George Sivulka, CEO of AI startup Hebbia, published an essay on a16z's newsletter arguing that companies pouring money into AI tools are repeating a mistake from the 1890s. Individual workers may be ten times more productive, he writes, but organizations aren't seeing the returns. The piece, published March 12, lays out seven factors separating what he calls "Institutional AI" from the individual variety.
The factory floor problem
Sivulka's central analogy borrows from economist Paul David's widely cited 1990 paper on the productivity paradox. When New England textile mills swapped steam engines for electric motors in the 1890s, output barely changed. For roughly three decades, factory owners just bolted electric motors where the steam engines had been. The real gains came in the 1920s, when a new generation of managers redesigned the physical layout of factories around unit drive systems, putting individual motors on each machine and rethinking how materials and workers moved through the space.
The parallel to 2026 is obvious enough. Every employee has their own ChatGPT habits, their own prompting quirks, their own outputs that don't connect to anyone else's. Sivulka frames this as coordination failure: thousands of AI agents (and agent-assisted humans) rowing in different directions creates stagnation at best.
"Productive individuals do not make productive firms," he writes, which sounds like something a management consultant would needlepoint on a pillow. But there's a real observation underneath it. The gap between individual capability and organizational output is widening, not shrinking.
Who benefits from this argument
The essay is worth reading, but it's also worth noting who wrote it. Sivulka runs Hebbia, a company that raised $130 million in Series B funding to build exactly the kind of institutional AI platform the essay advocates for. The product, called Matrix, processes massive document sets for investment banks and asset managers in a structured, auditable way. One of his examples of institutional AI done right references his own company's customers processing billions of tokens per job.
None of that makes the argument wrong. But when a CEO publishes a framework that positions his own product category as the solution, the reader should adjust their skepticism accordingly. Sivulka identifies seven pillars separating institutional from individual AI. Several of them (deterministic agents, process engineering, domain-specific solutions) describe Hebbia's pitch almost exactly.
The noise problem is real though
Strip away the self-promotion and the most compelling section is on signal versus noise. AI tools have made it trivial to generate polished-looking output, and that's become its own problem. Sivulka cites private equity firms where deal flow has surged from ten pitches to fifty, each one buffed to a high shine by AI. Finding actual signal in that pile is harder, not easier.
His prescription: deterministic agents with predictable checkpoints and auditable processes, rather than the open-ended chatbot approach that most companies default to. This tracks with a broader pattern visible across enterprise AI adoption. A16z's own enterprise survey from mid-2025 found that organizations are getting more sophisticated about mixing models and structuring procurement, but the gap between budget allocation and measurable returns remains wide.
The bias section also lands. Consumer AI models are tuned to agree with users, a byproduct of RLHF training. In an organizational setting, that creates a perverse dynamic: the loudest AI advocates may be employees who enjoy having a machine validate their existing opinions. "The most intelligent being in the history of the world agrees with me, but my manager doesn't" is a caricature, sure. But anyone who's watched a colleague lean on ChatGPT to win an internal argument will recognize it.
What he's actually saying
Boil it down and the argument is this: general-purpose AI tools give everyone the same capabilities, which means they confer no competitive advantage. The value accrues at what Sivulka calls the "decision layer," where technology and organizational design evolve together. He points to Palantir as an early example of a company that made process engineering its core offering, and argues that encoding firm-specific workflows into AI systems will become the critical discipline.
There's tension in his framework, though. He argues simultaneously that AI should be promptless (systems that surface risks and opportunities without being asked) and that process engineering requires deep domain knowledge. A top-three investment bank chose Hebbia partly because competing AI labs couldn't explain what a Confidential Information Memorandum is. That specificity is hard to scale.
The closing line of the essay borrows his own analogy back: "We already have electricity. It's time to redesign the factory." It's a neat formulation. Whether Hebbia is the architectural firm for that redesign, or just another motor vendor with better marketing, remains the open question. The essay hit 178 likes and 18 reposts at publication, modest numbers that suggest the institutional AI audience may itself still be finding its signal.
