Why the industry needs to trust in AI algorithms

Managers, executives and boards in the insurance industry need to trust in artificial intelligence algorithms, even if they don’t look like traditional models, an Accenture executive told Canadian Underwriter Thursday.

Machine learning models can better predict events with a greater degree of accuracy and a higher assigned level of confidence than with traditional analytical techniques. However, insurance is a very old industry and those in the industry may have a hard time trusting a new algorithm because it looks different than the results and predictions of the past.

As an example of the difficulties in getting management and board buy-in, Jodie Wallis, managing director for artificial intelligence in Canada with Accenture, provided an example of speaking to a data scientist from Shopify. The ecommerce provides cash advances to merchants. Shopify built machine learning models that are excellent predictors of likelihood to pay back in the case of a cash advance.

However, the models look completely different from those of traditional lenders because they have no information about assets and liabilities. “What they do have, though, is a whole bunch of transactional data that demonstrates what the cash flow is for the individual merchant,” Wallis said in an interview Thursday.

The machine learning models were originally taken to the company’s chief financial officer, who noted they didn’t look like any model he’s ever seen. “They proved over time to the CFO that, in fact, with some back testing, these machine learning models were excellent predictors of risk. So, the CFO said, ‘Let’s take it forward.’”

There was the same discussion at the board level, where they asked the CFO how he know the model was the right predictor of likelihood to pay back cash advances. “So, every time it went up a level, the story had to be retold,” Wallis said.

People are generally comfortable with the way things have been running and are unlikely to feel compelled to make a change. “If you don’t trust in the results and you’re not acting on them, the best algorithms in the world aren’t going to help.”

To change requires a cultural shift, at both the individual and management levels. “Managers, executives and boards need to have trust that the results of these new machine learning algorithms are going to yield better outcomes for the business,” Wallis said. “That is something that is culturally not that easy to shift.”

Besides an awareness and education component, buy-in can be achieved by “back-testing” AI. “You can take your models and apply it to historical data and see what the results would be, and compare it to actual results in order to give yourself a level of confidence it will yield equal or greater results. Part of it is awareness and education, and part of it is really taking the data and proving it. It is possible to prove that there’s benefit here.”

Wallis discussed the challenges of incorporating AI in the insurance industry days after Accenture launched the second season of its podcast series The AI Effect. The seven-episode series, including an episode on insurance, was hosted by Wallis and technology journalist Amber Mac.