Lessons from ChatGPT: How Automation Can Create More Impactful AI

As part of Solutions Review’s Contributed Content Series—a collection of articles written by industry thought leaders in maturing software categories—David Barber, a Distinguished Software Engineer at UiPath, outlines some of the lessons regarding automation and AI that companies can learn from ChatGPT.

AI is already within most of the software technology we use, whether it is obvious or not. With OpenAI’s ChatGPT and Bing’s GPT-4 taking the visibility of AI to new heights, leaders are preparing for a new era of AI in the enterprise.  

For much of the recent past, access to truly powerful AI solutions was restricted to advanced data scientists and engineers—and that’s because AI is not easy. The technology is difficult to train, expensive to maintain, and daunting to scale. On top of that, the skills gap to achieve game-changing AI was too high and too costly for most companies to attempt to tackle, especially building large language models (LLMs), the technology on which ChatGPT and GPT-4 are built.  

GPT heralds an era in which it is possible to democratize AI, making it accessible to every employee—no technical ability or special training needed. To take advantage of this new era, companies should consider the power of automation. Not only does automation help create low-code/no-code solutions that open access to LLMs, but it also can ensure AI technology is implemented responsibly within an organization—helping companies generate real business value faster. 

Understanding Large Language Models 

LLMs have revolutionized the ability of machines to understand human language. AI models like ChatGPT and GPT-4 can generate complete, almost human-like text responses to questions and inquiries that have fascinated users and engineers. And although these models have opened countless new AI use cases across the enterprise, CIOs and their teams must understand how LLMs work to determine how best to incorporate them into their digital strategy.  

LLMs are extensive deep neural networks trained by going through billions of pages of material in a particular language while attempting to execute a specific task, such as predicting the next word or sentence. This creates networks sensitive to contextual relationships between the elements of that language (words, phrases, etc.), allowing the models to help computers understand and respond to nuanced human conversations and questions. However, LLMs are still not entirely accurate, with some models operating anywhere between 65-90 percent accuracy. 

Nevertheless, companies are still starting to use LLMs as foundational models for a wide range of AI use cases. LLMs generate valuable insights from substantial language datasets. They can perform various tasks, such as accurately extracting emotions and important data—like dates, order numbers, and addresses—from mass communications like emails and IT tickets. But using LLMs to extract intent or import data is just the first step in a very long training process, and LLMs must be worked with carefully to be helpful in their total capacity. 

The idea is to create a continuous feedback loop with humans at the center. The more subject matter experts (SMEs) train the AI, the better it becomes. Eventually, with LLMs and active learning, models can be trained to be accurate enough for business automation, helping further drive transformation without human intervention. 

Deploying Automation  

Together, AI and automation are more powerful. While automation software has historically been used to complete routine and repetitive work, with AI, it can do a lot more. These two technologies can read documents and emails, analyze language and images, and understand the intent and content of communications. They can also be combined to answer deep domain questions, translate languages, comprehend and summarize documents, write stories, and compute programs. Ultimately, AI helps automations perform cognitive tasks, navigate uncertainty, and resolve inconsistencies, and the more that these automations can think and understand on their own, the more they can do, the faster they can do it, and the more significant the impact they can make.  

For example, simple automations could be an automatic email reply when on vacation or out of the office. However, what if someone asked a specific question in their email and they cannot wait for the recipient to return?  

With automation and LLMs, a custom response to the inquiry can be created. The solution can interpret the incoming email, tailor a custom response to any necessary questions, and even direct the individual to relevant resources—all customized and tailored to the specific scenario within seconds of an email coming into the inbox. Automation can also be created to have a person assess the accuracy of the response before any communication is sent.  

When automation and LLMs work together, they can function as virtual assistants—interpreting reports in everyday language, translating software from one coding language to another, and creating infinite possibilities of where the technology can help. Not only will this save employees time doing mundane tasks that clog up the workday, but it will give employees time back in their days to do more satisfying work, leaving time for creativity, brainstorming, and learning. Automation already saves its users from countless hours of work, but imagine what could be accomplished with an LLM, too.  

Business and technology leaders must understand that on their own, AI technology, like an LLM, has limits in its usefulness. It’s only when an LLM is embedded in a platform that organizations realize its total value. For example, an LLM that responds to a customer email is only somewhat helpful. Instead, an LLM within a platform can read databases, interact with CRMs, interact with other apps such as Salesforce or SAP, and then take action based on a holistic sense of the inbound request and next steps. The platform is the key to unlocking enterprise value, particularly as LLMs and open-source AI becomes more generally available.   

Eventually, this technology can even take on a customer-facing role, guiding consumers to the products and services that best suit their needs, answering their questions, and automatically deploying custom solutions that best fit them. However, for LLMs to reach this point where they can seamlessly work adjacent to automation, the AI must be dependable, balanced, and fair in its predictions and decision-making. 

Using Automation for Model Validation 

Model validation can help companies reduce biases in their AI, making for a more reliable product. Unfortunately, model validation is complex and takes time, and AI stakeholders often struggle to get executive support for responsible AI practices. According to the Massachusetts Institute of Technology, only 52 percent of companies practice some level of responsible AI, and 79 percent of those companies say their implementations are limited in scale and scope. But if we skimp on model validation, how can we have AI that we can trust? 

With the help of automation, users can train software robots to identify and detect potential biases in model training and rate the accuracy of model predictions. Automation software can even guide users to the exact actions they need to take to solve the model’s shortcomings. With a low-code, no-code automation platform, users do not have to dive into the code or the model’s data, saving time on wasted false positives or negatives. Automation allows the user to be guided through the complete AI training process from end to end. Eventually, automation will take over the LLM training process completely—keeping humans in the loop to sense check and ensure the AI is doing its job safely and appropriately.  

There’s a lot to be learned from this ChatGPT craze—the most important is that we need to make AI stronger, safer, and more widely accessible with the power of automation. Eventually, LLMs will be a seamless part of every enterprise, working alongside employees to deliver better customer experiences, increase sales opportunities, and improve operational inefficiencies. And companies with automation as their foundation for innovation will be the ones to lead the charge.