Raise your hand if you’ve ever felt personally victimized by Regina George the process of implementing GenAI tools at your company.
If you’ve raised your hand, you’re not alone. According to a recent report from AI company ABBYY, almost one-third (31%) of companies admitted training their GenAI models proved harder than they thought.
Companies admitted other challenges when implementing GenAI tools, including not having talent with the proper skillset (29%), tools unintentionally increasing time to complete tasks (18%), and hallucinations (16%).
ABBYY’s report is based on an Opinium Research survey that queried 1,200 senior managers across industries based in the US, UK, France, Germany, Australia, and Singapore.
Trainwreck. Markus Demirci, CEO of Rollio AI, told IT Brew part of the reason some companies struggle with training models is because of poor data quality, noting that data may be coming from legacy systems. He added companies that want to run “proper AI” and autonomous agents need to first provide them with context, circumstantial information that allow LLMs to make more informed responses.
He compared the task to teaching a child how to communicate their needs.
“When you have a very small kid, they cry when something’s wrong and once they grow older, they learn how to speak a language, and then they can communicate with you what’s really wrong,” Demirci said.
“At the beginning, the baby would just cry and that’s basically where we are with most of the companies and their contextual data they have available,” he added.
Top insights for IT pros
From cybersecurity and big data to cloud computing, IT Brew covers the latest trends shaping business tech in our 4x weekly newsletter, virtual events with industry experts, and digital guides.
Brian Shannon, CTO at cloud-management company Flexera, said difficulties during training can also occur because most LLMs are generic and meant to solve general problems as opposed to an organization’s specific needs and domains.
“They’re trying to solve general-purpose problems,” Shannon said. “But, most of us in the enterprise, we exist because we solve some particular problem in our industry.”
In the end, it was all worth it. While the senior managers fessed up to several hiccups when deploying AI, most believed the eventual comeback was greater than the initial setback. More than eight in 10 (82%) said they were either very satisfied or fully satisfied with the output of their GenAI tools within their company.
Just keep implementing. For companies that are in the middle of implementing GenAI tools and aren’t yet seeing the value of their AI investment, Shannon advised them to set measurable outcomes so that they can have accurate expectations of what to expect.
“For example, I might say, ‘Our goal is to increase revenue by 10% on this product line by introducing AI’…as opposed to, ‘Let’s use AI,’” Shannon said.
And in cases where those outcomes aren’t met by a set timeline, Demirci said not to fear: Those failures can be informative, and give you the opportunity to tweak an AI project.
“It saves you from investing 10x into it instead of just x.”