What is Generative AI?

Generative AI refers to a class of advanced machine learning systems capable of producing human-like text, images, audio, and video in response to prompts. Built on transformer-based deep learning architectures, these models—such as large language models (LLMs) like GPT-3.5—are trained on vast and diverse datasets to learn the structure and semantics of language and other media. Once trained, they can generate contextually accurate and coherent outputs across a wide range of tasks, including text summarization, question answering, translation, and code generation. Their versatility and ability to synthesize information make them especially valuable for enhancing decision-making, automating knowledge work, and modernizing workflows across industries.

While generative AI covers text-generation, image generation, and audio generation, for the purpose of this article I am going to focus on how large language models are changing cognitive labor in the business world. To better understand generative AI and large language models I want to take a look at how they emerged from the world of neural networks.

The Neural Network That Learned to Reason

Large language models (LLMs) like GPT-3 are the product of a long evolution in artificial intelligence, rooted in the broader field of deep learning. Deep learning, a subset of machine learning, builds on artificial neural networks with many layers—designed to learn increasingly abstract patterns in data. Early breakthroughs in computer vision and speech recognition demonstrated their power, but it wasn’t until the introduction of the transformer architecture in 2017 that deep learning began to reshape natural language processing at scale. Transformers replaced earlier models like recurrent neural networks, enabling systems to better capture the relationships between words, regardless of their distance in a sentence.

Transformers use a self-attention mechanism to determine which parts of the input matter most for understanding meaning. This architecture became the foundation for models such as GPT-3, which are capable of reading, interpreting, and generating human-like language. GPT-3’s architecture scales this design to extremes: its largest version includes 175 billion parameters, layered into 96 attention blocks. Such scale makes the model both powerful and costly—training it took hundreds of years of GPU time and millions of dollars in compute resources. But the payoff is clear: GPT-3 can generate language, interpret questions, and perform tasks it was never explicitly trained to do.

Among the most notable capabilities of these models is in-context learning—the ability to pick up a new task from just a few examples provided in the prompt, without retraining. This is a departure from previous approaches, which required reconfiguring the model for each new task. GPT-3 also shows signs of reasoning, planning, and tool use when chained with external systems like Python interpreters or search engines. These behaviors were not explicitly programmed but emerged from the scale of training and data exposure—what some researchers describe as “emergent abilities.”

The source of these capabilities lies in the vast amount of data the model was trained on—300 billion words from websites, books, and Wikipedia—and the size of the model itself. Language generation flows from the training objective of predicting the next word. World knowledge comes from the scale and diversity of the data. But the true origin of in-context learning—why the model can generalize from examples rather than updates to its weights—remains one of the open questions in the field.

For technology leaders the message is clear: LLMs are not just larger models—they represent a fundamental shift in how machines can engage with information.

Not Just the Same Old Machine Learning

Unlike older systems that needed to be trained or coded for every task, LLMs adapt through language.

They are reasoning engines built on the backbone of neural networks, now capable of parsing context, proposing strategies, and even helping teams work faster. As these systems become more integrated into enterprise tools, they will begin shaping how decisions are made, not just how data is processed.

Artificial Intelligence has long captured our imagination—what author Pamela McCorduck once called “an ancient wish to forge the gods.” Traditional machine learning methods, foundational to data science, rely on algorithms that extract patterns and structure from specific organizational datasets. These models are narrow in scope, trained to perform well within a constrained domain where the data distribution is familiar and relatively static.

Generative AI models, particularly large language models like GPT3 (and its line), represent a significant departure. Trained on massive corpora of text, they are built not just to detect structure but to reason across varied and unfamiliar contexts. Unlike traditional models, which are tailored to a specific dataset, generative models are broadly applicable, portable, and capable of handling a wide array of tasks through natural language input.

The emergence of reasoning over language with large language models, much like the ability to detect objects with the convolutional neural network architecture, allows LLMs to be applied in problem domains and on data they have yet to directly learn from. This is a key differentiator for LLMs as compared to traditional machine learning.

Why Large Language Models are Useful

This shift from pattern recognition to conceptual reasoning makes generative AI a strategic asset for enterprises ready to rethink how knowledge work is done.

Generative AI can produce text that is remarkably human-like in its fluency and coherence. This capability stems from the sophisticated patterns the models learn during their extensive training on diverse text data. Text generation is particularly interesting to F1000 enterprise customers as it provides a raw reasoning engine to make nuanced and complex analysis over data embedded as text in the input. LLMs take a string known as a prompt as input and then outputs another string to answer the question or request posed in language and data of the input string. Text generation output can range from simple question/answer to code generation, article writing, and even poem creation in distinct styles.

Large language models (LLMs) represent a foundational shift in computing, not just an upgrade. Their real value lies in embedding reasoning directly into software, enabling systems to handle complexity, make decisions, and adapt quickly.

Rather than simply adding features, LLMs open the door to rethinking entire user experiences. They lower the cost of complex tasks, increase execution speed, and expand what's technically possible. The result is a new generation of software that’s more intuitive, responsive, and capable of handling judgment-based tasks once reserved for specialists. Used well, they don’t just enhance workflows—they redefine them.

Types of Generative AI Applications

Examples of generative AI applications include:

1. ICD code translation from raw doctors text notes
2. Meteorologist specialist that can answer questions about current and past hurricanes, forecasting how they might impact a reinsurance portfolio
3. A medical sales assistant that pulls together prospect lists based on physician services specialty and physical location
4. visual analytics UX driven by natural language
5. complex scenario analysis based on natural language description
6. business management assistant that analysis current business metrics and recommends where to focus operational improvement

I generalize these applications of generative AI into 3 types:

• conversational AI
• workflow automation
• decision intelligence tools

There are many common use cases today, such as code generation, or chat bots, that I fold into these major grups. In the sections below I give details on what makes each group distinct.

Conversational User Interfaces

Generative AI-powered conversational interfaces are reshaping how organizations interact with both customers and internal systems. Text-based chatbots now automate routine service tasks, streamline semi-technical operations, and reduce the load on human agents—allowing teams to focus on higher-value work. Platforms like Bank of America's Erica, with over 1.5 billion interactions, demonstrate the scale and effectiveness of personalized AI-driven engagement. These interfaces not only enhance responsiveness but also provide meaningful, real-time feedback, improving both operational efficiency and customer satisfaction.

The rise of messaging apps—with over 3 billion global users—signals a decisive shift in communication preferences. Consumers favor messaging for its speed, accessibility, and ability to deliver rich, multi-modal interactions. With business messaging adoption accelerating, generative AI integrated into these platforms becomes a natural extension of digital service. Over the next decade, conversational AIs will evolve into persistent, memory-rich digital assistants that personalize experiences across voice and text. For insurers, this means scalable, 24/7 engagement capabilities and a future-ready approach to customer interaction.

Workflow Automation

Generative AI is redefining workflow automation by adding cognitive decision-making and adaptability to traditional robotic process automation (RPA). While RPA handles structured, repetitive tasks, generative AI expands the scope to include unstructured data and complex judgments. This fusion allows organizations to automate not just routine operations, but also tasks requiring contextual understanding—such as processing customer requests, analyzing claims data, or interpreting technical diagnostics—making automation more flexible, accurate, and scalable.

A key opportunity lies in integrating legacy systems into modern workflows. Generative AI can interpret and bridge data from older platforms, preserving existing investments while enabling smarter, end-to-end automation. Additionally, iterative feedback loops, combined with a multi-model architecture and human oversight, help mitigate reliability concerns by increasing transparency in AI-driven processes. The result is a more intelligent automation layer—one that accelerates execution, reduces downstream errors, and supports continuous improvement across diverse operational domains.

Decision Intelligence

Generative AI, particularly large language models, enables organizations to rapidly simulate complex business scenarios that once required cross-functional teams and extended analysis. These models support swift exploration of strategic options, allowing decision-makers to test contingencies and forecast outcomes with greater agility. By modeling operational dynamics and economic behaviors, LLMs introduce a new level of depth and speed to business strategy, supporting faster, more informed decisions.

Executives can synthesize many complex scenarios and eliminate the bulk of the ones that won't work before focusing on promising strategies. An example of this is how a CFO might ask a system "how much property damage is at risk if a category 4 hurricane hits a specific state?". This allows the CFO to build some ideas around what may be coming before she takes a few of them over to the company meteorologist for deeper validation.

The Mechanization of Cognitive Labor

Generative AI represents a turning point in the evolution of knowledge work, much like the mechanical loom reshaped hand weaving. Rather than replacing people, it mechanizes cognitive tasks—transforming how information is processed, decisions are made, and value is created. This shift mirrors earlier waves of transformation, where new tools didn’t eliminate labor but restructured it. In each case—from the invention of writing to the printing press and now AI—the core outcome has been the same: tools extend human capability, allowing organizations to scale coordination, accelerate insight, and focus skilled labor on higher-order problems.

This is not replacement, but realignment—freeing teams to focus on what matters most while the machinery of reasoning runs in the background.

The competitive edge lies not in cutting labor but in amplifying talent. The organizations that thrive will be those that use AI to enhance the judgment, speed, and output of their best people. Just as tools have always extended human reach, this new class of tools extends our ability to understand, decide, and act—faster and at greater scale. The stakes remain unchanged; only the means have evolved.