Microsoft's Chief Data Scientist On Ethics, Bias And Other Enterprise AI Considerations

As Microsoft’s chief data scientist, Ozge Yeloglu’s job is not to develop specific products but simply help further customer success with tools like artificial intelligence. There are some occasions, however, when she just has to say no.

Speaking at last week’s Elevate Tech Fest in Toronto, Yeloglu — who works for the software giant’s Canadian operation — recalled a request from a customer who used Microsoft’s Office 365 productivity suite to go through employee e-mails to see if they could improve retention rates.

“My reaction was, like, ‘Sorry, I can’t help you with that,” she said. “They said, ‘But it’s our data, right?” And I said, “yes, but don’t you think it would be ethically correct to get permissions from employees that your are allowed to do this?’”

The incident is just one example of what Yeloglu described as the growing interest and understanding of AI by large organizations, a process she said has particularly accelerated in the last two years.

“It was a shock coming into an enterprise world,” admitted Yeloglu, who had come from a background in research and startups. “I realized most of my time was spent on 101 conversations, starting at the very basic level.”

While corporate customers might have been nervous or skeptical about the impact of AI in the early days, however, one of Yeloglu’s main tasks as a chief data scientist today is to dispel the fantasies they may have towards the technology.

“The wrong expectation of AI is that you’re going to buy a product — some sort of magical black box — and that you’re going to plug it into your enterprise and it’s going to solve all your problems,” she said, noting that the depiction of AI in popular culture may be partly to blame. “Hollywood has never really helped us.”

Microsoft is steadily infusing AI into products Office 365, its Dynamics CRM and so on — Yeloglu said a new tool is added almost every other week — but her work is also about enabling others to build their own AI solutions. These tend to fall into one of two areas, she said: building an AI product or using it to optimize an operation.

“A product is an easier problem to solve — you just need good teams involved,” she said. “Operationalizing is about figuring out the right usage, training and comfort level.”

As might be expected, some of the key sector Yeloglu sees looking deeply into AI include financial services, government, utilities.

“It’s interesting to see how manufacturing is adapting AI super-fast compared to retail,” she said, arguing that “old school” and “plant-level” uses such as predictive maintenance should not be overlooked. “It’s about being able to see the ROI.”

Yeloglu also praised the emerging discussion about potential bias in AI, suggesting it should have started much earlier. As a chief data scientist, she said she not only needs to think about how data can be exploited, but the privacy and other issues that underlie use cases.

“There might be one data source and the bank owns it, like their customer account data, for example. It starts to get blurry if they want to use that data with third-party loyalty program data,” she said. “It’s a question of, are you allowed to do that or not? At that point we trust the customer, but we still ask. If the answer is, ‘I don’t know,’ my response is, ‘Can you please go check that?’”