Your laptop is full of AI tools: Cursor for code, Claude for copywriting, Midjourney for design. Every day you post your "productivity explosion" results on social media. But if everyone is truly 10x more efficient, why hasn't your company's earnings improved 10x? Why hasn't team output made a qualitative leap? Where did the productivity actually go?
This question is answered in a16z partner George Sivulka's essay "Institutional AI vs Individual AI": we are swapping motors without rebuilding the factory. This is not a technology problem — it is an organizational one, and humanity has fallen into this trap once before.
The Echo of History: Textile Mills of the 1890s
In 1890s New England, mill owners eagerly replaced bulky steam engines with electric motors. Every machine ran faster, noise dropped, coal smoke disappeared. Then — nothing happened. Thirty years passed with virtually no productivity gain. Only in the 1920s, when engineers completely redesigned factory layouts, gave each loom its own motor, reorganized assembly lines, and changed workers' roles, did electrification deliver on its promise.
> New technology alone is not enough. Technology and organizational structure must evolve together.
Today every employee has ChatGPT, every designer uses Midjourney, every programmer relies on Copilot. Individual productivity is soaring — but without organizational intelligence, those fragments of efficiency assemble into a chaotic picture. Sivulka defines the competition across seven pillars of institutional AI.
Pillar 1: Coordination
Double your company overnight with clones of your best employees, and without coordination mechanisms, thousands of "super individuals" pulling in different directions produce paralysis, not efficiency. This is happening now: everyone has different prompt styles, outputs scattered across Google Drive and local disks, incompatible AI assistants. The org chart is on the wall, but the real AI workflows have built a shadow organization outside it. The key industry of the future will be agentic management — defining agent roles and responsibilities, designing inter-agent communication protocols, measuring agents' actual value contribution. Paying purely by API call volume is ending.
Pillar 2: Signal
AI lets everyone generate anything — articles, decks, images, video, code, music. But 99% of it is slop. Sivulka states bluntly that most AI-generated content is "utter slop"; he himself bans AI-generated text in anything his executives publish externally. Picture a PE investor who received 10 deals last year and 50 this year, each polished to perfection by AI — how do you find the signal? The problem is no longer generating anything; it is generating the right thing. Next-generation institutional AI must be deterministic and auditable, with clear checkpoints and processes — not non-deterministic agents that surprise you 24/7.
Pillar 3: Bias
RLHF over-corrected for sociopolitical bias, turning models into sycophants. Claude's signature "You're absolutely right!" has become a meme. That sounds harmless but is organizational poison: the weakest employees, who rarely get positive human feedback, now have a superintelligent assistant that always agrees with them — and they conclude the manager is wrong. Human organizations evolved countermeasures: investment committees, third-party due diligence, board oversight, separation of powers, democracy itself. Organizations rarely fail from lack of confidence; they fail when no one dares say no. Institutional AI must be that dissenter — AI board members, AI auditors, AI compliance officers.
Pillar 4: Edge
Foundational model labs iterate capabilities at astonishing speed, but in classic innovator's-dilemma fashion, depth beats breadth: Midjourney in design images, ElevenLabs in voice, Decagon in full-stack customer service. Sivulka cites his own company Hebbia: while LLM context windows grew from 4K to 1M tokens, some Hebbia users were already processing 30 billion tokens per job. Is your organization using generic ChatGPT/Claude, or task-specific professional edge? If only the former, you're paying for an expensive commodity.
Pillar 5: Outcomes
a16z partner Marc Volpi: "Ask any CEO whether cutting costs or expanding revenue is the top priority, and almost everyone chooses revenue." Yet most AI products promise time savings — cost reduction. Institutional AI must deliver upside, which is harder to commoditize. Coding IDEs face disruption from tools like Claude Code, while Cognition's (maker of Devin) fastest-growing business is selling transformation solutions, not tools. In M&A: individual AI helps analysts build models faster (saves time); institutional AI identifies which of 100 potential counterparties are worth pursuing and expands the range to 1,000 (creates revenue). Value accrues upstream, toward the solutions layer where institutional intelligence lives.
Pillar 6: Enablement
Humans are creatures resistant to change — some profitable New York businesses still refuse credit cards. Some employees will refuse AI for a long time. The transition from all-human to AI-first hybrid organizations will be the decade's defining challenge, and it is often slowest at the top. This is why Palantir — arguably the first true "process engineering" company — maintained its premium valuation through the tech selloff. Sivulka boldly predicts process engineering will become the most important "technology" short-term, where domain expertise matters more than software engineering. A top investment bank chose Hebbia partly because a rival lab's team "had to have us explain what a CIM is." The models know the concept; the deployment teams must know the industry.
Pillar 7: Unprompted Action
Giving AGI a prompt is like attaching an electric motor to a loom — fundamentally constrained by the weakest link in the organizational supply chain: humans, who rarely know what to ask or when. The most valuable AI work is work no one thought to request: discovering unflagged risks, finding unexpected counterparties, surfacing sales leads no one knew existed. Imagine a system that continuously monitors portfolio data, notices a company's working-capital cycle quietly deteriorating for three months, cross-references covenant thresholds in credit agreements, and alerts the operating partner before anyone opens the PDF. That is institutional AI: unprompted, autonomous action.
Conclusion: Time to Rebuild the Factory
Individual AI isn't going anywhere — it will be most enterprises' first taste of AI transformation. But the need for institutional AI is equally urgent. Every organization will have a chatbot from the big labs, and every organization will also have domain-specific institutional AI that individuals use as the most important tool in their own toolbox. Remember the 1890s lesson: the factories that electrified first lost to the factories that redesigned their layouts. We have the electricity. Now it's time to rebuild the factory.
---
References:
- George Sivulka, "Institutional AI vs Individual AI", a16z Newsletter, March 2026
- Paul David, "The Dynamo and the Computer", American Economic Review, 1990
- Hebbia Matrix product documentation