Here are four takeaways for finance teams on AI implementation.
Focus on specific use cases and prove ROI before making large investments
There’s a considerable amount of pressure on organizations to invest in AI even when projects lack clear frameworks, So said. Child recommended trying “some small use cases” and proving return on investment first.
“Picking specific areas is probably the best approach. Put something out there, test it, figure out where it breaks,” and go from there, said Child, who offered an example from Arm Holdings: About 60% of Arm’s revenue comes from royalties on its technology, which are paid on about 8 billion chips per quarter. The company previously forecast royalties using Excel and other tools but now uses AI to build those forecasts.
“It’s really using the same approach that’s been around for quite a while. It just might be able to be done a little faster and a little easier, at a little lower cost and maybe with a simpler application,” Child said.
Understand the difference between automation, machine learning, and generative AI, and match the appropriate technology to your business problem
Hoffmeister said there’s no one-size-fits-all approach that companies can use when choosing an AI technology. What works in engineering won’t work in compliance and vice versa.
“There are so many different applications and use cases out there. I don’t have uniformity of any one specific application in any of my departments, with maybe a few exceptions,” Hoffmeister said.
Child advised finance professionals not to underestimate the power of simple automation. “Especially in the accounting world, a lot of the plumbing ends up being [improved by] automation. We found that with pretty good automation technology, you can basically solve a lot of those types of things,” he said.
The benefits from generative AI can be less straightforward for finance. “LLMs are probabilistic. Finance is deterministic: There’s an answer. An LLM is going to give you the highest probability of what might be the right number, but it’s not going to be exactly the right number,” Child said. “So I have to make sure I understand exactly how today’s LLMs work and where I can trust it and where I can’t. I think of today’s LLMs as mostly a first draft, but it can never be the answer.”
Ensure that your data is of the highest quality before deploying more advanced AI and machine learning tools
Good-quality data is crucial for AI and machine learning to be successful, but it doesn’t come easy. MIT Sloan associate dean for online education and artificial intelligence Dimitris Bertsimas has estimated that 60% to 80% of time spent on a data analytics project will be devoted to acquiring and cleaning up data.
Before doing anything else, ask yourself, “How clean is your data? Do you even have the capability to do just basic [robotic process automation] or comparing one spreadsheet to another?” Child said.
“I find that getting that clean data layer is really hard,” he said, but then “you can start to do the automation layer, you can do the machine learning — you can do a lot of different things.”
Hoffmeister acknowledged that CFOs often have to make business decisions with incomplete or noisy data or perhaps uncertainty about the future. “Every decision is obviously done at a point in time with a certain amount of information and a certain amount of context,” he said. “And almost definitionally, the very next day, that context changes.”
Whether in three months or three years, finance leaders should revisit their decisions down the road as both data and context evolve, Hoffmeister said.
Harness creativity and novel thinking when deploying AI
Experimentation among knowledgeable users is crucial to getting the most out of AI.
So said that AI experimentation aligns with an economics concept called “user innovation.” “A lot of the best technical applications are not developed in the lab [but] rather by individuals who know their particular circumstances and problems and what looks like a good outcome,” he said.
Often, that creativity lies in younger employees, provided that they have baseline skills and a respect for data compliance.
Child said that younger employees are good at using AI to solve problems. “When you ask someone who is under 25 [about AI use cases], you get incredible creative approaches that are really mind-blowing and are pretty novel,” he said.
But Repenning cautioned that an overemphasis on “the under-25 crowd” might lead to a workforce that lacks important skills that employees had in a less-automated world, such as critical thinking and judgment.
“At [MIT Sloan], we teach students how you can do basic finance models or system dynamics models, knowing full well that most of them are not going to be professional modelers when they grow up,” Repenning said. “I always justify this by saying, ‘We want you to be educated, hard-nosed model consumers when you leave here, and you need to know a little bit about what’s under the hood in order to ask the right questions.’”
