The Productivity Gap: Why Better Tools Have Not Made Better Institutions
A venture firm argues that AI has made individuals far more capable while their organisations stayed flat — and that the fix is structural, not another chatbot. The argument is sharp and worth reading. It is also opinion, from people who profit if you believe it.
Dr. Sven Jungmann
CEO
New England mills wired in electric motors in the 1890s and then watched their output sit almost flat for roughly thirty years. The motors worked. The gains did not arrive until the 1920s, when factories were finally rebuilt around electricity — a motor on each machine, the floor laid out for assembly rather than for a single steam shaft overhead. The new power source had not made the old factory faster. It had made the old factory obsolete, and the firms that prospered were the ones willing to tear up the layout.
This is the analogy at the centre of a recent essay on what artificial intelligence is doing inside organisations, and it is the part of the essay that is true regardless of who wrote it. The author's claim is that we are living in the equivalent of those flat thirty years: the tools are extraordinary, and the institutions around them have barely moved.
Individual AI, institutional AI
The essay is by George Sivulka, who runs the AI company Hebbia, published by the venture firm Andreessen Horowitz. Its organising distinction is between individual AI and institutional AI. Individual AI is a prompt and a reply — the assistant that helps one clinician draft one discharge letter faster. It is genuinely useful and changes nothing about how the organisation coordinates, decides, or remembers. Institutional AI is woven into the structure of the organisation: it works across roles and departments, acts on signals no single person could track, and is shaped by the institution's goals rather than by whoever happens to be typing.
Sivulka sets that distinction against a single sentence that does most of the rhetorical work: "AI just made every individual 10x more productive. No company became 10x more valuable as a result." Whether the real multiplier is ten or three is beside the point, and there is no measurement behind it either way. The gap it names — between a more capable person and an unchanged institution — is real, most of us have felt it, and stating it cleanly is the essay's genuine contribution.
A checklist dressed as a finding
The essay then lists seven properties it assigns to institutional intelligence: coordination, signal, bias, edge, outcomes, enablement, and acting unprompted. Each is built as a paired contrast — "Individual AI creates noise. Institutional AI finds signal"; "Individual AI saves time. Institutional AI scales revenue"; "Individual AI responds to human prompts. Institutional AI acts unprompted." As a checklist of what a standalone tool cannot do, the seven are useful. As a claim about the world they are unfalsifiable in the way every such framework is: no organisation is studied, no number is offered, there is no before and after. The list tells you where to look. It does not tell you what you will find when you get there.
It is worth being plain about the kind of document this is, because the answer changes how you read the seven labels. It is not a study. It carries no data, no method, no denominator. It is a well-argued opinion from a partner at a firm with very large AI investments, written by the chief executive of an AI company that sells precisely the kind of institutional system the essay concludes you should buy. None of that makes the argument wrong. It does mean we read it as argument rather than as evidence — and that we keep one hand on the wallet while admiring the reasoning.
The line most likely to surface on someone's strategy slide is the author's own: that even a superintelligence would want purpose-built tools for specific domains. Read coolly, it is a confident assertion about a system that does not yet exist, made by a person whose business is building purpose-built tools for specific domains. It may well be true. It is also exactly what you would say if you sold them.
“It is not a study. It is a well-argued opinion from people who profit if you act on it — which is a reason to read it carefully, not a reason to dismiss it.”
What survives the translation to a hospital
Nothing in the essay is about medicine; no hospital appears in it. The translation is therefore ours to make, and it should be made cautiously. What survives the scrutiny is modest and useful. A hospital that licenses a general chatbot and lets staff use it has acquired a tool, not changed a system. The harder work — linking a patient's record across episodes and departments, surfacing a population-level drift before any single clinician notices it, building decision support specific to one discipline rather than generic to all — is organisational, and most of its difficulty has nothing to do with the model. It is governance, data, workflow, accountability, and the regulatory care that any clinical software owes under the Medical Device Regulation (MDR).
That is the part worth keeping, and it is worth keeping whoever wrote it. The decision facing a health system is not which assistant to license. It is whether the institution is willing to be rebuilt around a capability — and whether that rebuild improves something a patient would recognise as better care. On the second question the essay is silent, because it was never asked it. The reader has to answer it alone, which is the more honest place to end than the author's.
Source: Sivulka G. Institutional AI vs Individual AI. Andreessen Horowitz, 12 March 2026. An opinion essay with no primary data, written by an AI-company chief executive for a venture firm with substantial AI investments; the productivity-lag analogy it rests on is a long-standing argument in economic history, and the framework it proposes is a lens, not evidence.
