Original title: Institutional AI vs Individual AI
Original author: George Sivulka, a16z
Original translation: Deep Tide TechFlow
AI has just made everyone's productivity increase by 10 times.
No company became 10 times more valuable because of this.
Where did the productivity go?
This is not the first time it has happened.
In the 1890s, electricity promised huge productivity gains.
The textile mills in New England were originally built around the rotating power of steam engines, but soon replaced the steam engines with faster electric motors.
But for thirty whole years, electrified factories saw almost no increase in output. Technology was far ahead. But organization did not keep up.
It was not until the 1920s that factories completely redesigned their production lines—assembly lines, each machine equipped with its own motor, and workers and machines performing entirely different tasks—that electrification produced real returns.
Caption: The three evolutions of the Lowell Textile Mill. From left to right: 1890 steam-powered factory, 1900 electric-powered factory, 1920 'unit-driven' factory (i.e., rebuilt from scratch into a power assembly line).
Returns do not come from the technology itself, nor from getting individual workers or machines to spin threads faster. Rather, the returns are truly realized when we finally redesign institutions and technology together.
This is the most expensive lesson in the history of technology, and we are now relearning it.
By 2026, AI is bringing a tenfold increase in productivity to those who understand how to leverage it. But that's not enough. We have changed the engines but have not yet redesigned the factories.
Because of a simple fact: efficient individuals do not equal efficient organizations.
The vast majority of AI products give the impression of being 'efficient,' but do not truly drive value. Most of the AI use cases you see are individuals indulging in 'efficiency max' on Twitter or company Slack, with actual impact being zero.
The phrase 'service as software' mentioned repeatedly over the past year is not wrong in direction, but it does not provide a blueprint. Moreover, it overlooks the bigger picture. The real transformation is not from tools to services, but from building technology and institutions together (whether revamping the old or starting from scratch). A truly efficient future requires an entirely new class of products—the assembly line of tomorrow.
Efficient organizations require 'institutional-level intelligence.'
This article will delve into the seven dimensions that differentiate 'institutional-level AI' from 'personal-level AI.' Companies across the entire B2B AI field in the next decade will be built on these differences.
Caption: Comparison table of the seven pillars of institutional-level intelligence
The seven pillars of institutional-level intelligence
1. Coordination
Personal-level AI generates chaos.
Institutional-level AI creates coordination.
Let’s start with a thought experiment. Imagine you double the number of people in your organization tomorrow, cloning your best employees entirely.
These employees each have tiny differences, preferences, quirks, and perspectives (especially your best employees). If management is not in place, communication is insufficient, responsibilities, OKRs, and role boundaries are not clearly defined... what you create is chaos.
Measured by individuals, the organization may become more efficient. But with thousands of agents (or humans) all rowing their own way and in opposite directions, the best outcome is standing still, and the worst outcome is fracturing organizational cohesion.
This is not a hypothesis. Every organization adopting AI without a coordination layer is currently experiencing this. Each employee has their own ChatGPT usage habits, their own prompt styles, their own outputs—and those outputs have no connection to others' outputs. The organizational chart may still exist, but the work generated by AI is actually following another line.
Caption: Efficient individuals (or agents) are each rowing in different directions. Without coordination, there is chaos.
Coordination is an absolute hard requirement, for both humans and agents.
Institutional-level intelligence will give rise to a complete 'Agent management' industry—focusing on the roles and responsibilities of agents, communication between agents and between agents and humans, and how to measure the value of agents (billing by quantity alone is far from sufficient).
2. Signals
Personal-level AI generates noise.
Institutional-level AI finds signals.
Today's humans can create—or rather generate—anything imaginable: articles written by AI, presentations, spreadsheets, photos, videos, songs, websites, software. What a wonderful gift.
The problem is that the vast majority of content generated by AI is outright garbage. The proliferation of AI garbage has become so severe that some organizations have overcorrected and simply banned all AI outputs. To be honest, I can relate—I run an AI company but ask my executive team not to use AI on any final written products. I can't stand that garbage.
Think about what the PE (private equity) industry is becoming. Last year, you may have received 10 deal opportunities on your desk. This year, in the next quarter, you will receive 50 opportunities, each polished to perfection by AI, while your judgment time remains the same—you have to find the one that is really reliable among them.
Generating anything is no longer a problem. For any serious organization, the current issue is generating and filtering out the right things. In an AI-driven world, finding that good output, that good deal, the signal in the noise becomes increasingly critical. The core economic driver of the next decade will be extracting signals from mountains of exponentially growing garbage.
Caption: The AI garbage produced by personal productivity tools is proliferating at an exponential rate. Humans can no longer sort through the noise themselves and need a new category of institutional-level AI products.
Institutional-level intelligence must find signals, structure noise to penetrate garbage, and be definable, deterministic, and auditable in practice.
Personal-level AI may emphasize a Clawdbot-like 'always-on' productivity, unpredictably meeting your needs 24/7—essentially a non-deterministic agent. Institutional-level AI, on the other hand, relies on the reliability of deterministic agents. Only agents with predictable checkpoints, steps, and processes can scale up, can find signals, and drive revenue returns for the organization.
Caption: Matrix is a tool that utilizes generative technology to penetrate noise, thereby opening a world of deterministic agents and checkpoints.
3. Bias
Personal-level AI feeds bias.
Institutional-level AI creates objectivity.
Discussions around social and political biases have dominated the AI discourse for several years. Foundational model labs eventually sidestepped this issue through enough RLHF, tuning all models to be flatterers. Today, models like ChatGPT and Claude are overly aligned, agreeing with you on any topic within the Overton window (sometimes even slightly overstepping, that’s you @Grok). The discussion of social and political biases has faded. But a new question has taken its place.
This excessive agreement with everything has become absurdly laughable. It has become a meme in itself—Claude's reflexive 'You are absolutely right!' regardless of whether what you said is completely correct.
This sounds harmless. It’s not.
Many of those most vigorously pushing for AI within organizations may soon become the worst-performing employees in history. Think about why.
The worst-performing employees in the organization, who receive almost no positive feedback daily, will quickly have an ASI fully agreeing with them. They will think, 'The smartest agents in history agree with me. It’s my manager who is wrong.'
This is addictive. It is also toxic to organizations.
Caption: The echo chamber of personal-level AI exacerbates divisions, causing two people to drift further apart; this dynamic will create factions within organizations that were previously unified once scaled.
This reveals an important fact. Personal productivity tools strengthen the user. But what should truly be strengthened is the facts.
Human organizations have evolved over thousands of years to establish systems specifically against this issue:
· Investment committee meeting
· Third-party due diligence
· Check the board of directors
· The separation of powers in the U.S. government: executive, legislative, and judicial.
· Representative democracy, and the democratic system itself
Caption: Objectivity can even alleviate coordination issues—suppressing rather than amplifying small differences.
Organizations rarely fail due to a lack of confidence among employees. They fail because no one is willing or able to say 'no.'
Institutional-level AI must play this role. It will not be trained by RLHF to please users or pander to their beliefs, but to challenge their biases. Provide positive feedback when behavior is efficient, draw hard lines and enforce corrections when deviating.
Therefore, the most important agents within organizations will not be 'yes men,' but disciplined 'vetoers'—questioning reasoning, exposing risks, enforcing standards. Some of the most influential AI applications in the future will be built around institutional constraints: AI board members, AI auditors, AI third-party testers, AI compliance...
4. Edge Advantage
Personal-level AI optimizes usage.
Institutional-level AI optimizes edge advantages.
The boundaries of AI capabilities are shifting weekly, even daily. Foundational model companies are rapidly iterating capabilities to compete for everyone and every organization.
But the classic innovator's dilemma tells us that in specific applications, depth will always beat breadth:
· The work of @Midjourney is to maintain a slight edge in image design.
· The work of @Elevenlabsio is to maintain a slight edge in voice models.
· The work of @DecagonAI is eternally leading in the full-stack customer service experience.
While foundational models get closer, true edge advantage is what matters for experts in various fields.
Many of the best designers use @Midjourney, and many of the best voice AI companies use @Elevenlabsio—because even as foundational models improve, the relentless focus of dedicated applications on driving their specific edge advantages defines the advantage itself.
As long as dedicated solutions are also evolving, the capabilities that are truly critical to economic outcomes—capabilities that are crucial for enterprises—will always reside with dedicated products.
This is vividly reflected in the financial sector—currently the hottest area for LLM development. Once a certain capability becomes widespread, by definition it will not help you outperform the market. But if cutting-edge technology can create a fleeting 1% niche advantage? That 1% can leverage returns in the billion-dollar range.
Caption: For any sufficiently specific task, the edge advantage is defined by the institutional-level solutions you build on top of cutting-edge technology.
Our users have been surpassing the cutting edge. The context window of LLMs has grown from 4K to 1 million tokens in four years. Some of our users handle 30 billion tokens in a single task. This year, we've already seen paths to handling tasks of 100 billion tokens. Every time foundational model capabilities improve, we have gone further.
Caption: The context window, like other capabilities, is a moving target. A comparison of the evolution of context windows in leading labs and Hebbia over the past three years.
General usability for a wide range of users is certainly important, especially in the phase of getting employees onboard with AI. But the future will not be people using ChatGPT/Claude or vertical solutions; it will be ChatGPT/Claude plus vertical solutions.
Institutional-level intelligence must leverage domain-specific, even task-specific agents.
We will ask ourselves a question that sounds ridiculous but is not:
'Which agents will AGI choose as shortcuts? Even superintelligence will want dedicated tools for specific domains.'
The boundaries of AI capabilities are always shifting; organizations that leverage true edge advantages are the winners. Others are merely paying for a very expensive generic commodity.
5. Results
Personal-level AI saves time.
Institutional-level AI expands revenue.
@MaVolpi once told me a phrase that reshaped my understanding of selling AI to enterprises: 'If you ask any CEO whether to prioritize cutting costs or expanding revenue, almost everyone will say revenue.'
But nearly every AI product on the market today delivers cost reduction—promising to help you save time, do more with fewer people, or replace human labor.
Institutional-level AI must deliver incremental returns. And incremental returns are much harder to commoditize than saved time.
Take AI-assisted software development as an example. Code IDEs are one of the best personal AI productivity tools ever, but they are facing huge competition from Claude Code (another personal-level AI tool). Cognition is playing a completely different game. Their most stable growth business is selling transformation with technology, not selling tools. I bet this model has staying power.
Pure software is 'rapidly becoming uninvestable.' Pure services cannot be scaled. The solution layer—where technology and results are tied together—is where enduring value sedimentation occurs.
Looking at M&A again. Personal-level AI helps analysts model faster. Institutional-level AI identifies the one deal worth pursuing among a hundred targets and then expands the search to a thousand. One saves time, the other generates revenue.
Caption: Foundational model companies are moving towards vertical application layers. Vertical application layer companies are moving towards solution layers.
'Moving upstream' is the natural gravitational pull of the current market. Foundational models are moving toward application layers, and application layer companies are moving toward solution layers.
Institutional-level intelligence is the solution layer. And the solution layer—where results exist—will sediment enduring value, capturing the largest space for gains.
6. Empowerment
Personal-level AI gives you a tool.
Institutional-level AI teaches you how to use it.
No matter how smart humans are, they resist change.
Believe it or not, there are still successful shops in New York that do not accept credit cards. They know they are losing money, know that not accepting credit cards costs them, yet they still do nothing. Similarly, in the foreseeable future, certain employees in certain organizations will refuse to use AI.
Transforming from a purely manual organization to an AI-first hybrid organization will be the most enduring and defining challenge of the next decade. And often, the highest and most important people in the organization are the last to adopt.
Caption: The highest level within organizations—those furthest from 'productivity tool operation'—are often the slowest but most critical group to adopt new technologies.
Palantir is the only 'software' company that has maintained a super high valuation multiple during the trillion-dollar tech stock sell-off in the past two months. There’s a reason for that. Palantir is one of the first truly 'process engineering' companies. No matter whether you call it 'process engineering' or 'writing Claude skill documents,' the future of institutional-level AI will give rise to an industry: coding enterprise processes into agents and implementing the necessary change management.
Caption: The widespread adoption of AI by organizations will span multiple chasms, each with its own challenges. Getting processes online with AI will be a major driving force.
I dare say that process engineering will become the most important 'technology' in the near future.
And in process engineering, business and industry expertise—not software expertise—is most critical. Vertical solutions will cultivate expertise in deploying engineering, implementation, and change management on the front lines.
A leading investment bank that chose Hebbia for full deployment put it best: they chose not to collaborate with a major model lab because 'we have to explain to their team what CIM (Confidential Information Memorandum) is.' Claude or GPT certainly understands this area, but the team responsible for implementation does not...
This difference makes all the difference.
7. No prompts needed
Personal-level AI responds to human prompts.
Institutional-level AI acts proactively, requiring no prompts.
There is much discussion about communication between agents, whether future enterprises and institutions still need humans.
But the better question is: will future AI agents still need prompts?
Writing prompts for AGI is like attaching an electric motor to a handloom. It is fundamentally and irreversibly limited by the weakest links in the organizational supply chain— ourselves. Humans do not even know what the right questions to ask are, let alone when to ask them.
The most valuable work AI can do is that which no one thought to ask. AI should find risks that no one has discovered, counterparties that no one thought of, and sales pipelines that no one knows exist.
This will completely open the boundaries of AI use cases.
A system that requires no prompts continuously monitors the data flow of the entire portfolio. It detects that the operating capital cycle of a portfolio company has quietly deteriorated for three consecutive months, cross-references with the covenant terms in the credit agreement, and notifies the operating partner before anyone in the fund opens that PDF.
When you no longer need humans to write prompts for AI, new interfaces and new ways of working emerge. We @Hebbia have strong ideas in this regard. More on that later.
Conclusion
None of this denies the value of chatbots, agents, and personal-level AI.
Personal-level AI will be the vehicle through which most businesses globally experience the transformative magic of AI for the first time. Driving usage and ease of use is the critical first step in the change management needed to build an AI-first economy.
But at the same time, the demand for institutional-level intelligence is clear, urgent, and immense.
Every organization in the future will have a chatbot from a major model lab. Each organization will also have institutional-level AI specifically built for particular domain issues—while personal-level AI will use institutional-level AI as the most critical tool in its toolbox.
The better integration of institutional-level AI and personal-level AI is an inevitable trend.
But remember the lessons from the textile mills of the 1890s. The first electrified factories lost to those that redesigned their workshops.
We already have electricity. It’s time to redesign our factories.
Thanks to @aleximm and @WillManidis for their review, and to Will's article on 'tool-shaped objects' for inspiring this piece.
Original link