After several years of marketing hype over artificial intelligence (AI) technology, it turns out a major challenge in deploying AI successfully is managing customer expectations inflated by that hype.
At last week’s RISE conference in Hong Kong, a panel dedicated to ‘unlocking the true potential’ of AI generally agreed that while there are numerous challenges to AI, a key obstacle is cultivating realistic expectations about what AI can and can’t (and won’t) do.
A common example, said Ping An Technology CEO Ericson Chan, is the age-old discussion about whether AI will replace humans in the workplace. Taking healthcare as an example, the fact that machine learning can scan images to diagnose lung cancer with around 95-96% accuracy – which is around the same accuracy as a human physician – doesn’t mean that the algorithm makes the final decision on the diagnosis, he said.
“When you have the AI technology assist the physician, research shows the accuracy to detect and the efficiency goes up to 99.5%,” Chan said. “So to set expectations properly, AI is not for replacing humans but enhancing what everybody can do.”
Winnie Lee, COO of Appier – which helps enterprises adopt AI in various applications – agreed, noting that many of Appier’s customers often look at AI as being an all-purpose cure for solving their business problems overnight, which is decidedly not the case.
“In fact, the adoption of AI is a journey – experimenting, as well as iterating based on the results that you’re seeing, and it all has to go back to the data that you have in place,” she said. “Most companies will say, ‘Hey, I don’t have enough data, but I have so many key things I want to achieve at once,’ but that’s not very realistic. What we have to do is to focus on one key challenge that you have at the moment that can actually leverage the data that you have internally. This is not going to be fixed overnight – it will be a continuous effort.”
Closing the data gap
Another key challenge with AI has to do with data infrastructure. For example, Lee said, there tends to be a significant degree of difference between the training data used to train algorithms and the real-world data that will eventually be fed into it.
“This is a key area that everyone is focusing on right now in the industry – how to design algorithms so they can be robust enough to adapt to the gap between the real-world data versus your own training data,” Lee said.
Another concern with data is, of course, related to privacy, anonymity and security, although Ping An’s Chan said this isn’t that big a concern in the sense that sensitive data usually isn’t necessary for proper AI training.
“Going back to the health tech example, where we are able to identify the lung nodules through the CT scan and X-ray [images] way, way better than physicians – when you do the training, you only look at the image of the CT scan,” he said. “We don’t need to know anything about which individuals the image comes from and so forth.”
Chan added that companies can safeguard data privacy by complying with data privacy laws, encrypting data and not sharing sensitive data with third parties.
AI pro tips
Winnie Lee of Appier offered several pro tips for organizations gearing up for AI, starting with having a clearly defined goal: “Identify a key challenge that you are facing right now, and just dedicate to solving that.”
Second: have full team support. “That’s including your management team and the people who are going to operate it day to day, as well as the people who are going to help you to build the data infrastructure.” Speaking of which, she added, take time to build out your data infrastructure to ensure that you have the data that’s relevant to the questions that you want to answer.
And finally, she emphasized the importance of understanding that this is not a one-off solution. “You actually need to continue to iterate, and even reshape both your data strategy and the ultimate goal that you’re trying to achieve, based on the results that you’re seeing.”
Chan of Ping An offered some additional success factors for AI, including a long-term investment strategy, hiring enough engineers to build and refine new models, and – most importantly – giving those engineers enough use cases to work on.
“In order to attract the top range scientists, it’s not just about the pay scale, it’s also about giving them meaningful things to do,” he said.
Be the first to comment