Why AI Isn't Scaling in the Enterprise
The real constraint is economics, not intelligence
By Donovan Crewe · · Enterprise AI, AI Strategy, Context Engineering, System Architecture, Artificial Intelligence
Over the past two years, enterprise AI has largely been measured by one metric: adoption.
How many developers are using AI? How many lines of code are AI generated? Which model performs best? Every quarterly report seems to include another statistic demonstrating that AI has become embedded within software engineering.
These metrics tell us that organisations are using AI, but they tell us very little about whether they're using it effectively. More importantly, they ignore what I believe is becoming the biggest constraint on enterprise AI: not intelligence, but economics.
Anyone who spends time in engineering communities has seen the same demonstrations. A developer rewrites an authentication platform over a weekend. Another migrates an ageing monolith to a modern framework in a matter of days. Entire applications appear to materialise from little more than a well-crafted prompt. These stories create an understandable expectation among business leaders. If AI can produce these kinds of results for individuals, why aren't organisations seeing the same transformation across thousands of engineers?
The answer is uncomfortable because it has very little to do with the capability of today's models.
It has everything to do with what enterprises can afford.
The demonstrations that capture our attention are typically powered by AI that has been given everything it needs to succeed. The model has access to extensive documentation, large context windows, multiple rounds of reasoning and enough time to refine its understanding before generating an answer. It isn't simply writing code. It is building a mental model of the system before attempting to change it.
That process is incredibly powerful, but it is also incredibly expensive.
Every architectural document added to the context consumes tokens. Every design discussion, previous pull request, engineering standard and production incident that helps the model understand the problem adds computational cost. As reasoning models become more sophisticated, that cost increases further. For an individual developer, the expense is barely noticeable. For an enterprise supporting thousands of engineers, it quickly becomes one of the largest operational considerations in an AI strategy.
This is where many organisations unknowingly begin working against themselves.
Finance introduces sensible controls. Token budgets are established. Expensive reasoning models are restricted. Context windows are shortened. Conversation history is truncated. Developers are encouraged to start fresh sessions rather than maintaining long-running conversations. Each decision reduces operational spend, but each decision also removes a piece of the context that made the AI effective in the first place.
Imagine asking one of your most experienced architects to redesign a distributed system after reading only a fraction of the documentation. Or asking a senior engineer to forget every design decision they made yesterday before continuing their work this morning. No engineering leader would deliberately create those constraints for people, yet we routinely create them for AI because every additional token has a measurable cost.
The result is a pattern that many organisations misinterpret. AI adoption continues to rise, but productivity improvements begin to plateau. The common conclusion is that the technology has reached its limit. In reality, many organisations have simply reached the limit of what they are willing to spend on allowing AI to think.
I believe this is where our industry has started asking the wrong question.
Most organisations are asking how they can reduce AI costs.
AI-native organisations ask how they can reduce redundant thinking.
There is an important difference.
If every engineering task begins with explaining the architecture, uploading the same documentation, describing the same business rules and reconstructing months of previous decisions, then organisations are repeatedly paying for knowledge they already own. Every new conversation effectively hires a brilliant consultant and spends the first half of the meeting bringing them up to speed before any real work begins.
That isn't a model problem.
It's a systems problem.
The future of enterprise AI won't be defined solely by more capable frontier models. It will be defined by the systems organisations build around them. Persistent project state, organisational memory, architectural knowledge, engineering standards and business context should not be recreated every time an engineer opens a new conversation. They should become durable assets that accumulate over time.
This is what I believe AI-native engineering should optimise for.
Not simply intelligence, but Intelligence Efficiency.
Traditional engineering disciplines optimise for efficient use of CPU, memory, storage and network resources. AI introduces a new resource that deserves the same attention: reasoning. Every time a model reconstructs information it has already learned, it consumes computational budget without creating new value. Every time that understanding is preserved, the model can spend its reasoning budget solving the next problem instead of rediscovering the last one.
The implication is significant. Persistent state is no longer just a productivity feature. It becomes an economic strategy.
When AI already understands your architecture, previous implementation decisions, coding standards, deployment history and business terminology, every request becomes cheaper. Not because the model itself costs less, but because it no longer needs to rebuild its understanding from scratch. Organisations reduce token consumption while simultaneously improving the quality of the output. Those two outcomes are often viewed as competing priorities, but with the right architecture they become complementary.
This represents a fundamental shift in how we should think about enterprise AI. The competitive advantage won't necessarily belong to the organisations with the largest AI budgets or access to the newest frontier models. It will belong to those that build systems where intelligence compounds. Every project, every deployment, every incident and every architectural decision becomes part of an expanding organisational memory that makes the next task faster, cheaper and more informed than the last.
Perhaps this is why measuring AI maturity through adoption alone feels increasingly inadequate. Knowing that every developer has access to AI tells us very little about whether that AI retains understanding, whether it learns from previous work or whether the organisation has engineered its knowledge to persist beyond a single conversation. Those are the questions that will increasingly separate companies experimenting with AI from those fundamentally transforming how software is delivered.
Enterprise AI will always operate within economic constraints. Models will continue becoming more capable, but richer reasoning, larger contexts and increasingly autonomous agents will continue demanding more compute. That isn't a temporary limitation. It is a design constraint that every technology leader must embrace.
The organisations that thrive won't do so by making AI think less. They will succeed by ensuring it never has to think about the same thing twice.
That is the difference between deploying AI and becoming truly AI-native.