Companies looking to leverage generative AI to improve productivity and revenue are struggling with expensive, immature technology that doesn’t promise a return on investment.
This is according to recent reports from KPMG, McKinsey & Company, and Goldman Sachs, which reflect the GenAI experiences of many large organizations. A key challenge with this technology is generating revenue that justifies the high costs of deploying a large-scale GenAI infrastructure.
A recent AI study by management consulting firm McKinsey found that only 11% of companies surveyed have widely adopted GenAI, Aamer Baig, a senior partner at the firm, told attendees at the MIT Sloan CIO Symposium in May, and only 15% of respondents reported improved revenue as a result of GenAI.
Cost justification is important because AI infrastructure is expensive: Nvidia's previous generation H100 AI GPUs, for example, cost around $30,000, and the numbers needed can range from hundreds to thousands, depending on the model and size.
“The technology is so new that it barely works and is very expensive,” said Anshul Chaturvedi, managing director at IT services provider Worldwide Technology (WWT).
Large-scale language models (LLMs) are currently “black boxes” that are difficult to keep accurate and ensure consistent responses, said Rob Mason, CTO of Applause, a company that tests software's AI results from a user's perspective.
“Inside the company that's developing GenAI, we're studying what we've built and trying to think about what it will accomplish, rather than what we want it to do,” Mason said.
One Fortune 100 company that used Trustbit, an AI applications specialist acquired this year by IT consultancy Time2Act Group, had to implement three different models in-house to ensure response accuracy of at least 95%, said Rinat Abdullin, a technology consultant at Trustbit.
If all three models give the same answer, it is considered accurate. If less, a human determines which answer is correct. Accuracy is very important because customers can hold companies liable if they give the wrong answer.
“Companies don't necessarily want 100% accuracy, but they want the confidence that when the model gives you an answer, if you're not sure about it, you'll know,” Abdullin said.
The shift to revenue-generating AI
Some companies are strategically repositioning their use of GenAI from employee productivity to revenue generation, according to a second-quarter survey of 100 U.S.-based C-suite and business executives conducted by KPMG. Respondents come from organizations with more than $1 billion in annual revenue.
According to a KPMG survey, in Q1, 51% of respondents said using GenAI to improve employee productivity was their most important ROI metric. In Q2, revenue generation ranked first at 52%, while productivity gains dropped to third place at 40%.
“Leaders are beginning to view investment and adoption of GenAI as a minimum requirement,” said Steve Chase, vice chairman of AI and digital innovation at KPMG. “They're now focused on how to translate these investments into competitive advantage.”
Chaturvedi, who has noticed a similar shift among WWT customers, attributes it to confusion about how to get the most value out of GenAI. He recommends starting with employee productivity and customer service applications before tackling higher-revenue use cases.
“Using this to improve employee productivity means creating a small, safe playground where if you make a mistake it's not a big deal,” Chaturvedi said.
Immaturity of the model
The hype about GenAI's business-transforming potential hides the technology's immaturity and the need for more research: Daron Acemoglu, a professor of economics at the Massachusetts Institute of Technology, predicts that GenAI is at least a decade away from being ready to dramatically change how businesses operate.
“Many tasks performed by humans today, for example in sectors like transportation, manufacturing and mining, are multifaceted and require real-world interactions, and AI is unlikely to significantly improve them anytime soon,” Acemoglu said in the Goldman Sachs GenAI report.
Acemoglu said the current architecture used for LLMs needs to be modified to mimic humans' various cognitive processes, ability to process sensory input and reasoning abilities.
“Today's large-scale language models have proven to be better than many expected, but it still takes a big leap to believe that an architecture that predicts the next word in a sentence will ever achieve capabilities as smart as HAL 9000 did in 2010. 2001: A Space Odyssey,” He said.
Antone Gonsalves is Editor-in-Chief of TechTarget Editorial, where he reports on industry trends that matter to enterprise technology buyers. He has been in technology journalism for 25 years and is based in San Francisco. Have a news tip? Email us!
