Expectations about AI are so high that people are anticipating civilization-changing things from the technology. Google CEO Sundar Pichai said earlier this year that he expects AI's impact to be more powerful than the discovery of electricity or fire.
But, in reality, enterprises are tempering their expectations about the impact of AI on business operations. While Pichai's prediction for AI may turn out to be true in the long term, businesses today are taking a considerably narrower view.
In particular, experts in the field are urging others to be mindful of the limitations of AI today. Relying too heavily on the technology could lead to significant business problems.
"Right now, there is so much hype around data, but we have to be careful about how we use it," said Deeksha Joshi, managing director of corporate strategy and research at Liberty Mutual Insurance, based in Boston.
Don't expect machine learning to lead your business
In a panel discussion at the AI World Conference and Expo in Boston, Joshi said AI algorithms can uncover interesting insights in data that were previously unknown to the business. What AI can't do is tell the business what to do with these insights.
Joshi talked about how a lot of businesses today are interested in using natural language understanding to mine social media comments to gauge how the public feels about their brands. This is a powerful application of machine learning and gives businesses a way of understanding that was previously unavailable.
But it's less clear what a business should do with that understanding, an ambiguity that can limit the impact of AI on business operations.
"We can't change our strategy every time we get a customer complaint," Joshi said. "AI can help us identify patterns to inform our strategy, but it can't create our strategy."
There's still a lot that algorithms don't know
While machine learning -- and, in particular, deep learning -- has made significant progress in the last few years, it still faces a fundamental limitation in training data.
Algorithms need to be shown hundreds or, in some cases, thousands of examples of a phenomenon before they can make predictions. However, if business leaders don't understand the impact of an event because it's rare, AI will be similarly powerless.
"There are certain myths and inconvenient truths," Anthony Scriffignano, chief data scientist and senior vice president at Dun and Bradstreet, based in Short Hills, N.J., said in an interview. "One of the myths in AI is that you can AI your way out of something you don't understand."
Anthony Scriffignanochief data scientist and senior vice president of Dun and Bradstreet
He used the example of a supply chain company trying to predict the impact of Great Britain leaving the European Union -- referred to as Brexit -- on global shipping and trade operations. If you simply ask a machine learning algorithm to foresee the impact of Brexit, it will be unable to return a useful answer, because we've never seen a comparable situation before. The algorithm has no meaningful data on which to train.
But if the supply chain company were to ask about specific conditions within the supply chain and how the firm could respond to specific changes, there's more likely to be relevant training data, and machine learning can help. In this way, the impact of AI on business operations is more tangible.
This type of specificity is still important in AI. Scriffignano said many companies are looking for AI to give them answers to big, broad questions. But this approach often ends up simply reinforcing existing knowledge and misperceptions about business conditions, because the assumptions used to train the model come from humans, rather than data.
"Data science and AI can help you make mistakes a lot faster," he said.
Think big, but start small
So, while AI may produce huge changes in some distant future, businesses need to be realistic about the impact of AI on business operations. That's why Robert Bogucki, CTO at software vendor deepsense.ai, based in Palo Alto, Calif., said businesses should "think big, but start small."
In a panel discussion at the conference, Bogucki said a customer in the insurance industry came to his team and asked them for a software tool that would use computer vision to analyze pictures associated with insurance claims to see if the damage in the pictures matched up to the claims.
Deep learning algorithms can theoretically do this now, and it's a project worth considering, because it could save insurance companies substantial expense. But the application would take a lot of training data that may not be available immediately.
Instead, Bogucki recommended starting with a smaller initiative that involved analyzing metadata tags associated with images to make sure they were taken at the location of the claim. It's not as technically complicated, but it can be just as effective from a business perspective in catching some instances of fraud.
"This is something machine learning can help with right now," Bogucki said.