The technology industry is spending extraordinary sums building a new computational infrastructure. GPUs, data centers, high-speed networks and increasingly capable models are making machine intelligence available at a scale that would have been difficult to imagine only a few years ago.
The assumption underlying much of this investment is straightforward: more capable intelligence, applied more broadly, will create more economic value. I believe that assumption is correct. But it also leads to a problem that has received far less attention.
The industry is building abundant machine intelligence. The next constraint will increasingly be the architecture that intelligence is asked to operate against.
The First Age of Enterprise AI: Depth #
Most enterprise AI adoption today begins with the world as it already exists. Companies have employees, applications, databases, workflows and decades of accumulated technology. AI is introduced into that environment to make it operate better.
Agents can investigate exceptions, search across systems, generate reports, navigate applications, reconcile information, write code, communicate with customers and automate activities previously performed by people. The economic opportunity is enormous, and it creates enormous demand for tokens. Intelligence is being applied more deeply to existing activities, and in many cases the return on that intelligence is compelling.
But there is an important characteristic of this first stage: AI is a service to the operating model rather than a constituent of the operating model. The underlying enterprise remains largely unchanged.
The Legacy Dividend #
Legacy technology is frequently criticized, but it exists for a reason: it works. A mature financial system may contain thousands of capabilities accumulated over decades. Some may be inelegant. Some may depend upon batch processes, reconciliations, duplicated data and extraordinary operational complexity. But the features exist and organizations depend upon them.
AI agents offer an extraordinary opportunity to extract additional value from this installed base without replacing it. In fact, architectural complexity can initially create additional opportunities for AI. A human who previously needed to navigate five systems to understand an exception can be assisted—or perhaps replaced in that activity—by an intelligent agent capable of navigating all five. The organization receives substantial productivity improvements without confronting the risk and expense of replacing its core systems.
This is rational. It is also finite.
When Agents Begin to Encounter the Architecture #
As increasingly capable agents are applied to increasingly difficult problems, something interesting happens. The intelligence itself often ceases to be the limiting factor.
The agent can reason, call tools, understand language and navigate applications. But it is still operating against systems whose fundamental architecture was designed decades before machine intelligence became a participant. Additional intelligence is then increasingly consumed compensating for the architecture underneath it. More context is required, more systems must be traversed, more reconciliation is necessary and more exception handling appears. Human intervention may even be reintroduced when the economics or reliability of the agent deteriorate.
This can easily be interpreted as evidence that AI has reached its limit. I believe that is often the wrong conclusion.
Diminishing returns from AI agents often indicate that the constraint has moved from the intelligence layer to the underlying application architecture.
Putting the human back into the process may address the immediate problem. It does not remove the architectural constraint.
From Depth to Breadth #
This leads to what I believe will be the second major stage of enterprise AI.
Token demand today has extraordinary depth: increasingly capable intelligence is being applied to activities organizations already perform. The longer-term opportunity is breadth.
There is a specific mechanism worth naming here. Intelligence does not become an economically consumable resource simply because models get cheaper. It becomes economically consumable when it is applied against an architecture that does not force it to compensate for underlying complexity. The evidence that we are not yet in that world is visible today: in some enterprise deployments, the pendulum has begun swinging back toward humans as more cost-effective than the tokens required to navigate legacy technology. This is not a failure of intelligence. It is intelligence being asked to spend most of its cognitive budget on the wrong problem.
A coherently designed architecture consumes less depth of tokens per unit of value produced. That is what makes intelligence economically justifiable in the first place. And paradoxically, once that shift occurs, organizations begin to use more tokens, not fewer — because the tokens now generate continuous return as the organization discovers how an AI-embedded architecture can be exploited. Cheap intelligence against bad architecture produces cheap waste. Coherent architecture is what unlocks the compounding returns that make intelligence economically transformative rather than economically marginal.
Instead of asking, How can AI perform this existing activity more efficiently?, organizations will increasingly ask, What activities should we perform now that intelligence itself has become an economically consumable resource?
That distinction is subtle but profound.
An organization may eventually deploy machine intelligence not merely to automate work currently performed by people, but to perform valuable activities that no organization could economically perform with people. Intelligence can continuously investigate operational activity, simulate alternatives, interrogate state, examine every transaction rather than a selected sample, and identify questions that humans have not even thought to ask.
At that point AI is no longer simply improving the operating model. AI has become part of the operating model.
This also changes the economics of token consumption. The objective is not to maximize the number of tokens consumed, nor simply to replace the greatest possible number of people. The meaningful economic measure is the value of the cognition produced relative to the cost of producing it. Sometimes the right answer will be deterministic software, sometimes a human, and sometimes enormous quantities of machine intelligence. The opportunity lies in applying each where it creates the greatest value.
Architecture in a World of Abundant Intelligence #
This changes what we should demand from enterprise software. Systems historically evolved around interactions such as human to screen to application to database. Later, APIs allowed applications to interact directly with other applications. The first generation of enterprise AI adds agents to this existing structure, typically allowing them to interact with applications through APIs, tools and interfaces designed around systems that already exist.
But that will likely prove transitional.
The more consequential architectural question is: What should an enterprise system look like when humans, applications and intelligent agents are all first-class participants?
This does not mean replacing deterministic computation with probabilistic AI. Quite the opposite. Systems responsible for establishing authoritative truth may need to become more deterministic, transparent and reconstructable, while simultaneously becoming far more intelligible to machine intelligence.
The objective is not to have AI manufacture truth. The objective is to construct truth in a form AI can understand, interrogate and act upon.
Compute Cannot Repeal Causality #
Financial accounting provides a particularly revealing example. Modern computing encourages decomposition: break a large problem into smaller independent problems, distribute them across processors, execute them concurrently and aggregate the results. Massive computing infrastructure makes this increasingly attractive.
But accounting contains a deceptively simple problem:
An accounting event can cross any logical boundary that we assume can be processed in isolation.
Those boundaries can include portfolios, accounts, securities, currencies, legal entities, reporting structures and even time. The economic event does not care which boundaries are convenient for the technology architecture. Worse, the meaning of a subsequent event may depend upon the authoritative state established by the events preceding it.
This creates both cross-sectional and temporal dependencies. The correct architectural question therefore isn’t How do we decompose accounting? It is: What is actually safe to decompose?
Massive computational power does not eliminate these dependencies.
Compute cannot repeal causality.
Parallelism can be enormously valuable around authoritative financial state. Analytics, simulation, interrogation, reporting and countless other activities can exploit extraordinary computational resources. But establishing financial truth must first respect the economics and temporal integrity of the events that created it.
Domain semantics must precede computational optimization.
An architecture that respects this ordering treats events as first-class citizens, preserves temporal integrity as a structural property, and makes the authoritative construction of state legible to whatever intelligence operates on it. That is a different kind of architecture than what most enterprise systems currently provide.
The Coming Architectural Ceiling #
This suggests a different way to think about the enormous investment currently being made in AI infrastructure. The first generation of returns can come from making existing systems and existing activities dramatically more productive. That alone may justify extraordinary investment.
But eventually some organizations will encounter diminishing marginal returns from applying additional intelligence to architectures that were never designed for it. At that point the question changes.
It is no longer: How do we make the agent better at operating the system?
It becomes:
What would we build if the system itself were designed for an era of abundant machine intelligence?
That is a much larger transition.
Building for What Comes Next #
For the past several years I have been developing a financial architecture called Visibility, designed to operate within and exploit the modern trillion-dollar computing infrastructure now being built around us. Visibility did not begin as an attempt to predict the explosive evolution of AI. It began with a fundamental architectural question: whether financial books and records could be represented with radically greater transparency, determinism and reconstructability. As both Visibility and AI evolved, the convergence became increasingly apparent.
What has emerged is not simply a different way of implementing the same financial applications. It is an architectural foundation intended to make financial truth directly accessible to humans, applications and machine intelligence while preserving the deterministic integrity required of books and records.
That distinction matters.
A young architecture cannot reasonably compete with decades-old systems on accumulated feature count. Legacy technologies have spent decades answering thousands of requirements, however inelegantly some of those answers may have been implemented. In a feature-by-feature comparison, age itself becomes an enormous competitive advantage.
But that may be the wrong competition.
If machine intelligence is becoming an integral participant in the enterprise operating model, the more important question is not whether a new architecture reproduces every feature accumulated by the previous generation on day one. The question is:
Which architecture provides the better foundation for what comes next?
Visibility is ready for that question to move from architectural proposition to early adoption. The right early adopter does not need to abandon the systems that operate its business today. It needs to be willing to prove, against sufficiently difficult real-world financial problems, whether a fundamentally different architecture provides a better foundation for the intelligent operating model that is emerging.
The Next Constraint #
We are constructing infrastructure capable of delivering machine intelligence at extraordinary scale. That infrastructure will create tremendous value by improving what already exists. But success itself will expose the next constraint.
As intelligence becomes more capable, more available and ultimately less scarce, the architecture through which that intelligence operates becomes increasingly important. Some organizations will respond by continuing to improve the layer between intelligent agents and legacy technology. Others will begin reconsidering the layer underneath it. This architectural transition will unfold over years, but the organizations that begin now will have an important advantage: instead of pursuing incremental AI improvements whose value eventually diminishes against the constraints of legacy architecture, they can begin building an operating model in which each advance in machine intelligence expands what the underlying architecture makes possible. If machine intelligence is becoming an integral participant in how organizations operate, rather than merely a service used to improve existing operations, the organizations that recognize this transition early will be positioned to lead it.
The industry is building abundant machine intelligence. The next constraint will increasingly be the architecture that intelligence is asked to operate against.