The Pin Factory Paradox

In 1776, Adam Smith described a pin factory to explain the efficiency of the division of labor. He noted a skilled generalist working alone could make ~20 pins a day. However, by dividing the labor so that one man draws the wire, another straightens it, a third cuts it, and so forth, ten men could produce 48,000 pins a day.1

Smith wasn’t just describing a factory. He was describing a patch for the human mind. He recognized that while humans are “general intelligences,” we suffer from high “switching costs”. The metabolic tax of shifting our attention from one context to another is so high that we must organize ourselves into rigid, specialized slots to get anything done.

For the last three centuries, we have viewed this bureaucracy as a necessary evil. It was a biological limitation we hoped to eventually transcend.

Then we built the “God Model.”

The Death of the Monolith

When GPT-4 arrived, the tacit assumption in Silicon Valley was that Artificial Intelligence would be the ultimate generalist. We imagined a single, monolithic neural network that could do it all. We thought it would write the code, design the UI, and manage the database without the messy overhead of departments or managers.

But look at the state-of-the-art in 2026. The “God Model” dream is dead. Instead of a single digital genius, the industry has converged on architectures that look suspiciously like the very thing we tried to escape: bureaucracy. We are building “Orchestrators” (managers), “Workers” (specialists), and “Gateways” (compliance officers).

It turns out that intelligence, whether biological or silicon, is subject to the same laws of physics.

Subscribe now

Lessons from the Flat Organization

I’ve been tracking the “organizational design” of AI agents since the chaotic days of 2023. Back then, we had the “ReAct” paradigm (chain-of-thought reasoning with external actions) with simple loops like AutoGPT. These were the AI equivalent of a “flat organization”. Every agent could talk to every other agent. There was no hierarchy, and there were no managers.

Just like the famous “flat” experiments at Zappos2 or Valve3, it resulted in chaos. These flat swarms suffered from “loops of death” and massive hallucination spirals. Without a hierarchy to filter information, the noise amplified until the system crashed.

By late 2025, the trend line shifted aggressively toward structure. The release of Claude Opus 4.5 and GPT-5.2-Codex didn’t just give us smarter models. It gave us models capable of submitting to a “boss”.

We aren’t seeing the liberation of intelligence. We are seeing the industrialization of it. The cutting-edge AI architecture of 2026 is essentially a digital org chart. It consists of containerized pods of specialized agents strictly governed by an orchestration layer.

The Transaction Costs of Thought

Why is this happening? Why can’t a super-intelligent model just “figure it out”?

To understand this, we have to look at the Transaction Cost Theory of the firm, introduced by Ronald Coase in 1937.

Coase asked why companies exist at all. Why don’t we just contract everything out in a free market? His answer was that coordination is expensive. There are costs to finding the right person, negotiating the price, and enforcing the contract. When those costs are high, you build a firm (a hierarchy) to reduce the friction.

AI is facing its own “Coasean Moment.”

1. The Cognitive Transaction Cost. Just as humans have “Bounded Rationality” (a limit on how much we can process), AI models have “Context Windows”. Even with “Context Compaction”, dumping every piece of information into a single model creates “Context Drift”. The model gets confused. Specialization is the fix. By breaking a complex objective into small, specialized “MCP Servers” (Model Context Protocol), we lower the cognitive load on any single agent. We have one server for the database, one for Slack, and one for the file system. We are essentially creating “departments” to handle the information overload.

2. The Principal-Agent Problem. In economics, the “Principal-Agent Problem” occurs when a worker (the Agent) doesn’t perfectly align with the owner’s (the Principal) goals. In AI, we call this “Alignment” or “Safety.” A rogue agent with root access is a security nightmare. The solution in 2026 mirrors the solution in 1920: Middle Management. We now use “Orchestration Layers” to act as the digital middle manager. This layer doesn’t do the work. It audits the work. It ensures the “Worker Agent” is not hallucinating or trying to execute a malicious command. We’ve reinvented the supervisor because trust is not scalable.

3. The Scaling Laws of Agency. Research from DeepMind in 2026 formalized this with the “Scaling Laws of Agency”. They found that adding more agents to a flat swarm does not linearly increase performance. Instead, it exponentially increases coordination friction.

This is mathematically identical to “Dunbar’s Number” in sociology, which suggests human groups fall apart without structure once they exceed roughly 150 members. The optimal topology for AI, it turns out, is a hierarchy.

Structure is a Feature

There’s a profound irony here.

For decades, technologists have viewed the corporation with its org charts, memos, and managers as a relic of the past. We thought silicon would liberate us from structure.

But it seems that structure is not a bug of human biology. It is a feature of general intelligence.

As we scale AI toward AGI, we are not building a god. We are building a firm. The limiting factor of the future will not be raw compute. It will be organizational design.

The question for us, as leaders, shifts from “How do I prompt this model?” to a much older, more familiar question:

“How do I structure this team?”

A lot of enterprise knowledge is hiding in plain sight. Think of the rich data captured in “unstructured data”: PDFs, PowerPoint presentations, Word documents, images, charts, and even our communication systems (Slack, Teams, email, etc.). If we’re being honest, despite massive investment in the storage of “structured data” (databases, data lakes, business intelligence tools, etc.), much of the knowledge in this data is inaccessible as well.

This is a huge challenge for businesses, and innovation in particular. Christian Szegedy, an AI research scientist and co-founder of xAI, stated:

Many have come to believe a key to unlocking this hidden knowledge, finding these obscure connections, and, as a result, accelerating innovation is the creation of next generation enterprise search capabilities powered by GenAI. Most of these solutions are leveraging an approach called retrieval augmented generation (RAG), a specific type of GenAI enabled by large language models (LLMs). For the brave among you, I’ll dive into RAG more deeply in the Technology Wise feature below. First though, let’s discuss LLMs.

Four LLM superpowers.

Previously I wrote about the different types of AI. A specialized type of the deep learning models described there is the large language model, or LLM. LLMs are “very large deep learning models that are pre-trained on vast amounts of data”1 and are notable for their “ability to achieve general-purpose language generation and understanding.”2 Getting to know a little more about how LLMs work reveals their superpowers and, as importantly, their limitations.

Scale: LLMs are trained3 on an enormous volume of text to predict the next word (or “token”4) in a sequence. In the process of learning to predict efficiently and effectively, LLMs learn patterns and structures representing the vast array of human knowledge expressed in text across the Internet. The learned patterns and structures provide LLMs with a capability which can be thought of as a contextual understanding - an understanding of how things work, work together, and relate to one another - albeit one based on statistical relationships and dependencies rather than genuine comprehension. One of the surprising characteristics of LLMs, like the ones powering OpenAI’s ChatGPT and Meta’s Llama, is how well they scale to massive quantities of training data. Llama 3, for example, was trained on 15 trillion tokens, or approximately 12 trillion words of text. Bottom line: The scaling superpower provides LLMs with a very broad and deep contextual understanding, including an understanding of extremely rare patterns and structures.

Similarity: Traditional search technologies relied on keyword-based methods like inverted indexes, which were limited because they depend on exact term matching and could miss the semantic meaning of queries. In 2013, Google introduced Word2Vec, an embedding5 model that produced vector (or numeric) representations of words, positioning similar words closer together. Embeddings became feasible for use in search applications recently due to advancements in computing power and the development of the transformer6 architecture in 2017. Transformers, which enabled modern LLMs, use a mechanism called self-attention7 to understand the relationships between words in much longer sequences, regardless of their position. Bottom line: This superpower enables more accurate and semantically relevant search results, making it easier to find specific information in vast datasets consisting of both structured data (like databases and spreadsheets) and unstructured data (such as PDFs, emails, and social media posts).

Subscribe now

Summarization: LLMs excel at distilling large quantities of text into concise summaries. To some extent they learn this capability during training on next word prediction - an example of a learned pattern (or abstract concept). However, they can also be specifically trained in this skill. LLMs can be fine-tuned8 by exposing them to examples of text paired with a human- (or AI-) curated summary. These examples teach the LLM key aspects of great summarization, aligning their output with their user’s preferences. Bottom line: Whether inherent or fine-tuned, this superpower makes LLMs incredibly useful in business settings, where they can condense lengthy reports, research materials, and unstructured text into clear, actionable insights that enhance the decision-making process.

Synthesis: There’s no reason to believe the high-level abstractions created during LLM training are domain specific. Anthropic’s work on interpretability seems to support this idea. Think of something like the exponential function. It would be far more efficient for a model to learn and store a concept once, rather than once for bacterial growth and once for radioactive decay. As the LLM predicts the next word, it draws on these abstractions and, in the process, can produce outputs which apply them across domains in a way that might not be obvious to human thinkers. Bottom line: This superpower makes LLMs invaluable tools for brainstorming sessions, strategy development, and innovation workshops, where generating a broad range of ideas quickly can be crucial to success.

Two LLM limitations.

While LLMs have amazing superpowers, new technologies rarely come without caveats. Let’s look at two key limitations of LLMs.

Hallucinations: It’s important to remember that LLMs are not information/fact retrieval systems; they are pattern and structure detectors that predict (or “guess”) the next best word in a sequence. Imagine an LLM is happily going along predicting the next word and it comes to a situation where the patterns and structures it learned during training for a context were weak, likely caused by limited occurrences in the training data. It will select the highest probability word, but there is no guarantee it is the correct word. Building on this incorrect word it continues selecting words with great confidence which are plausible but nonfactual given the original context. Researchers call these fabrications “hallucinations”9. LLMs can generate incorrect dates, numbers, names, web sites, and quotations. A partial solution for hallucinations is the RAG architecture I describe below, but this characteristic is not always negative or a limitation. Human creativity often appears like an out of bounds prediction at first. Many uses of LLMs outside of business (such as writing poetry, songs, fiction, etc.) and some uses within business (such as the cross-domain synthesis we discussed earlier) benefit from hallucinations. The key for business leaders is to understand they occur, and architect systems in ways which minimize hallucination risks and instead leverage their creative and innovative aspects.

Safety and Security: As powerful as LLMs are, they come with significant safety and security concerns. One major issue is bias. LLMs can unintentionally learn and propagate biases present in their training data, leading to outputs that reinforce stereotypes or discrimination. This can have serious societal impacts, from spreading misinformation to influencing hiring decisions unfairly. To minimize biases, it's important to use diverse and representative training data and implement bias detection and correction mechanisms. Security is another critical concern. LLMs can inadvertently expose personally identifiable information (PII) or protected health information (PHI), posing risks to individual privacy. Additionally, they might misuse or leak intellectual property. Implementing strict data handling protocols and access controls helps safeguard against these risks. Recognizing and addressing these safety and security issues helps businesses harness the power of LLMs safely, responsibly, and effectively.

Balancing superpowers and limitations unlocks innovation.

Large language models offer businesses breakthrough capabilities in scale, similarity, summarization, and synthesis, unlocking hidden insights and driving innovation. However, it's important to be aware of their limitations, such as hallucinations and the potential for biases and security risks. By understanding and mitigating these challenges, businesses can effectively leverage LLMs to transform vast amounts of data into actionable intelligence, while minimizing the risk of hallucinations, maintaining ethical standards, and ensuring regulatory compliance.

Subscribe now

As always, the quick section below will dive deeper into the technology, so keep reading but only if you want to geek out!

One of the most widely adopted approaches to minimizing LLM hallucinations and increasing the groundedness of their generated responses is retrieval augmented generation (RAG). RAG is a method that combines the strengths of information retrieval and LLM text generation. In addition to hallucinations caused by limited exposure to concepts in their training data, RAG addresses two other important LLM limitations. First, LLMs are trained using data as of a certain point in time. Unless the model is further trained, its parametric knowledge stops as of this date and it will be unaware of events subsequent to this point in time. Second, LLMs are trained on data freely available on the Internet. Data which is proprietary to an organization will (hopefully!) not be held in the model’s parameters. Let’s walk through how RAG addresses these limitations.

First, contextual data is transformed into dense, low-dimensional vectors using an embedding model and stored in a specialized vector database. Think of this as taking the information the user wants to complement the model’s knowledge with (be it timely text, a PDF, an Excel sheet, etc.), transforming it into a purely numeric representation, and storing it for semantic search later.

Next, the user asks a question which is passed through an embedding model to create its vector representation. This vector is compared against all stored vectors in the database using metrics such as cosine similarity10. This determines how close or far the vectors are from each other in the vector space, effectively identifying the most relevant documents.

Finally, both the user’s question and the retrieved documents are fed into a language model, which uses the context and knowledge from the documents to generate a detailed and (hopefully!) accurate response. This process helps ensure that the generated answer is grounded in real, up-to-date information, making it more reliable and informative.

Let me hear your “natural language”.

What questions, criticisms, or suggestions do you have? Where would you like me to go from here? Please feel free to engage with me and the rest of the community in the comments below.

Postscript.

If I told you a new solution was using artificial intelligence, what would that mean to you? You may think of a chatbot like OpenAI’s ChatGPT or an image generator like DALL-E, but AI is a diverse set of technologies, each with unique capabilities. Despite its recent rise in popularity, AI is not new; applications of AI have been adding value in businesses since the 1970s. As a leader responsible for defining strategy and purchasing technology, answering the following questions will help you make informed decisions and avoid falling prey to sales and marketing hype.

Is it Artificial Intelligence?

First, let's define artificial intelligence. The ‘artificial’ simply means we’re talking about the “intelligence of machines or software, as opposed to the intelligence of humans or animals”.1 Notice the definition doesn’t state the magnitude of the intelligence the machine/software must possess. AI can be anything from a set of if-then rules written by an expert, to an artificial general intelligence (AGI) where the machine/software “could learn to accomplish any intellectual task that human beings or animals can perform”.2 Given this broad definition you can see why so many vendors add AI to their websites! The key question for leaders becomes: what’s the nature of AI in the solution I am evaluating?

Is it Machine Learning?

It’s safe to assume most vendors who say they’re leveraging AI today are using machine learning (ML). Unlike the if-then solution described above, ML does not require an expert to predefine the rules. Instead it learns the rules by analyzing carefully prepared data (called features3) and identifying patterns and structures within. ML algorithms are valuable to businesses because they can identify these structures across large quantities of data, define rules which represent the identified structures, and then apply these rules to new data to make predictions. Practical applications from long before the current AI hype cycle include customer segmentation, forecasting, recommendation systems, fraud detection, algorithmic trading, route optimization, credit scoring, sentiment analysis, and many more.

Is it Deep Learning?

Deep learning (DL), a key driver of the current AI revolution, is a sophisticated form of machine learning. It's built on artificial neural networks4 (ANNs), which are inspired by neurons in the human brain but are much simpler. The “deep” simply means the network consists of multiple layers between the input and output layer. Each layer transforms its input data into increasingly abstract and complex representations. This structure allows DL to automatically extract important features from data, removing the need for manual feature engineering. Breakthroughs in algorithms and hardware over the past decade have enabled DL to improve on multiple ML techniques, particularly in highly complex and unstructured data, and led to widely known examples like AlphaGo5. Interestingly, for tasks that involve making predictions on structured data, like the tabular data in a spreadsheet, DL models tend to be outperformed by traditional ML methods.

Is it Generative AI?

Generative AI (GenAI) has existed for many years, but recent advances in deep learning, highlighted by the development of models like the transformer6, have significantly increased its capabilities. These advances enable GenAI models to efficiently learn patterns and structures from input data and create new, similar content. GenAI stands out in machine learning for its nascent abilities to seemingly model the world, reason, plan, and take action, venturing into realms once exclusive to human creativity. While many amazing consumer technologies are being created using GenAI, such as sophisticated virtual assistants and generated art, music, and video, the implications for businesses are enormous. By lowering the marginal cost of creativity and intelligence, GenAI amplifies human capabilities across industries dramatically increasing the productivity of both the white- and blue-collar workforce.

Now, be curious and ask questions.

The most important advice I could give business and technology leaders evaluating a new technology is to be intellectually curious and ask questions. Obviously the first question you need to answer is, “What business problem(s) does this AI solution solve?” For help answering this, please take a look at my previous post on building an AI Initiative Inventory. Assuming this is answered favorably, and equipped with the basic understanding of the types of AI provided above, you should now be comfortable asking the question, “What type of AI does your solution utilize?” When they answer, listen carefully for the following:

• Do they refer to specific types of AI like machine learning, deep learning or generative AI? This is a good indicator there is something real codified in their solution. Ask follow up questions like, “Can you give me examples of this type of AI being successfully applied in other fields or consumer applications?” Also, “What are the limitations of this type of AI?”.

• Do they explain the application of this AI in terms of how it amplifies human capabilities, translating to an increase in productivity? How would a capable human solve the problem their solution does, and in what ways is their methodology better?

• Do they talk about how their models are “continually learning and improving” as they receive new data? If they do, ask them how often their models are retrained or to explain how their reinforcement learning7 approach works. In many business sales scenarios this will be marketing/sales speak. Other than through complicated, emerging, and/or computationally expensive techniques, most business solutions leveraging AI only incorporate new learnings when they are retrained.

Finally, make sure you ask them if your proprietary data will be used to train their AI models for other customers. It's crucial to understand how your data might be utilized beyond your own use case, especially in terms of privacy and competitive advantage. This question will help assess the vendor's commitment to data confidentiality and whether their solution might inadvertently benefit your competitors.

Subscribe now

As always, the quick section below will dive deeper into the technology, so keep reading but only if you want to geek out!

Most popular applications of AI improve along with advances in the underlying ML, DL, and GenAI technologies. This makes sense when you understand how they relate to and build upon each other. Lets walk through three highly visible and valuable examples business leaders should be aware of.

Notice how computer vision8, itself a type of AI, overlaps in our diagram with ML, DL, and GenAI. Computer vision enables computers to interpret and process visual information from the world, similar to human vision. Using digital images from cameras and videos, along with ML, DL, and GenAI models, it can accurately identify and classify objects, and then react to what it "sees." Advances in ML and DL have significantly improved the accuracy and efficiency of computer vision systems, enabling them to process complex images with greater precision. GenAI breakthroughs contribute by creating new, varied training data, enhancing the robustness and adaptability of computer vision models in real-world scenarios. State-of-the-art examples include autonomous drones using advanced image recognition for efficient delivery, AI-driven quality control systems in manufacturing that detect minute defects with high accuracy, and real-time facial recognition technologies enhancing security and personalized customer experiences.

Natural Language Processing9 (NLP) is another type of AI that combines computational linguistics with ML, DL, and GenAI and enables computers to understand, interpret, and respond to both written and spoken human language. Recent advances are revolutionizing NLP by enabling more complex language models that can better grasp context, subtleties, and nuances in human communication, leading to more effective and natural interactions between humans and machines. State-of-the-art examples include advanced chatbots that provide highly personalized customer service experiences, AI-driven sentiment analysis tools for real-time market and consumer insights, and sophisticated language translation services that enable seamless international business communication.

Robotics10 is a broad field of technology that involves the design, construction, operation, and use of robots, integrating concepts from computer science, engineering, and AI. This field focuses on creating machines capable of performing tasks in various environments, often those that are too dangerous or tedious for humans, by mimicking human actions or autonomously making decisions based on programmed instructions or AI. Advances across ML, DL , GenAI, computer vision, and NLP are enabling robots to learn from experience, recognize and interpret visual and auditory inputs more accurately, and make complex decisions, leading to more dynamic and autonomous applications. State-of-the-art examples include autonomous warehouse robots using DL for efficient inventory management, robotic surgical assistants enhanced with precise computer vision for medical procedures, and customer service robots equipped with advanced NLP for engaging and helpful interactions in the retail and hospitality sectors.

Each of these higher-order technologies existed long before ChatGPT and have proven use cases in business and industry. The challenge for most leaders today is the breakneck pace of innovation. With all of the context presented here, it might be worth your time to revisit the AI Initiative Inventory I described in my previous post and make sure your strategy incorporates the full spectrum of potential use cases. Given the pace of change this should be a regularly updated document!

Let me hear your “natural language”. 

What questions, criticisms, or suggestions do you have? Where would you like me to go from here? Please feel free to engage with me and the rest of the community in the comments below.

Postscript.