This is the web version of Eye on A.I., Fortune’s weekly newsletter covering artificial intelligence and business. To get it delivered weekly to your in-box, sign up here.

What’s the new normal going to look like? It’s a vexing question for data scientists and engineers working on artificial intelligence systems.

So many of these systems are designed to learn from historical data. But, as I’ve noted in previous Fortune Eye on A.I. newsletters, when the present stops looking like the past, these A.I. systems can get into serious trouble.

What’s more, it’s unlikely that the future—when lockdowns ease and business resumes, but perhaps with some social distancing and travel restrictions still in place—will resemble today, either. So how can A.I. software be trained to function correctly with all this changed data?

I put this question to Ahmer Inam, the chief A.I. officer for Pactera Edge, a Redmond, Washington-based technology consulting and services firm that spun out from Chinese IT company Pactera Technology International in January. It helps businesses, including many Fortune 500 companies, implement A.I.

Inam answered that one way to make sure an A.I. system doesn’t run amok is to ensure a “human-in-the-loop” is always reviewing its recommendations and making final decisions. Today’s A.I. systems function best when supporting human decisions, he says, not fully automating them.

Humans should also keep a close watch on the data that A.I. software is being fed. That’s important even when there isn’t a pandemic. “Model drift,” in which data gradually changes over time, is “a common problem even with standard business data,” he says. “The dynamism of the business itself causes that.”

But another possible solution is to use a different kind of machine learning altogether. Instead of supervised learning, in which an algorithm learns from historical data, Inam says businesses could use reinforcement learning, in which an algorithm learns from experience, usually in a simulator.

This is the kind of A.I. that researchers have successfully used in the past five years to teach software to beat humans at games, such as Go and poker. But businesses have been slow to adopt these techniques for a number of reasons.

Building a reliable simulator can be expensive and time-consuming. The level of machine learning expertise needed to successfully design and train a reinforcement learning algorithm tends to be greater, and that talent is in short supply. The amount of computing power needed to train an algorithm using reinforcement learning can be massive and costly too.

Inam says the expense and time, however, are well worth it. Using a simulator, you can expose an algorithm to a wide range of potential scenarios and teach it how to handle them all. The result is a much more robust A.I., ready for whatever the world is going to throw at it.

Inam has built simulators to see how a coffee shop chain might be able to mitigate climate change’s affects on the reliability and pricing of coffee supplies. In another case, he simulated how a hurricane could impact an automotive retailer’s regional sales. More recently, he built a simulator to help a logistics company optimize its routing. The project took six months, he says, but it ultimately enabled the company to save millions in fuel and labor costs.

Inam thinks the coronavirus pandemic is likely to accelerate companies’ adoption of these more sophisticated A.I. techniques. “It’s fast-tracking adoption of stronger A.I.,” he says. Add that to the list of the pandemic’s unanticipated effects.


While you’re here, we’d like your opinion on how we’re doing with the newsletter. Please spare the time to answer some quick questions for us. It will take no more than 2 minutes to complete.


We take your privacy seriously. The data we collect in this survey will be used for research purposes only.

Read on for the rest of this week’s A.I. news!

Jeremy Kahn

This story has been updated to correct the relationship between Pactera Edge and Pactera Technology International.