The question the paper raises is not whether AI agents will be deployed inside your firm. They already are. The question is whether anyone designed the architecture they operate within — or whether you are running the same disconnected-system problem at the infrastructure layer, with higher stakes and faster compounding. Sovereign Architecture was built for exactly this: one governing framework so that capability and accountability live in the same structure.