t was just a few months ago that NRF looked at AI in retail, questioning whether 2020 might be the year that it finally took off. How quickly things can change.

Artificial intelligence and its companion, machine learning, have been upended along with every other aspect of retail. And while AI’s ability to anticipate the future might have been damaged by toilet-paper hoarding, bread-baking shoppers, it also might provide an unexpected roadmap for the future.

Machine-learning based AI has been “thrown for a loop by the coronavirus,” says Nikki Baird, vice president of retail innovation at Aptos. “We’ve seen some emerging examples of companies that have had to turn forecasting AI off, because it was not capable of understanding and incorporating the complete shift in consumer behavior and the impact on the supply chain. When you’re relying on any kind of prediction model, it will run out of data that it can use to make a prediction.”

Still, Baird believes the anomalies will quickly cycle out. “We’ll start getting back to predictability, hopefully soon. But there’s an implication to it and one that we’re all trying to figure out. One of the biggest challenges of machine learning is there isn’t a good mechanism for what to do if the AI learns the wrong thing.”


Investments in AI certainly will be impacted by budget cuts in 2020 and “deprioritized,” says Sucharita Kodali, vice president and principal analyst for Forrester. “The pandemic is proof that AI doesn’t solve everything, doesn’t predict everything, can’t save you from catastrophe.”

Understanding nuances

One of the biggest challenges was the unexpectedness of the pandemic, Baird says. She points to Texas grocer H-E-B, which “has had a lot of experience with emergency situations like hurricanes and flooding. They said they felt they understood what was going to happen, but they didn’t predict the toilet paper issue.”

There are key differences between a hurricane — in which people evacuate — and a pandemic, which causes stay-at-home orders. But will AI understand those nuances and apply that learning to future disruptions?

“If machine learning believes that in March of next year, everybody shuts down, the conclusion that AI will make is to stock up on toilet paper,” Baird says. “We have to take actions to help AI learn that this is a one-time event. There’s not a mechanism to do that.”

Even though the industry has been talking about AI for a long time, she says, it’s still “a pretty immature technology and the pandemic has exposed the immaturity.”

Pointing the way out

Kodali recently examined AI and the impact of the pandemic on its use, looking at customer-facing AI, knowledge worker-facing AI and field worker-facing AI. “The biggest advantages are in knowledge worker-facing AI solutions that can help to speed up their jobs or help them observe patterns faster,” she says.

That’s exactly how AI and machine learning can lead retailers out of the pandemic, says Anil Kaul, co-founder and CEO of consulting analytics and research firm Absolutdata. One scenario is using AI to digest the wealth of information about what post-pandemic life might look like.

Kaul’s team recently gathered more than 200 research projects and let AI “weed through the documents out there to infer what these papers and articles are saying about a post-COVID consumer,” he says. “Then you’re able to generate insights and those insights can be further analyzed to see patterns.”


Yes, a human could do the same thing, but not as quickly. “From a decision-making perspective, this is reducing the time, effort and resources you would have to put into formulating a view. You can see a consensus starting to build on what the post-COVID consumer is expected to do.”

It also can return to its more traditional role of demand prediction, Kaul says. “This is where you can leverage past data. If I know that health is going to be more important, I can see which consumers are most sensitive to this topic and use that to project what the demand is likely to be. Instead of just projecting one single point, which we do in traditional analytics, we are able to project the most likely scenario.”

Baird says some are “responding to the black swan nature of the pandemic by proposing ways to train AI on the data that we have from the financial recession, the Great Depression, just trying to see if there are things they can do to help AI better weather when something goes completely off the rails versus the more typical shifts that are hidden in the noise.”

And while the data — and any AI-induced output — might be mistrusted, Greg Buzek, founder and president of IHL Group, says that can be a peril. He points to a vendor who uses AI as part of their inventory solution. “All of a sudden, the system was telling the buyers, ‘You need to buy 20 to 30 times the normal that you buy.’ They had one buyer who said, ‘This is wrong.’ And they cancelled the order. The other buyer let the order go through. One side lost all those sales due to being out of stock.”

Factoring out bad data

There is no doubt that any sales data for 2020 has already been tainted by the series of anomalies and might not be trusted. But it can help in the short term in a few key areas. Buzek believes AI can help with the coming tidal wave of returns.


“That clothing is coming back, and you have to hold it for 24-72 hours. You can’t bring that back in and put it back in inventory. If your AI was already doing this — What products are we sending back for evaluation? What are we putting back on shelves? — it can start making recommendations, where when it scans items back in, it immediately tells the personnel what to do so it optimizes the process.”

AI also can be useful to notice trends that might start to signal another immediate shift. “As a result of the peaks and troughs from the pandemic, you’re going to have systems that will learn and tweak those algorithms and look at the markers that indicated something was coming,” Buzek says.

“There were plenty of markers along the way. If we ever see them again, alarm bells will need to go out and orders changed appropriately.”

Kaul suggests using AI to create a “digital representation of whatever information you have about those customers. Once you are able to create a digital trend of a customer, you can start creating predictions based on customer behavior. Where are they likely to shop? What are they saying about what they’re likely to buy and not buy?”