Why the AI Boom Needs More Than Better Models: It Needs Better Ways to Change the World

Apr 29, 2026

10 min read

3 views

68%

The Strange Problem at the Heart of the AI Moment

What if the biggest thing standing between AI and world-changing impact is not intelligence, but institutional friction?

That sounds backwards. The story we keep telling ourselves is that once models become capable enough, transformation follows automatically: better tools, cheaper labor, more output, fewer bottlenecks. But the reality is more stubborn. AI may be astonishing at generating text, code, images, and plans, yet still fail to produce broad productivity gains because real organizations are messy, expensive, and full of human oversight. A system that can draft a perfect memo is not the same as a system that can safely run a hospital, a school district, or a government agency.

This creates a deeper tension that matters far beyond AI itself. We are living through a period where many people are trying to pour exponentially improving technology into systems that were built for linear change. That mismatch explains both the hype and the disappointment. It also reveals a more interesting question than whether AI is overvalued: How do you identify the kinds of interventions that can actually move complex systems?

The answer is not just “build more powerful models.” It is to understand where leverage truly lives: in technology, yes, but also in institutions, incentives, norms, and the fragile places where one small change can cascade through a large system.

The real bottleneck is rarely intelligence in the abstract. It is the ability to convert intelligence into durable change.

Why Capability Does Not Automatically Become Impact

There is a seductive assumption behind every technology boom: if a tool can do something humans can do, then businesses, governments, and societies will naturally adopt it. But adoption is not the same as substitution. A tool can be impressive and still fail to matter economically if it creates as much work as it removes.

Consider the familiar office example. An AI system can write a first draft of a contract in seconds. That sounds like replacement. But if the lawyer still has to verify every clause, check for hallucinated citations, catch edge cases, reconcile risk, and maintain responsibility for the final document, then the workflow changes less than the headlines suggest. The task shifts from creation to supervision. Productivity gains become partial, delayed, and uneven.

This is true in almost every complex domain. A clinic cannot simply “automate” diagnosis if the cost of mistakes is catastrophic and the workflow requires accountability. A city government cannot adopt AI for permitting if the underlying process is already tangled, politically sensitive, and full of legal constraints. Even when the model is competent, the surrounding system may be the real source of cost.

That is why so many technologies plateau at the level of novelty. They produce demonstrations, not transformations. They impress individuals, but organizations are not individuals. Organizations are collections of incentives, liability, coordination costs, and legacy structures. A new tool only becomes a breakthrough when it reduces total system friction, not just task difficulty.

This is the first key insight: capability is not leverage unless it compresses an entire workflow, not merely one step inside it.

The Hidden Geometry of Complex Systems

If you want to understand where change can actually happen, stop asking only, “What can this technology do?” Ask instead, “Where is the system most vulnerable to a small intervention?”

Complex systems have a geometry. Some parts are thick and resistant, like concrete walls. Others are thin, like load-bearing joints. Most people focus on the largest visible structures, but the highest leverage often lies in the narrow seams: a procurement rule, a distribution bottleneck, a licensing requirement, a data standard, a social norm, a single interface between institutions.

This is why many of the most effective interventions do not look dramatic at first. A new educational format can outperform a flood of expensive devices. A change in default settings can outperform endless persuasion. A better coordination mechanism can outperform a larger budget. These are not “smaller” changes in importance. They are smaller in surface area, but larger in system impact.

A useful way to think about this is to distinguish between thick interventions and thin interventions:

Thick interventions try to overpower a system directly. They add more money, more labor, more features, more policy detail.
Thin interventions change the system’s rules, interfaces, or incentives so that the same resources start working differently.

AI often gets trapped in the thick category. We celebrate raw model capability, but the bottleneck is often integration, oversight, trust, and adoption. That means the highest-value AI applications may not be the flashiest. They may be the ones that quietly reduce coordination costs, standardize decisions, or make complex tasks legible enough for institutions to act on them.

This is why “AI will replace jobs” is usually the wrong frame. The more precise question is: Which parts of a system can AI thin out, so the remaining human effort becomes more valuable rather than merely busier?

The Bubble Question Is Really a Mispricing of Transformation

If a technology is overhyped, the usual explanation is that investors expect too much too soon. That is probably part of the story here. But there is a more interesting possibility: the market may be mispricing not just the speed of adoption, but the kind of change that matters.

A lot of AI investment assumes that intelligence itself is enough. If models keep improving, they will naturally unlock enormous economic value. Yet the evidence from real workplaces suggests a more complicated reality. Even when AI can perform a task, businesses may not save much unless they can remove oversight, simplify workflows, and trust the outputs at scale. That means value is constrained by organizational redesign, not just model quality.

This creates a familiar boom risk. The excitement is centered on visible capability, while the hard part is invisible institutional adaptation. If the visible layer gets priced as if it automatically implies the invisible layer, the result is a bubble. Not because the technology is fake, but because the pathway from demos to impact is longer and narrower than expected.

The lesson here extends beyond AI markets. We often overinvest in what is easiest to measure and underinvest in what actually changes systems. We fund the tool, not the adoption architecture. We celebrate the breakthrough, but neglect the policy, training, governance, and operational redesign needed to make the breakthrough matter.

A model that can do 80 percent of a task but requires 200 percent of the oversight is not transformative. It is impressive overhead. The mistake is treating apparent competence as economic substitution. In reality, substitution only happens when a system can absorb risk cheaply enough to restructure itself.

The market often prices the machine, but transformation depends on the scaffolding around the machine.

The Real Opportunity: Grants, Innovation, and Institutional Design

This is where a different kind of ambition becomes visible. If novel technology alone cannot reliably produce system-level change, then the highest leverage work is often the work of designing better pathways for change itself.

That means seeking interventions that combine three qualities:

A genuine mechanism for leverage
A plausible route through existing institutions
A feedback loop that makes success scalable

Think of a charity that does not merely fund more of the same, but tries a new financing model, a new coordination structure, or a new policy template. Or a civic technology project that does not just build software, but changes how agencies share information. Or an educational reform that does not merely add content, but changes how schools allocate attention and accountability.

The most promising projects in complex systems are often the ones that make systems easier to learn from themselves. They create a better signal. They reduce ambiguity. They lower the cost of coordination. They change defaults. They turn repeated improvisation into reusable structure.

AI can play an important role here, but not as a standalone savior. Its best use may be as a system amplifier: a tool for making institutions more legible, more adaptive, and more capable of targeted action. That might mean helping regulators process comments, helping nonprofits identify which interventions deserve more funding, helping researchers synthesize evidence faster, or helping local governments detect bottlenecks before they become crises.

In other words, the biggest prize may not be automating isolated tasks. It may be building the machinery for better institutional evolution.

This is why the question of “what kind of charity, startup, or policy project matters most?” is really the same question as “what kind of AI will matter most?” The answer in both cases is: the kind that changes the system, not just the output.

A Mental Model for Spotting Real Leverage

Here is a practical framework for evaluating whether an intervention is likely to matter in a complex system.

The Four Friction Test

Ask whether the idea reduces one or more of these frictions:

Cognitive friction: Do people understand what to do better?
Coordination friction: Do people and institutions align more easily?
Verification friction: Is it easier to know what is true or safe?
Transition friction: Can the system change without massive disruption?

Many shiny technologies reduce cognitive friction but not the others. That is why they create excitement but limited adoption. A good intervention reduces all four, or at least enough of them that the system can move.

Now apply this to AI. AI may drastically reduce cognitive friction by generating drafts, summaries, and suggestions. But if it increases verification friction because outputs need extensive checking, the total effect may be modest. If it lowers coordination friction by helping different teams share a common model of a problem, then it may be much more valuable than the raw task benchmarks suggest.

This framework is useful outside AI too. A policy reform that is technically elegant but hard to administer will fail. A nonprofit strategy that is morally compelling but hard to evaluate will stall. A startup that saves time but increases trust risk will face resistance. Real leverage is not about the beauty of the idea in isolation. It is about how much total friction it removes from the system.

What To Do With This Insight

The temptation in moments like this is to become either cynical or euphoric. Cynical people say the hype is fake and the market will collapse. Euphoric people say the model will eventually solve everything. Both miss the deeper point.

The point is that impact is an engineering problem in the broadest sense. It requires understanding not only what is possible, but what is absorbable by a real system. The best builders, investors, donors, and policymakers should not ask whether a technology is magical. They should ask where it changes incentives, reduces friction, and unlocks second-order effects.

That is why the most valuable opportunities may be hiding in plain sight. Not in the loudest demo, but in the quiet redesign of a workflow. Not in replacing people, but in making institutions capable of doing what they already claim to want. Not in maximizing raw intelligence, but in turning intelligence into coordination.

If there is a bubble, it may be a bubble of misplaced expectations about where change begins. The correction would not mean AI is useless. It would mean we finally stop confusing possibility with deployment, and deployment with systemic transformation.

Key Takeaways

Do not confuse model capability with real-world leverage. A tool matters only if it reduces the total friction of a workflow, not just one task.
Look for thin interventions. The highest-impact changes often alter rules, defaults, interfaces, or incentives rather than adding more resources.
Evaluate AI by its effect on oversight. If a system creates as much checking as it removes labor, the productivity gain will be small.
Think in systems, not products. The most valuable uses of AI may be the ones that make institutions more legible, coordinated, and adaptive.
Use the Four Friction Test. Ask whether an idea reduces cognitive, coordination, verification, and transition friction.

The Reframing

We are not just trying to build smarter machines. We are trying to discover which kinds of change can actually survive contact with the world.

That is a much harder challenge, but also a more interesting one. It suggests that the future will not belong merely to those who build the most powerful tools. It will belong to those who understand how to design systems, institutions, and interventions that can absorb power without breaking.

In that sense, the real question is not whether AI will change everything. It is whether we will learn to build the kinds of structures that can turn intelligence into lasting civilization-scale improvement.