Generative AI can affect managerial decision-making in “a transformative way” by boosting value generation, according to Prasanna (Sonny) Tambe, Wharton professor of operations, information and decisions. Tambe is also faculty co-director of AI at Wharton, which fosters AI activities across the University of Pennsylvania. He was speaking at a conference hosted jointly by Wharton’s Mack Institute for Innovation Management and AI at Wharton in November 2023, titled “Driving Innovation with Generative AI: Strategies and Execution.

Its unique strengths in translation, summation, and content generation are especially useful in processing unstructured data. Some 80% of all new data in enterprises is unstructured, he noted, citing research firm Gartner. Very little of that unstructured data that resides in places like emails “is used effectively at the point of decision making,” he noted. “[With gen AI], we have a real opportunity” to garner new insights from all the information that resides in emails, team communication platforms like Slack, and agile project management tools like Jira, he said.

Those insights will be helpful in a variety of ways, such as more accurately predicting delivery times for say, software development projects, Tambe said. In recent work, he and his research colleagues found “enormous potential” in one specific use case, where they processed raw patent texts and gained more accurate “blue-ocean” insights than was previously possible. They brought a superior understanding of “where firms are innovating, and where there’s room for an entry-level firm to innovate,” he added.

“With generative AI, one important use case is to take these millions of documents in any context and try to boil them down into a small set of factors that managers can understand,” Tambe said. “Generative AI tools can be used to create intuitive answers to questions, and the technology is better at representing ideas in a way that’s intuitive for people to understand.”

“Generative AI tools can be used to create intuitive answers to questions, and the technology is better at representing ideas in a way that’s intuitive for people to understand.”— Prasanna (Sonny) Tambe

For enterprises, gen AI’s power in providing personalized learning will “fundamentally allow people to learn on their own terms, and meet them where they are,” said Scott Snyder, a senior fellow at the Mack Institute and chief digital officer at EVERSANA, a provider of commercialization services to the life sciences industry. He shared those perspectives as he moderated a conference panel that delved into how businesses can leverage large language models (LLMs) using their proprietary data for training and fine-tuning commercial and open-source foundation models.

”As a digital leader, you’re always looking for the burning platform, and we had it handed to us with the pandemic,” Snyder said. “It forced us all to operate completely differently as companies. All of a sudden we were distributed virtual companies.”

“I see gen AI as the same kind of burning platform,” Snyder noted. “In fact, it’s caught the attention of executives like nothing I’ve ever seen. Eighty percent of executives surveyed now say this will impact their company and industries significantly, but only about 50% think they have the capabilities to fully realize its potential; 92% of Fortune 500 companies are doing something or building something with OpenAI’s ChatGPT. Now everybody is a data scientist in some ways.”

Gains in Strategic Planning and Customer Service

Gen AI can help enterprises become more efficient strategic planning in new ways. Gen AI’s ability to process millions of text documents also helps identify “actionable factors” for organizations, Tambe noted. For instance, it could help companies analyze competition dynamics in their industries and plan on allocating their resources and investing, said he said. Or, it could find uses in performance reviews and instilling corporate culture.

“If you want to take 30,000 performance reviews every year over 10 years and boil it down to a small number of factors that people most care about at your company, such as culture or fairness, what are those few things?” he asked. “How can you boil information from say, thousands of customer service conversations, down into an actionable number of factors? Gen AI can help us distill all that data and represent it back to decision-makers in a way that they can start to act on it.”

Enterprise-level learning is another area where gen AI has big promise. Chris Callison-Burch, professor of computer and information science at the University of Pennsylvania said, “These pre-trained models can do amazing work with learning.” As it happens, his research areas include natural language processing, from where sprung large language models.

Callison-Burch pointed in particular to a feature called RAG, or ”retrieval augmented generation,” which allows users to post web queries to retrieve information and summarize it. Enterprises could also use that tool to upload their internal documents and index them for retrieval via semantic search. “Those are super exciting,” he said.

“[Gen AI] is about marrying the AI and the humans, and the companies that figure out how to unlock that are going to get there the fastest.”— Scott Snyder

A Measured Adoption Curve

Businesses are not rushing in to use gen AI, and their adoption curve is dictated by the risk sensitivity of their activities, among other factors. Avi Patel, chief marketing officer and chief data scientist at Fulton Bank, said companies are experimenting with gen AI, but at a measured pace. “Companies should stay current with gen AI and learn what works and what doesn’t work for them.”

In especially tightly regulated industries, companies will try out gen AI based on the risk sensitivity of their activities, Patel continued. For instance, they might begin by experimenting with gen AI in relatively lower-risk tasks such as document summarization, which would enable their teams to be more effective in their daily jobs, he said. One concern would be the risk of sensitive documents getting leaked out in the process of summarization with third-party tools, he explained.

Other early adopters of gen AI are focusing initially on activities with low complexity, such as Automation Anywhere, which provides automation services to businesses. Tejasvi Devaru, vice president of business applications and data at the firm, is encouraged by some early success with gen AI. His firm had rolled out more than 20 use cases in the six months prior to the conference. In one case involving robotic process automation in the customer service area, his firm was able to automate 60% to 65% of workflows, which freed up the team to focus on escalated emails and provide better customer service. That amounted to savings of nearly 10,000 hours, he said.

In another instance, Devaru’s team tapped GPT to extract specific information from purchase orders to ensure accuracy between sales orders and customer purchase orders. It allowed them to extract “structured information from unstructured documents such as purchase orders,” he said, noting that it was challenging to sift through product information and other details for more than 20,000 purchase contracts with each customer having a different format. Traditional methods were too expensive or time consuming. Devaru’s team is using GPT to process information for about 80% of the purchase contracts, which happen to be relatively less complex. But that shift to GPT is already making a big impact in improving cash flows by about $850,000, he said.

Automation Anywhere is counting gains also in its customer service “deflection rate,” which is a measure of customer support requests that are resolved through self-service mechanisms like chatbots and tutorials, without human intervention. “Right now we have a 30% deflection rate, and we want to increase that to 60%,” Devaru said.

One big challenge for users of gen AI is its so-called “hallucination” problem, where inadequately trained data can produce output that is inaccurate or biased, and does not match real-world settings. “It becomes a problem if you want to solve a business case that requires higher accuracy, and having a human-in-the loop in these scenarios is useful,” said Devaru.

“Companies should stay current with gen AI and learn what works and what doesn’t work for them.”— Avi Patel

Early Questions Facing Gen AI Users

Businesses that want to use gen AI will necessarily have to make some choices based on their specific requirements, Devaru noted. One is to pick the technology that works for them from among the roughly 17,000 large language models that currently exist; ChatGPT is only one of those. “We need to think about what the business use case is and which language model to use for that,” he said. Some, like Google’s Bard, are especially useful in dealing with security threats, while others like OpenAI Ada are good at summarizing documents, he added.

Another question for business users is to decide whether they should use a public model like ChatGPT that is on the cloud versus using an in-house model. Even if a company were to use a public model, it could incorporate security features such as ensuring that its proprietary data is not used by its gen AI provider to train language models, or anonymizing its information before sending it to the gen AI provider, Devaru said.

As companies get more and more comfortable with gen AI and begin to see tangible gains, they would use the technology for higher-level or more sensitive activities. “[For now], companies are likely to think about very, very low-risk items,” Patel said. “But the biggest impact will be when companies use their tabular data and the power of context learning in large language models to understand risks relating to customers, or their likelihood of purchasing the next product,” he added.

Another gen AI feature that Devaru is excited about is the ability to translate from conversation or text to SQL (structured query language), which allows access to databases. “The use case that we are thinking about is exposing a conversational user interface to our leaders where they could get responses to questions like ‘What’s our sales data for the last quarter? Or what are our biggest deals in a quarter? How is it trending?’” he said. “That’s the power we want to unlock.”

Snyder, whose company EVERSANA is in the life sciences industry, sees even bigger possibilities ahead. “There are so many that I get excited about, like giving a voice back to people that have lost it because you can now generate it from their previous history and conversations. Or sight,” he said.

“Ultimately, I think of AI not as artificial intelligence, but augmented intelligence,” Snyder said. “It’s about marrying the AI and the humans, and the companies that figure out how to unlock that are going to get there the fastest.”