Structure: the Key to Understanding AI

AI is one of the most transformative technologies of our time, but its success often depends on something deceptively simple: structure. Structure is the framework that gives raw data shape, context, and meaning. It’s what transforms a jumble of information into something actionable and understandable. For leaders, understanding structure is imperative for assessing where AI works well—and where it doesn’t.

What Is Structure?

Structure is what connects and organizes data. Without it, data is just noise. For example:

• Raw Data: 11052021. What does it mean? It could be an ID, a product code, or even a random sequence of digits.
• Providing Shape: 11/05/2021. Formatting adds structure, suggesting it’s a date. But is it November 5th or May 11th? Structure improves clarity but doesn’t always resolve ambiguity.
• Context as Structure: On a form in the United States, where dates are typically formatted as MM/DD/YYYY, the context provides additional structure, helping us interpret the data as November 5th, 2021.

This example highlights two critical aspects of structure:

1. Rigidity: Shape and rules that organize the data.
2. Context: The environment or assumptions that help interpret the data.

Without structure, even the most advanced AI struggles to deliver meaningful results.

How Does Structure Relate to AI?

AI doesn’t "see" the world like humans. Instead, it identifies relationships, patterns, and connections within the data. Structure gives AI the framework to:

• Recognize patterns (e.g., edges in an image, trends in sales data).
• Interpret relationships (e.g., correlations between user behavior and purchase history).
• Make predictions (e.g., what a customer is likely to buy next).

The stronger and more consistent the structure in the data, the better AI performs. But when structure is weak or misleading, even sophisticated AI models struggle.

A Real-World Example: Parsing Receipts

Let me share an experience that illustrates the importance of structure. My team was tasked with building an AI to interpret arbitrary receipts and extract key information: items purchased, prices, and dates. At first glance, it seemed similar to tools already available from companies specializing in receipt parsing. However, understanding structure demonstrates why our task was fundamentally different.

Unlike those companies, which worked with well-defined constraints—specific vendors, consistent layouts, or curated datasets—we were dealing with arbitrary receipts. This meant:

• Highly Variable Formats: Each receipt had a unique layout, font, and organization.
• Noisy Data: Many receipts were poorly scanned or partially illegible.
• Inconsistent Product Information: Item names, groupings, and price calculations varied dramatically between vendors.
• Limited Training Data: Our dataset didn’t have enough real-world examples to represent this variability.

The variability was far larger than our dataset could elucidate. I was unable, at the time, to accurately explain this to stakeholders, but the lack of clear structure and sufficient data made the task effectively impossible. AI models can’t learn patterns when the data lacks consistency or the examples are insufficient to generalize.

This experience taught me an important leadership lesson: understanding structure is critical to setting realistic expectations for AI.

Where AI Works Well: Strong Structure

AI thrives in domains where data has a clear, consistent structure. Examples include:

1. E-Commerce Recommendations: Structure: Relationships between user behavior and product attributes (e.g., "people who buy tents often buy sleeping bags").Example: Amazon uses purchase and browsing history to recommend products effectively.
2. Fraud Detection:Structure: Patterns in transaction data, such as unusual locations or amounts.Example: Credit card companies detect fraud by analyzing deviations from typical spending habits.
3. Predictive Maintenance:Structure: Patterns in sensor data (e.g., vibration levels, temperatures) that indicate wear and tear.Example: Manufacturing firms predict equipment failures to minimize downtime.

Where AI Struggles: Weak or Misleading Structure

AI struggles in situations where data lacks clear organization, has too much variability, or is incomplete. Examples include:

1. Random or Rare Events: AI can’t predict random events (like lottery numbers) or rare events (like pandemics) because there’s no consistent structure to learn from.
2. Biased Data: When training data reflects historical biases, AI learns misleading patterns. Example: Amazon’s hiring algorithm penalized female candidates due to biases in historical hiring data.
3. Ambiguous Contexts: Without clear context, AI can misinterpret data. Example: A language model trained on US English might struggle with British idioms, failing to recognize their intended meaning.

A Final Thought: From Noise to Insight

AI’s power lies in its ability to find patterns we can’t see. But its limitations stem from the same place—it can only find patterns where structure exists. As leaders, understanding structure allows us to set realistic expectations, guide our teams, and maximize the impact of AI in our organizations.

Structure is what turns chaos into clarity, and understanding it is key to unlocking AI’s full potential in your organization.