Leading in the age of AI

Every day, we come across radically opposing predictions about the impact of artificial intelligence. Some believe it will eliminate a large number of jobs and profoundly transform society. Others warn that the technology is overrated, prone to hallucinations and far less capable than the headlines suggest. Between these two extremes lies a confusing landscape for leaders trying to understand what AI really means for their organizations.

To navigate this debate, it is worth taking a step back and separating hype from reality. Artificial intelligence is not exactly new. In fact, many forms of what we might call “large-scale AI” have existed for decades. Robotics, computer vision, automation and expert systems have been used in manufacturing and engineering for more than 40 years. These technologies operate in the physical world (machines that move, grasp, inspect and process real materials). We can think of this as the “torque side” of AI, as I like to call it: the domain in which machines interact directly with the physical environment.

What took the world by surprise in recent years was something different: the rapid emergence of large language models, or LLMs. When tools such as ChatGPT became widely accessible, millions of people suddenly came into contact with a new kind of artificial intelligence. Instead of controlling robots or industrial equipment, these systems began to interact through language, producing texts, explanations, summaries and programming code.

This distinction is important. Traditional AI dealt with physical actions and mechanical systems. Generative AI operates in a completely different world: the world of words and language patterns. Understanding this difference helps explain both the enthusiasm and the confusion surrounding modern AI.

The problem of hype versus reality

Part of the challenge is that most people interact with AI only at its most basic level. Many users, for example, come into contact with large language models only through simple chat interfaces. At this level, the results can seem unpredictable. Sometimes the answers are impressive; at other times, they are incorrect, incomplete or strangely out of context.

This inconsistency occurs because language models do not actually "think". They generate answers by predicting the most likely next word, based on patterns present in the data used to train them — mainly large volumes of content available on the internet. Some researchers describe them as “stochastic parrots”: systems that simulate reasoning but have no real understanding of the world.

Despite this limitation, language models are surprisingly capable. When used properly, they can perform many tasks that previously required considerable human effort. The key is learning to go beyond the most basic interactions.

In practice, there are several levels of AI use. At the most elementary level (the chatbot interface), almost anyone can take part. However, as we move up the scale of maturity and capability, the number of people able to use these tools effectively decreases drastically.

The next level involves prompt engineering, in which users carefully craft instructions to guide the model's responses. Above that is interaction through APIs, in which developers connect directly to AI models using programming tools such as Python. This makes it possible to integrate AI into applications and automate more complex workflows. Beyond this level, there are advanced techniques, such as retrieval-augmented generation, known by the acronym RAG, which connects models to external data sources, and systems fine-tuned to perform specific tasks.

Although these advanced approaches can produce impressive results, only a very small share of users operate at these levels today. Most remain at the basic stage of interaction. As a consequence, the real capabilities of AI often seem exaggerated or are poorly understood.

Where AI excels... and where it struggles

To understand how leaders should approach AI, it helps to consider the differences between the way humans and machines carry out their thinking processes. As Stanford researcher Balaji Srinivasan pointed out, humans by nature think end to end: faced with a problem, we are able to define the situation, develop hypotheses, carry out solutions and reflect on the results. In lean thinking, this cycle is often represented by PDCA (plan, do, check and act). Language models work differently: they perform best in the intermediate stages of the thinking process. In Srinivasan's words, they are “middle-to-middle thinkers”.

At the beginning of the problem-solving cycle, the Plan stage, human beings excel. We frame the problem, gather the context and decide which direction to take. AI struggles at this stage because it does not have a concrete understanding of the real world. Without clear guidance, it can easily generate irrelevant or misleading suggestions.

At the end of the cycle, the Act stage, humans also predominate. This phase involves judgment, reflection and the ability to adapt strategies based on experience. Once again, these are areas in which AI has clear limitations.

However, in the intermediate stages (analysis, synthesis and structured reasoning), AI can be extremely powerful. Once the task is clearly defined, language models can process large volumes of information, produce summaries and suggest possible solutions.

This pattern becomes even clearer when we look at different types of problem solving, a topic I addressed in my book Four Types of Problems. AI performs best on procedural tasks, in which both the problem and the solution are already well understood. In lean terms, this resembles Type 1 problem solving, which involves quick action and response to abnormal conditions. These are tasks such as writing routine code, summarizing documents or troubleshooting known problems.

For diagnostic problems (Type 2), in which the situation needs to be investigated, AI can help organize information and propose hypotheses. However, humans still need to verify the facts and determine the root cause.

In designing a future state (Type 3), AI becomes less reliable. Defining a compelling target state requires observation, imagination and an understanding of the context. These capabilities remain essentially human.

Finally, in the field of true innovation (Type 4), AI faces even greater difficulties. Although it can recombine existing ideas in interesting ways, it rarely produces genuinely original breakthroughs without strong human guidance.

Technology alone is not enough

For leaders considering the adoption of AI (and, let's be honest, who isn't?), the most important lesson may come from lean thinking itself:

Technology alone does not create transformation.

A simple framework can sum up this idea: Impact = Technology × Behavior × Management.

If any one of these elements is missing, the total impact will be reduced to zero.

Organizations have already learned this lesson many times. Enterprise systems, such as ERP platforms, promised significant improvements but often failed because companies neglected the behavioral and management changes needed for them to work.

AI runs the same risk. Even the most advanced algorithms cannot generate value if professionals do not know how to use them or if the organization does not have a management system that guides their application. Consider a well-known example from lean manufacturing: the andon system. Sensors and monitoring tools collect data from machines and production lines, but the technology, on its own, produces no results if people do not respond quickly when a problem occurs. Operators need to trigger the andon, supervisors need to investigate the problem, and leaders must ensure that it is solved and prevent future occurrences. Without these behavioral responses and these management routines, the technology loses its meaning.

Four leadership priorities for the AI era

So how can we ensure that the adoption of artificial intelligence generates real value? For me, there are four fundamental points to consider.

Redesign workflows, not just isolated tasks. The organizations that obtain the greatest benefits from AI are those that rethink their workflows as a whole. Automating isolated tasks rarely brings significant gains. Instead, leaders should analyze how work flows from start to finish and redesign the process based on the new capabilities available.

Do not treat AI as an IT project. AI adoption cannot be delegated entirely to technical teams. It requires strategic thinking about where the technology adds the most value and how its application connects to the organization's objectives. This is a leadership responsibility.

Focus on augmenting capability, not on replacing it. Although some people fear that AI will replace large portions of the workforce, the most promising approach is to use it to enhance human capability. When professionals have access to tools that help them work faster, analyze information better and solve problems more effectively, organizations can grow, instead of simply shrinking their structures.

Build trust through transparency and experimentation. Top-down mandates rarely succeed. Employees need to understand how AI can benefit both their work and customers. Starting with practical applications aimed at solving real problems helps build credibility and encourages broader adoption.

A future of cautious optimism

Artificial intelligence will certainly transform many aspects of work, but its impact will probably occur in a more gradual and uneven way than the headlines suggest.

Generative AI is powerful, but it is still a tool that requires conscious and careful integration into human systems. When leaders combine technology with behaviors and good management practices, this technology can accelerate learning, improve decision-making and raise productivity.

In the end, true transformation does not come from the technology itself, but from the way organizations learn to use it. For lean organizations, this means adopting disciplined experimentation, practicing continuous improvement (kaizen) and maintaining a commitment to developing human capabilities. Our ability to bring together technology, behavior and management will determine whether AI will be just another wave of passing hype or a genuine force for progress.

This article is based on a presentation by Art Smalley at Lean Global Connection 2025 on the challenges of leadership in the age of artificial intelligence.

This discussion will also be at the center of Lean AI Summit 2026, an event that will bring together leaders and experts to explore how to integrate lean management, agility and artificial intelligence to generate real value.

The gathering will take place on September 24, 2026, at Cubo Itaú, in São Paulo, with in-person attendance and global online streaming.