Unlocking Generative AI for Content Understanding in Enterprises 

AI is shaping up to be a true paradigm shift across all industries. OpenAI's groundbreaking GPT technology is the most widely known example of how AI is reshaping our daily experiences in ways we couldn’t predict just a few years ago.

Although AI has been around for decades, its practical application to real-world problems has only recently become feasible due to advancements in computing power and the availability of data at a large scale. We are now entering the era of "applied AI," where the breakthroughs achieved by academic research groups are finally becoming accessible for businesses to unlock real value. With the latest developments in generative models, adopting AI has become easier than ever, transforming the landscape of how we utilize this technology.

This article will explore the tangible impact AI is already making in businesses and discuss key factors needed to unleash its full potential.

Applying AI in Business

Enterprises that have been able to leverage AI to solve complex business problems frequently emerge as “category leaders” and command a premium because of their ability to innovate. However, even the most advanced companies have mostly focused on structured data – anything that you would find in a spreadsheet or a database. As it turns out, the vast majority of data (and its potential value) is unstructured and lives in documents, images, audio, or video.

How we deal with this unstructured data has changed over time. Traditionally we relied on template-based methods, where handwritten rules were crafted for each document type. These rules proved reliable as long as the documents were, which only worked for well-structured documents like tax forms that change at the pace of tax codes.

More recently, deep learning techniques have gained popularity to address variability and complexity in content. Some of the models used for this task have been pre-trained on millions of pages, which makes them good at many things but not great at anything in particular. In other words, they are able to comprehend any data, as long as they are fine-tuned on relevant training examples. These examples have to be created by humans who laboriously teach models complex tasks, like processing invoices. If the task changes, those same humans have to go back to the classroom to teach a new generation of models.

Fast forward to today, the latest innovation in Large Language Models (LLMs) promises yet another step-function in performance. GPT-4, trained by OpenAI, is several orders of magnitude larger than previous models, and can be said to be great at many things out-of-the-box (it passes most standardized tests and even the bar exam). These models exhibit a remarkable aptitude for comprehending the nuances of intent and the intricate relationships embedded within text. The implications are twofold: (1) Natural language is now a viable form factor for most products. You can express your intent in natural language and GPT writes code, SQL queries, poems, emails, even entire books. (2) For most use cases, GPT is able to answer content-related questions without any fine-tuning, eliminating the need to laboriously create example data. The value of applying GPT-like models to content hasn’t gone unnoticed by OpenAI’s competitors, who are launching models that are better suited to skim through thousands of pages.

As a result, we expect a tidal wave in business value coming from unstructured data. Enterprises used to have to pick just a handful of use cases where the laborious task of fine-tuning models was justified by a costly-to-create business case. With GPT-like models, new products will emerge that are both easier to use (because of natural language interfaces) and more powerful (because they can comprehend all of your content out of the box). This will transform legal processes, streamline medical practice management, accelerate financial analysis, expedite mortgage approvals, and more generally supersede most legacy workflows. For the first time in history, this technology will impact the day-to-day operations not only of big tech but also of traditional enterprises, small businesses, and professionals, truly ushering in the era of applied AI.

Challenges and Looking Forward

As businesses look forward to applying these technologies, they should keep a few things in mind to truly derive value:

• Align AI with Business Knowledge: To implement AI in practice, it’s important that impactful applications come from the overlap of what’s possible with AI and deep knowledge about the business. Otherwise, there is a risk of numerous small AI projects that fail to deliver significant value. AI initiatives should not be viewed merely as technology center driven endeavors; they must be closely aligned with the business lines they serve and be driven out of business units.
• Assess and Optimize Current Processes: To begin, companies should start by obtaining a good understanding of their current processes. Take for example a process like customer onboarding which requires complexity of systems, sensitive data, and multiple processes. Map out such processes to identify the parts which can be automated. Next, evaluate multiple processes to identify which one would gain the most from the application of AI.
• Manage the Roll Out: Once ready to roll out the solution do so in small experiments. It's critical for the success of AI initiatives that there is no financial or reputation damage. Invest in technologies that enable quick integration of new LLM models, validations, and human checks by your teams to ensure there is a low risk of major hallucinations. Additionally, almost all serious applications of AI should have policy layers that sit on top of the AI system to enforce guardrails. For example, when you generate text, you should probably ensure that it doesn’t include profanity. Or whenever AI is used to predict important economic values like house prices, it is important to have humans review whenever the new prediction deviates significantly from a previous estimate.

As we embrace the era of applied AI, businesses that navigate these challenges with foresight will unlock the full potential of AI, driving value creation across their operations and have agility to adopt the next wave of AI innovation that enters the market. By embracing this new technology, enterprises and small businesses alike can thrive in the transformative AI landscape.

About the Author

Clemens Mewald is the Head of Product at Instabase. With over 15 years of experience in the industry, Clemens Mewald has built a successful track record as a product and technology leader in the AI and machine learning space. Previously Clemens held leadership positions at Databricks, where he spent more than three years leading the product team for Machine Learning and Data Science. Before Databricks, Clemens served on the Google Brain Team building AI infrastructure for Alphabet, where his product portfolio included TensorFlow and TensorFlow Extended (TFX). Clemens holds an MSc in computer science from UAS Wiener Neustadt, Austria, and an MBA from MIT Sloan.