Machine learning is no longer limited to the domain of math and coding experts. Its influence permeates every corner of modern business, from optimizing supply chains to personalizing customer experiences. But while leaders are adept at leveraging AI as a tool, they often overlook the powerful mental models that drive these systems.
Always fascinated by the intersection of code and creativity, I built my own AI from scratch last year. I really wanted to better understand what it could offer the organizations and leaders we work with. The biggest learning was not so much its strengths and limitations, but its strengths and limitations. Rather, it was a way in which the foundational principles of mathematics developed to solve complex computational problems provided a surprisingly fresh and effective lens through which to view and solve persistent organizational challenges.
The core concepts of machine learning are more than just the technical building blocks of an algorithm. These are sophisticated strategies for navigating uncertainty, finding the signal in the noise, and promoting adaptive learning. Translating from code to the boardroom provides a new vocabulary for leadership and a robust framework to improve decision-making, clarify strategy, and intentionally build a stronger culture.
Here are four fundamental concepts in machine learning that can help us all tackle complexity more clearly and effectively.
1. Cross-entropy: Clarity requires contrast
In machine learning, models are trained to make predictions. Is this image a cat or a dog? Is this transaction fraudulent or legal? An important function used to measure and modify model performance is called cross-entropy loss. There are heavy penalties for ambiguity. If the model is uncertain and predicts a 50% chance of a cat and a 50% chance of a dog, a high loss score is given. The model is rewarded for making a decisive choice, even if that choice is initially wrong. Because clear feedback enables learning.
Business leaders often face similar challenges. When faced with ambiguity, there is a natural tendency to avoid risk. They launch products with ambiguous positioning to appeal to everyone, provide vague feedback to avoid offence, and develop strategies that try to be all things to all people. This approach seems safe, but like any uncertain machine learning model, it introduces significant “losses” to the organization. Ambiguous messages confuse customers, and half-baked decisions without clear direction demotivate your team.
Adopting a cross-entropy mindset means seeking clarity through contrast. It is a mental model that forces you to choose between competing options and communicates that choice with confidence. When determining market positioning, do more than just describe what your product is. Define what it is not. As you articulate your new strategy, be clear about which initiatives you’ll stop in order to focus on the ones you’ll start. This determination can feel dangerous because it closes the door to other possibilities. However, it provides the clarity your team needs to work together and execute effectively.
2. Attention mechanisms: focus on what is most important in the current context
Large-scale language models and advanced image recognition systems, such as ChatGPT, rely on a concept called “attention.” In long blocks of text or complex images, not all information is equally important to the task at hand. Attention mechanisms allow models to dynamically weight the importance of different parts of the input and focus on specific words, pixels, or data points that are most relevant in a given context. This ability to selectively prioritize allows for nuanced understanding and sophisticated output.
In business, leaders are exposed to a constant stream of information: market trends, competitor trends, internal metrics, customer feedback, and team concerns. A common mistake is to treat all these signals with equal importance, leading to strategic paralysis and a reactive, unfocused agenda. [A recent study by found that 85% of business leaders have suffered from ‘decision distress’ questioning decisions they have made in the past year, with 72% admitting the sheer volume of data stopping them from making any decision – Oracle, 2023]
In fact, not all signals need to be treated equally. A good strategy is to ask, “What should I focus on now?” It’s not about seeing everything. It’s about knowing what to focus on and what to ignore.
Applying an attention mindset means dynamically shifting your focus based on changing goals and circumstances. For example, during a product launch, customer acquisition metrics may be the most important. During times of financial downturn, attention may shift to improving operational efficiency. This is not about abandoning other metrics, but rather intentionally giving more influence to the signals that are most important to the task at hand. It also enables a type of organizational multitasking, allowing different teams to apply different “heads of attention” to the same company-wide data set to focus on what’s most relevant to their job.
3. Gradient descent: Small steps down a slope are better than big jumps off a cliff.
At the heart of how most machine learning models are trained is an optimization algorithm called gradient descent. The goal is to minimize the “loss function,” which measures how inaccurate the model’s predictions are. Gradient descent works by taking small iterative steps rather than attempting one complete leap to the lowest point of error. At each point, it calculates a “gradient” (the direction of steepest descent) and increments in that direction. This process is repeated thousands or millions of times, gradually bringing the model to a state of minimum error.
In contrast, the business world is often obsessed with the idea of large-scale transformation. The company is pursuing a large-scale, multi-year “digital transformation” project, a fundamental reorganization, and a big bang product launch. A recent Harvard Business Review article called this the “transformational treadmill” (HBR – Magazine 2026). These efforts are the equivalent of jumping over a valley and trying to find the lowest point. They are expensive, risky, and often do not yield the desired results because the business environment changes before they are completed. A shift in mindset from large-scale transformation to continuous, incremental improvement is often desired. It is safer and more effective to walk down the slope little by little than to take a big jump off a cliff.
Employing a gradient descent model means breaking down large goals into smaller testable experiments. Instead of completely overhauling your website, you might run dozens of A/B tests on your existing site. Instead of a major corporate restructuring, you can also pilot a new team structure in one department. The key is to have a clear “loss function” that is the specific metric you are trying to improve, such as customer churn, employee engagement, or conversion rate. Each small experiment provides feedback to let you know if you’re on the right track.
Leaders can foster this approach by celebrating learning over perfect execution. Frame new initiatives as hypotheses rather than postponing conclusions. Ask yourself, “What is the smallest and fastest experiment I can run to test this idea?” This approach reduces risk because the cost of a single failed experiment is low. More importantly, it accelerates learning, builds a culture of adaptation, and enables organizations to navigate complex challenges by continually adjusting their path based on real-world feedback.
4. Self-care: parts of the system need to understand how they relate to each other.
Another key innovation behind the power of modern large-scale language models is a mechanism called self-attention. This allows each element in the input sequence (for example, every word in a sentence) to examine every other element in the same sequence and compare its importance relative to itself. The word “it” in a sentence acquires its meaning by noting the noun mentioned earlier in the text. This process creates a rich, context-aware understanding in which the meaning of each part is defined by its relationship to the whole.
This can provide a powerful model for breaking down one of the most persistent organizational challenges: internal silos. In many companies, teams and departments operate with limited awareness of each other’s priorities, pressures, and capabilities. Marketing teams launch campaigns without fully understanding the product team’s roadmap. Engineering teams make technical decisions without understanding the impact on customer support. This lack of internal context leads to friction, duplicated work, and missed opportunities. (For example, a recent survey of HR leaders showed that more than 8 in 10 companies reported significant misalignment between specific department efforts and broader business efforts, leading to an estimated $8.9 trillion in annual economic losses.) 8x AI, 2024. )
Applying the idea of self-attention means building a system for shared awareness. It’s about creating an environment where each team can effectively “participate” in the work and context of other teams, leading to more intelligent and coordinated actions. This is more than a simple status update. It requires building empathy and a deep understanding of how each function contributes to the larger mission. When sales teams understand engineering constraints and engineering teams understand market pressures, both parties can make smarter, more aligned decisions.
Take the lead with a new model set
By borrowing some of the logic that powers today’s most exciting technologies, you can equip yourself and your organization with more robust, flexible, and insightful ways to address business challenges. The job of a leader is not to be a programmer or an LLM expert, but to explore and implement the best AI tools for your business, as well as learn from the elegant strategies that make algorithms so powerful.
