In the past year, Generative AI (GAI) has transitioned from a purely academic interest to a major talking point in global news, largely thanks to the democratization of tools like ChatGPT and Stable Diffusion. These platforms have reached the general public with free access and easy-to-use interfaces, sparking interest across various sectors. But from a business perspective, is this wave of attention merely hype, or are we in the early stages of a significant technological revolution? Moreover, is Generative AI capable of creating new and innovative use cases that can transform industries?
The Era of Large Language Models (LLMs)
Generative AI focuses on creating new content, marking a new stage in artificial intelligence. Traditionally, statistical analysis could only provide descriptive and diagnostic insights. With AI, particularly through machine learning (ML) techniques, businesses can now benefit from predictive and prescriptive analytics. These techniques forecast patterns or situations and suggest recommendations or alternatives.
Generative AI goes beyond this, offering “creative analytics,” capable of not only analyzing existing data but also generating new information. While Generative AI applies to all forms of human creativity—text, audio, images, video—its most well-known application is Large Language Models (LLMs), such as OpenAI’s ChatGPT, which has the honor of being the fastest-growing app in history. Unlike traditional methods that required understanding machine languages, users can now interact with LLMs using natural language.
The Shift from Traditional AI to Generative AI
In the last decade, “traditional” AI (anything more than five years old) has transformed business models. Much like the democratization of computers in the 1980s and the internet in the 1990s, AI has optimized processes, making them more efficient and secure, and enabling data-driven decision-making. These digital transformations are becoming foundational pillars for businesses worldwide.
Generative AI, however, introduces a new layer of capability. For businesses, this could unlock untapped potential. But what are the practical use cases for Generative AI in the business landscape?
Use Cases for Generative AI
- Internal Documentation Search: With large volumes of unstructured text data, businesses often struggle to find relevant information efficiently. Traditionally, Ctrl+F was the go-to tool for searching through documents, which was neither efficient nor reliable. Generative AI allows users to query large datasets in natural language, extracting relevant concepts, strategies, and answers from vast databases.
- Virtual Assistants and Chatbots: With the precision and versatility achieved by LLMs, chatbots can now respond more comprehensively to user questions. One of the main advantages of LLMs is their memory, enabling them to maintain coherent conversations and pick up from where the user left off, even after interruptions. This means that most customer problems can be resolved quickly, only escalating to human intervention for the most complex issues. Additionally, sentiment analysis on previous interactions can help understand what strategies and solutions were most satisfying to customers.
- Synthetic Data Generation: Generative AI allows for the creation of synthetic data, which can be useful for expanding datasets when obtaining real-world data is time-consuming or resource-intensive. With GAI, businesses don’t need advanced statistical knowledge to achieve this, making it accessible for smaller organizations.
- Proposal Writing: Generative AI can craft new value propositions based on existing documents. The AI can align these proposals with the company’s strategy, client type, team dynamics, deadlines, and other factors.
- Executive Summaries: Using AI to generate summaries from large documents, or even transcriptions from voice recordings, is another powerful application. Many companies have to deal with lengthy and complex texts—technical reports, legal documents, or even European regulations—that are difficult to digest quickly. GAI can summarize the most important points, adapting the style to the user’s needs.
- Personalized Marketing: Traditional AI allows for customer profiling and segmentation, helping to design customized campaigns. GAI takes this a step further by enabling personalized marketing at an individual level. This could mean that no two customer interactions are the same, while still adhering to the company’s values and existing campaign style. Much like auto-generated YouTube captions, GAI opens the door to creating content at an impossible scale through manual efforts.
Developing GAI for Your Business: Build or Buy?
Given the growing use cases, how can a company implement a customized LLM? The most obvious answer is to build one from scratch. However, it’s no coincidence that the best LLMs come from hyperscalers like OpenAI, Google, Meta, and Amazon. Training these models requires enormous amounts of data, time, and resources.
For example, training GPT-4 reportedly cost around $100 million—an astronomical figure for most businesses. Furthermore, training models based solely on existing company data may result in outdated models. GPT-3, for instance, only contains information up to September 2021. While GPT-4 integrated conversations with GPT-3 into its training, this practice led to some privacy concerns. In certain instances, users could inadvertently share confidential information that was accessible by others simply asking the right questions. As a result, many companies have restricted the use of these models to prevent data leaks.
Given these challenges, training an LLM from scratch may not be feasible. Instead, businesses can explore two alternative strategies: Fine Tuning and Retrieval-Augmented Generation (RAG).
Fine Tuning vs. Retrieval-Augmented Generation (RAG)
Both strategies offer ways for businesses to tailor LLMs to their specific needs. The choice between them depends on various factors.
- Fine Tuning: This approach involves refining an existing LLM, like ChatGPT, by incorporating company-specific data into its training. While it can be costly, the cost is orders of magnitude lower than training a model from scratch. Fine tuning essentially adapts a pre-trained LLM to understand the company’s unique language, customer data, and case scenarios.
- RAG: With RAG, the LLM references a structured database with pertinent company information. When a user asks a question, the model pulls relevant data from this database to formulate an answer. The advantage of RAG is that the model can also show the user which documents it referenced to generate the response, making the information more traceable and verifiable.
Fine Tuning or RAG: Which is Right for Your Business?
Deciding between fine tuning and RAG depends on several factors, such as:
- Data Volume: If the company has a small amount of data, fine tuning may not be effective. The model could suffer from overfitting, where it learns too much from a small dataset and becomes less generalizable. In such cases, RAG may be more suitable, as it allows the model to consult a structured database without extensive retraining.
- Privacy: Privacy is a key consideration for businesses. If employees in different departments need access to different types of information, RAG allows the company to restrict access based on user roles. In contrast, fine tuning incorporates all company data into the model, which could lead to potential data leaks, similar to what was experienced with GPT-4.
- Interpretability and Versatility: LLMs are prone to hallucinations, where the model generates incorrect or fabricated responses. If a business needs a model that can provide traceable and accurate answers (e.g., for summarizing legal documents), RAG is the better choice. However, if the use case requires flexibility (such as a chatbot handling diverse customer queries), fine tuning would be more appropriate.
- Data Freshness: If the business environment requires regular updates, such as incorporating new data on a daily or hourly basis, RAG is more efficient. Fine tuning “freezes” the model at a certain point in time, so incorporating new data would require retraining, which is costly and inefficient.
Combining Fine Tuning and RAG
For businesses seeking the best of both worlds, a hybrid approach is also possible. Fine tuning can be used to incorporate a large corpus of historical data, while RAG can handle more recent data. This would enable the model to address a wide range of scenarios based on historical data, while also staying up-to-date with current events and trends.
For instance, a chatbot trained on historical customer service data through fine tuning could manage a variety of common inquiries. Simultaneously, RAG could ensure that the model stays informed of any newly opened service tickets, enabling it to address ongoing customer issues effectively.
Conclusion: A Personalized LLM for Every Business
With careful consideration of the factors mentioned above, businesses can choose between fine tuning, RAG, or a combination of both, to implement customized LLMs for their needs. Generative AI opens up a new frontier of possibilities, offering businesses the chance to harness creative analytics to enhance operations, improve customer service, and innovate at scale.
The current wave of Generative AI attention is not merely hype—it represents a significant shift in how businesses can generate value. By understanding the capabilities and limitations of tools like fine tuning and RAG, businesses can start building personalized AI solutions, providing a competitive edge in a rapidly evolving digital landscape.