Artificial intelligence (AI) can produce surprisingly accurate analyses and predictions, but these make little sense if companies don't act on them.
“Unless you're actually having trouble implementing it, AI is just a party trick and is very impressive, but it doesn't help anyone,” said Daniel Saksenberg, Head of AI at Machine Learning Company Emerge ML.
He was the keynote speaker at the “AI in the Financial Services Industry” event held in Johannesburg on July 24th.
According to Saxenberg, the topic ranged from “from data to value: unlocking AI in financial services,” but there are two main reasons why AI can provide the value that financial institutions want and even damage losses. “One thing is that they are interesting but not valuable. The second thing is that the output of AI is merely information. You need to act on this information.”
Saksenberg, an actuary, academic instructor and entrepreneur who has worked with AI for 23 years, cited an example of an AI model built for South African insurance companies that have lost their clients. Looking three months away, the model predicted which customers would cancel their policy.
“It was horrifyingly accurate. We're talking about 90% accuracy.
Mostly to 1,000 people, and almost all three months left,” he said, adding that this accuracy had not left the insurers better because they had not taken steps to act on the insights they produced.
“It's important to devise interventions to operate your data,” he said, providing other more successful examples of the benefits AI can bring to financial services companies that effectively use AI Insights.
Success and fighting fraud in credit applications
In one case, the AI-based credit application scorecard reduced the default for South African banks by 42%, increasing the volume of their business by allowing high-credit customers that would have been screened previously. “Previous credit scorecards were dull,” he said.
Another very exciting area of AI in financial services is fraud prevention and detection. Traditional rules for fighting credit card fraud are often failed because they are designed by “honest people who don't think they're criminals” and are rarely updated. Furthermore, the vast amount of transactions that have been accidentally flagged as fraudulent outweighs the ability of financial institutions' call centres to follow up flagged purchases.
In contrast, machines trained to recognize fraud in certain South African banks could touch 94% of actual fraud cases compared to previous 20%.
Furthermore, machine learning models cannot be criticized, so even if employees are paying fraud syndicates, they cannot grasp the rules the machines are applying. “Machine learning models are very complicated. They turn into black boxes. Building these models doesn't know what the machine learned at the end of the day. That's surprising many people, but in the case of scams, this is an incredible advantage.”
Why hallucinations occur
He also described the phenomenon of AI hallucination and mentioned the generation of incorrect or manufactured AI outputs.
Hallucinations of AI occur when machines are instructed to provide answers that are not informed, Saksenberg said. Machine learning models need to give answers when instructed to do so, he noted, adding that the best way to ensure that a machine does not hallucinate the answer is to provide data directly from the source.
He briefly spoke about the risks of AI. “I don't want to give the impression that all AI is rainbows and unicorns. There's a real risk,” he said.
On the job issue, Saxenberg said that while jobs are lost to AI and new jobs are born, these are usually not available at the same time as the job is lost and are not necessarily the same type of work.
Despite the challenges, he said it was “unthinkable” to continue having a business that doesn't focus on AI. “People who don't want to embrace AI will become relics of the past.”
Dealing with resistance
Yet, large financial institutions, especially face major challenges when introducing AI into organizations, including the phenomenon of “corporate antibodies.”
In his experience, the most effective way to implement AI at a financial institution is to carve out a dedicated team of business decision makers, data decision makers, and operational staff responsible for implementation. Starting with low-bending fruits, the team needs to incubate the innovation, release it to the organization, and advocate for its success.
Saksenberg warned that it is important to get the support of employees who are expected to carry out the work. “You can't preach from the high that COOs want to implement this. You need buy-in from the people on the ground. You need to bring people along.”
