Gen AI Could Start Transforming Businesses in 2025

Almost two years after the introduction of OpenAI’s chatbot ChatGPT, investor enthusiasm for generative AI (Gen AI) has continued to grow with increasing focus on its real-world use cases and applications.

Technology solutions for large-scale companies – or enterprise software – has a $900 billion global available market, where investors believe that the proliferation and adoption of Gen AI could more than double the total available market to $2 trillion - $3 trillion.

On August 20, Goldman Sachs Research held a Silicon Valley AI field trip led by George Tong, head of US Business Services Research, and Kash Rangan, head of US Software Research. The team spoke with companies, venture capitalists, and academics to explore advancements in Gen AI and its challenges.

Here are some of the highlights on what could shape the future of Gen AI.

Gen AI could become table stakes 

Venture capitalists believe that Gen AI will ultimately be table stakes – a minimum requirement – for enterprises to be competitive, as customers will come to expect large language model (LLM) features in products.

Investors noted that as Gen AI becomes more ubiquitous, those who do not deploy it effectively may miss their ability to monetize on Gen AI features and fall behind competitors who ramp up their base offerings with the technology.

This is already the case for hyperscalers – the largest cloud providers that drive global infrastructure to handle enormous amounts of data and computing power. They are making a deliberate decision to invest a growing portion of their revenue into Gen AI. In their view, not doing so would risk the loss of their market share to competitors that are investing to service their customers’ higher performance computing needs.

As for other use cases, Gen AI could revolutionize knowledge work and become part of our day-to-day workflows, according to Stanford University’s Dr. Monica Lam, a professor in computer science. She believes the technology will be adopted by most people in academia as well as the public and private sectors.

Data is the most important enabler of Gen AI

AI models are trained by data, and without quality data, Gen AI would have little added value. That’s why for many investors, quality data is seen as an essential requirement for the successful deployment of Gen AI.

Investors believe industries that closely guard data or keep them behind paywalls (e.g., financial services) may face challenges in implementing Gen AI unless a third party agrees to license out the data. The exception would be if the enterprise owns all the proprietary data in-house and does not need external data to train its model.

While training data can be difficult to discover, some companies discussed methods of deploying proprietary code to scrape and process public data to train AI models.

One company noted that having proprietary data to train a LLM would give an enterprise a competitive advantage over models trained on commoditized or publicly available data.

Costs of AI expected to come down as it becomes more efficient

A byproduct of Gen AI deployment is the amount of energy it consumes in order to retrieve answers to each user query. Many models call on LLMs to generate a response, known as inferencing, and the provider would typically charge by a certain number of queries (inference cost). This cost includes the consumption of computing resources such as energy, data, and memory every time an LLM is given a prompt to produce an answer.

One company believes that inferencing costs will come down over time, as closed-source foundational models (where the source codes are not publicly available) face growing competition from open-source models (where the source codes are publicly accessible). 

Gen AI struggles with creative work but is efficient in rote tasks

Gen AI is currently most applicable to work that involves more rules-based tasks such as legal, tax, and accounting and less helpful when it comes to more creative or subjective tasks. For fields that tend to require more labor-intensive rote tasks and have a greater need to find human capital efficiencies through technology, Gen AI can drive efficiencies with document drafting, performing legal research and summarizing complex documents.

Investors believe Gen AI’s potential growth to become a $2 trillion - $3 trillion available market will be largely driven by applications that make more efficient use of human labor. 

Hallucinations are still a key risk to LLM accuracy

Based on academic studies, ChatGPT-3, the third version of a transformer LLM developed by OpenAI, was found to be accurate only two-thirds of the time. Note, GPT stands for generative pre-trained transformer and it’s the driving force behind the chatbot. Stanford University’s Dr. Lam said it’s challenging for Gen AI users to detect LLM hallucinations – content that is inaccurate, nonsensical, or irrelevant to the user's prompt – because we just don’t know what we don’t know.

She noted LLMs are prone to hallucinate since they apply only speech rules, instead of actual knowledge, when generating answers. For instance, a typical model responds with the next word that has the highest probability of occurrence based on the data it’s trained on.

Dr. Lam believes hallucinations could be minimized by applying a human approach to answer questions, which involves filtering out irrelevant data, checking whether each claim is true, and then summarizing those claims.

What’s next for Gen AI in 2025

Investors told Goldman Sachs Research that more consumer applications would be developed and coming to the market in 2025 – namely those focused on sales and customer service. However, they also flagged a potential risk that one of the top open-source LLMs might face scrutiny from misuse next year.

Goldman Sachs Research noted from their meetings in Silicon Valley that nearly all enterprises are aiming to put Gen AI into production in 2024 and that the integration of Gen AI in the operational workflow could begin to occur in 2025.