Is AI Optimization More Important Than Bigger Models?

Enterprises have become very good at building AI models. But what they have not yet solved is how to translate that intelligence into consistent business results.

The gap is no longer just about model capability. It is about what happens next: whether intelligence leads to action, whether action produces measurable outcomes, and whether the enterprise learns from those outcomes over time.

For Juan Morán, CEO and Co-founder of TuringDream, closing that gap requires more than adding AI to existing workflows. It means rethinking enterprise architecture, giving AI carefully governed autonomy, and measuring success by business performance rather than task-level productivity.

In this conversation, he explains why the next major AI breakthrough may not come from larger models but from systems that learn from experience, adapt as conditions change, and improve with each use.

Beyond better models: Why AI optimization must start at the enterprise level

Better models are necessary. But they are not sufficient. Morán argues that the bigger opportunity lies in redesigning the systems that connect intelligence to business performance. He explains that businesses must rethink how AI systems, people, data, and workflows work together to improve enterprise performance.

Has the industry become too focused on making AI models more capable while overlooking the systems required to make enterprises more intelligent? What needs to change in how leaders evaluate AI investments?

Morán: Absolutely – and the problem is particularly pronounced on the supply side. AI vendors have poured extraordinary resources into making models more capable. While paying far less attention to the systems that must turn those capabilities into enterprise-wide results. Imagine a motor-racing industry obsessed with improving engines while neglecting the chassis, transmission, steering, and brakes. A spectacular engine does not, by itself, make a winning car.

This is a familiar pattern in technological revolutions. We first improve a component, then discover that the real transformation requires redesigning the system around it. Today’s paradox is especially striking. We have extraordinarily sophisticated, locally intelligent models embedded in architectures largely orchestrated by conventional deterministic software. The surrounding systems have little capacity to learn how to coordinate, allocate, or improve the use of those models.

Enterprise leaders should therefore demand progress not just in model benchmarks. But in the architectures that integrate models into business operations-including agentic platforms, orchestration, governance, and critically, mechanisms for learning and optimization across workflows. The investment question must shift from ‘How intelligent is the model?’ to ‘How much more intelligent does the entire enterprise become?’

Many AI initiatives are considered successful when they improve productivity, accelerate tasks, or generate better outputs. Are these metrics insufficient to establish whether AI is genuinely improving enterprise performance?

Morán: They are insufficient, but the problem goes beyond measurement. Much of enterprise AI is designed, deployed and evaluated primarily as a cost-reduction technology.

We celebrate tasks completed faster, hours saved and lower operating expenses. Those can be real gains, but they are not the ultimate measure of business success. The bigger opportunity is competitive advantage: serving customers better, responding faster to changing conditions, making better coordinated decisions and discovering entirely new ways to operate.

If we optimize the wrong objective, apply AI mainly to isolated tasks and leave the surrounding organization unchanged, more flattering productivity metrics will not solve the problem. A local improvement can simply move a bottleneck elsewhere. Genuine enterprise performance requires systems capable of coordinating improvements across functions and learning from their impact on the business as a whole.

Leaders should still measure efficiency. But they should judge AI investments above all by durable outcomes: superior customer results, growth, adaptability, resilience, and sustainable competitive differentiation.

That change in metrics must be accompanied by a change in technology and in where AI is applied.

If AI models are becoming increasingly capable, why hasn’t enterprise performance improved at the same rate? What does that disconnect reveal about the industry’s assumptions about intelligence?

Morán: Just like in the 80s, when companies were putting computers on every desk while the systems, processes, and organizational structures were not yet redesigned to turn individual computing power into collective productivity, we are repeating the pattern with AI.

Models become dramatically more capable, and individual tasks can improve dramatically. Yet enterprise performance does not rise at the same rate. The mistake is to treat intelligence as a property of individual components. Rather than of the larger system in which they operate. A collection of intelligent pieces is not automatically an intelligent organization.

The eventual answer to the PC-era paradox was not merely a better PC. It was new software, networking and organizational architecture that connected individual gains. AI needs an equivalent technological leap.

It needs a new layer that coordinates models, humans, data, and workflows, observes outcomes, and learns how to improve the enterprise as a whole. That layer is still missing from mainstream enterprise AI architectures.

AI pilots often perform well in controlled environments. But they struggle when exposed to fragmented data, legacy infrastructure, exceptions, and real-world operational complexity. What are the most consequential assumptions that break when AI moves from demonstration to enterprise-scale deployment?

Morán: I would challenge the premise. Fragmented data, legacy systems and operational exceptions are difficult, but they are not the deepest reason pilots fail at scale. The fundamental problem is that today’s AI systems do not genuinely learn the enterprise in which they operate.

A frontier model’s original training data contains little of a company’s private internal reality. Once deployed, its weights are generally frozen. Companies then provide instructions, prompts and procedures that are themselves largely static. In a pilot, that can look impressive because the environment is constrained. In production, the real world keeps unfolding: customers behave unexpectedly, exceptions accumulate, priorities shift and business processes evolve. Yet neither the model nor the surrounding system necessarily improves from those experiences.

Imagine hiring an army of brilliant graduates from the world’s best universities, giving them an operating manual on day one and then preventing them from learning anything new. Their only way to ‘improve’ would be to reread the manual. That is surprisingly close to how much enterprise AI is deployed today.

The missing capability is not another integration patch; it is continuous learning from actual operations and outcomes.

Continuous learning: The engine of AI optimization

AI systems must do more than follow instructions. To enable continuous AI optimization, they need to learn from operational feedback, respond to changing conditions, and refine their decisions as business needs evolve.

Business environments are rarely static. Customer expectations, regulations, market conditions, and organizational priorities change over time. How can a system continuously improve without becoming overfitted to historical patterns or optimizing yesterday’s objectives?

Morán: The answer is to build systems that learn from the present, rather than merely execute policies derived from the past. We already have a powerful family of techniques for doing this: reinforcement learning built around world models. These systems can represent an evolving environment, explore the consequences of possible actions and update their decisions as new evidence arrives.

The key architectural shift is to combine those learning and planning capabilities with the generative and reasoning strengths of large language models. In such a system, a world-model-based learning layer continuously evaluates what works, which objectives matter and how the environment is changing; LLMs help interpret information and carry out the actions selected within defined permissions. Objectives must also be revisable, with human governance setting constraints and priorities.

This is fundamentally different from deploying an LLM with a static prompt. And hoping that more context will somehow produce organizational learning. Avoiding overfitting requires ongoing feedback, testing against current conditions and the ability to revise both policies and goals. I see that as the foundation of a second wave of enterprise AI.

How should enterprises design feedback mechanisms when outcomes are delayed, influenced by multiple variables, or difficult to attribute to a single decision?

Morán: Delayed rewards and uncertain attribution are not unusual edge cases.

They are central problems in sequential decision-making, and reinforcement learning has addressed them for decades. A world-model-based reinforcement learning system can represent many interacting variables, estimate how actions affect future states and learn from rewards that arrive long after the original decision.

For enterprises, that means designing feedback around persistent business processes and their outcomes, not just evaluating each isolated model response. The system needs a history of actions and decisions, an evolving representation of the relevant business entities, explicit objectives and a way to connect delayed results to earlier choices. It should account for uncertainty and competing explanations rather than pretending that every outcome has a single, obvious cause.

LLMs are useful components, but by themselves, they are not an adequate framework for this kind of long-horizon optimization. The opportunity is to combine LLMs with world models, reinforcement learning and a runtime that records what happened. Feedback then becomes part of the system’s operation, rather than a retrospective reporting exercise.

Granting AI autonomy without losing control

As AI moves from answering questions to taking consequential actions, enterprises must ensure that AI optimization does not come at the expense of security, accountability or human oversight.

What level of operational autonomy can enterprises safely grant AI?

Morán: There is no single safe level of AI autonomy. Asking how much autonomy an enterprise should grant is like asking at what speed a racing driver should take every corner of a circuit. Without an accelerator and brakes for each corner, any universal speed will be either unnecessarily slow or dangerously fast.

What matters is a precise, understandable and enforceable system of permissions. For every AI actor, action and context, an enterprise must be able to define what can be read, changed, initiated, approved or delegated; humans must be able to inspect and audit those boundaries. Autonomy should be granular and conditional rather than a binary choice between ‘human in the loop’ and ‘fully autonomous.’

Crucially, AI should not simply inherit the working environment and permissions designed for human employees. Models need their own technological substrate: a secure virtual machine and runtime designed for AI, with native identity, access controls, observability and execution limits. Only then can organizations grant substantial operational autonomy where appropriate while restricting it precisely where risk demands.

The question is not whether to trust AI in the abstract, but whether the system makes its freedom of action governable.

The human advantage: Intellectual flexibility in an AI-driven world

As machines take on more analytical and operational responsibilities, the enduring challenge for leaders may be less about protecting a fixed set of human tasks and more about keeping their own thinking adaptable.

As AI takes on more analytical and decision-making responsibilities, what skills should the next generation of technology leaders develop that cannot be easily delegated to intelligent systems?

Morán: Adaptability and intellectual flexibility.

I would go further: the most important leadership capability may be the deliberate preservation and development of our capacity to change how we think. The human brain is extraordinarily plastic, yet as we accumulate experience, we also accumulate habits, assumptions, and intellectual comfort zones. Expertise can gradually become an obstacle to seeing a genuinely new possibility.

In a period of repeated technological upheaval, leaders need to cultivate divergent thinking: the willingness to consider alternatives that initially look implausible, question established categories and revise their own convictions when evidence changes. That means remaining curious, experimenting and resisting the instinct to defend yesterday’s successful mental models.

I would not frame the objective as finding one permanently ‘human-only’ task that machines can never perform. The boundary of what AI can do will keep moving. The enduring leadership skill is the ability to adapt to that movement-to keep learning, to rethink the problem and to remain intellectually flexible throughout an entire career.

The future: Rise of self-optimizing enterprises

The next stage of enterprise AI may extend far beyond today’s copilots, chatbots and agents. Morán believes the decisive shift will come when AI becomes part of how businesses learn, make decisions and improve their operations continuously.

Looking ahead five to ten years, how do you envision the relationship between AI and the enterprise evolving?

Morán: I expect at least one-and quite possibly two or three-major technological revolutions over the next five to ten years.

We tend to imagine that the latest breakthrough defines the road ahead. After ChatGPT, it is tempting to assume that the future is essentially more powerful chatbots and agents. That would be like looking at Yahoo, AltaVista and Infoseek in the early web era and assuming internet search had reached its final form-just before Google changed the rules.

One major shift will be from LLM-centric applications to integrated systems that combine language models with world-model-based reinforcement learning, persistent operational state and continuous optimization. Today’s models are extraordinary components, but their limitations as stand-alone enterprise systems are substantial. The architecture around them will matter as much as, and perhaps more than, incremental advances in model capability.

The ultimate transformation is that an enterprise becomes a self-optimizing machine. Every operation generates information about what works; that information improves the next decision; the organization becomes more effective simply by operating.

AI will no longer be an external productivity tool added to a business. It will become part of how the business continually learns and evolves.

Having spent more than three decades examining how technological shifts reshape business and society, what lessons from previous technology transitions do you believe today’s AI leaders most need to understand?

Morán: The most important lesson is that the pioneers of the first phase of a technological revolution are not necessarily the winners of its second phase.

The history of technology is full of transitions in which a new architecture displaced the companies that seemed to define the market: Google changed internet search after the early portals and search engines; the iPhone changed the mobile industry after Nokia’s era of dominance.

The reason is that the first generation often solves the most visible part of a problem, while the next generation reorganizes the whole system. By the time that architectural shift is obvious to everyone, much of the competitive advantage has already been captured.

Leaders should resist the comforting assumption that the truly disruptive changes are behind us. They are not. The greatest opportunities are often still ahead-and frequently begin at the margins, where today’s leaders are not looking.

Competitive advantage comes from recognizing those emerging shifts early and acting before they become conventional wisdom, not from adopting a revolution once it has already become the industry standard.

Mentorship: Looking beyond today’s AI optimization paradigm

The next generation of leaders must look beyond dominant AI approaches and remain open to ideas that challenge conventional thinking. The competitive advantage lies in recognizing the next technological shift before it becomes mainstream.

If you could offer one piece of advice to the next generation of leaders preparing to lead enterprises through the next decade of technological change, what would it be – and what conventional wisdom would you urge them to question?

Morán: Question the LLM monoculture. Large language models are an astonishing breakthrough, but treating them as the only meaningful path forward is a dangerous form of tunnel vision. The history of technology shows that the next decisive innovation often arrives not from extending the dominant approach, but from combining it with ideas, disciplines or architectures that the mainstream has overlooked.

My advice is simple: keep looking at the margins of the road you think you are traveling. Study what is happening outside the current consensus-in reinforcement learning, world models, systems architecture and other approaches to intelligence. Ask not only how to make today’s tools better, but what entirely different system might make their limitations irrelevant.

For the next generation of leaders, competitive advantage will not come from following the crowd more efficiently. It will come from noticing, earlier than others, when the road itself is about to change.

In brief:

For CTOs and enterprise leaders, the mandate is clear: do not just ask how AI can make existing operations faster or cheaper. Ask how it can learn from every decision, adapt to meaningful change, and improve its performance over time. The companies that get this right will not merely deploy AI more effectively.

They will build a new kind of organization – one where intelligence, action, and outcomes form a continuous loop, and where competitive advantage compounds with every decision.