Forecasting AI Adoption in Retail: A Mixed Bag
You don’t have to look far to see the impact that artificial intelligence is having on the world around us. Across multiple facets of work and play, we’re surrounded by smart devices and applications that are strangely prescient at anticipating our wants and needs. But one industry where AI adoption has been surprisingly slow is retail — particularly around demand forecasting, where AI’s potential has scarcely been scratched.
A 2018 report from the McKinsey Global Institute concluded that AI has the potential to boost global GDP by 16% by 2030. In dollar terms, that’s a gain of $13 trillion, which is a huge number, to be sure. Companies today are scrambling to get their piece of that AI bounty – and at the same time, to avoid being devoured as the AI party unfolds.
Large retailers certainly are aware of what’s at stake. Retail is one of the largest sectors of the U.S. economy, accounting for $2.6 trillion in sales in 2016, which is nearly 15% of the country’s output. The retail sector employs nearly 29 million people directly, and supports another 13 million jobs indirectly.
Clearly, retailers are aware of AI, and have moved to adopt AI to improve their business processes. While they’ve had success in some aspects of the retail equation, few of them have figured out how to use AI to address some of more complex aspects of their businesses, such as demand forecasting and merchandise planning.
That’s the opinion of Nikki Baird, who’s the vice president of retail innovation at Aptos, a developer of enterprise software solutions for retailers. Baird has dozens of years of experience in the retail world, and has seen AI’s roll-out go surprisingly slow.
“You hear about all kinds of evolutionarily algorithms and genetic algorithms and all of the more hard-core future of AI kinds of stuff, and none of that is currently getting applied within the retail space,” Baird tells Datanami. “There’s even a question about whether some of the neural net kinds of algorithms and AI actually are applicable in the retail space.”
That’s not to say that retailers haven’t tried to adopt AI technology, or even found some success with it. Baird says the most visible impact that AI has had on the retail business is around natural language processing (NLP) technology, such as chatbots and automated customer service, as well as customer sentiment analytics, often driven from social media.
But considering that inventory is by far the biggest investment that retailers make, the implementation of a few flashy chatbots or sentiment analysis programs pales in comparison to the big game that top retailers are hunting, which is demand forecasting she says.
“It’s not that the other stuff doesn’t deliver value,” Baird says. “It’s just that there’s so much more money and so much more error that’s happening on the product side of the business, that this is where I think AI is going to get real.”
Those Retailin’ AI Blues
Retailers that have implemented AI for demand forecasting are not letting the AI forecast demand, Baird says. Rather, they’re using the AI techniques to pick the best of the traditional forecasting methods, she says.
“You’re running it through a multi-step process and having a weight come through at the end of every step,” she says. “But that’s about as sophisticated as it’s getting. It’s not actually doing a new kind of forecasting. Its’ really just deciding, here are all the forecast methods we’ve known for a really long time. These have all been around for decades. So which one is the best one to apply in this circumstance?”
Some retailers have gone a bit further and allowed AI to make the actual product demand predictions, but they have a hard time trusting the predictions. It turns out there’s a natural human resistance among retailers to letting machines make high-value decisions involving millions of dollars’ worth of inventory.
“Especially in merchandise planning, there’s a bit of a mystique around buyers and the art of retailers, picking the right products and knowing the mind of the customer,” Baird says. “For companies that are trying to bring in AI to some of those processes, they definitely run into the brick wall of user adoption, the ‘That thing doesn’t know more about my business than I do’ kind of a response.”
The black-box nature of AI models has thwarted a good percentage of attempts at implementing AI in demand forecasting, Baird says.
“‘We took these inputs and we’re telling you you need to buy this much.’ Well, that’s great, but why is it telling me this much and not that much? When that doesn’t happens, it raises the risk that the person on the receiving end of those recommendations or predictions kind of just rejects them out of hand.”
Despite the slow adoption that AI has had up to this point in demand forecasting, Baird believes that AI eventually will make its way to the demand forecasting table. There’s simply too much at stake for AI not to be applied here.
For example, in the fashion side of the retail house, the forecast error sits between 30% to 40% on un-promoted items, or products that the retailer has not marked down and advertised at a lower price, Barid says. AI could reduce those error rates by 10% to 20%, she says. “There’s still a lot of opportunity there,” she says.
Those high error rates stem from the fact that retailers face fickle consumers at the end of a large supply chain. Companies that sell to other businesses – such as the manufacturers or distributors who sit further up the supply chain – are working with more complete data and lower risk, Baird says.
“In a B2B world, all of your customers are known,” she says. “If you’re a supplier to a retailer, you already operate at a much more aggregated fashion, because that retailer is aggregating that consumer demand before they’re passing it on to you. The opportunity for error is smaller because you’re operating at a higher level in the forecast model.”
Predicting demand for a retailer is much more difficult than it is for distributors and manufacturers because of the granularity of the products and the unpredictable nature of individual customers.
“Literally, you’re trying to predict for the next five people who walk into the store, what are they going to walk out with?” she says. “That level of prediction is extremely difficult and prone to an enormous amount of error.”
Retailers have tried to protect themselves from that error in the past by emulating their suppliers and aggregating the demand, Bair says. Instead of trying to predict exactly what those five people will buy, they figure out which five products will move, and then average it out across a group of stores.
“That probably protects you from some error,” Baird says, “but then you’ve got to figure out which stores to put it in.”
That puts merchandizers up against another dynamic at play: long-term strategic planning versus short term tactical planning. “That’s been the biggest challenge for retailers: balancing the two,” she says. “It all depends how much am I going to send to store versus how much am I am going to send to my ecommerce distribution center and try to balance it all at a more immediate execution level.”
Retailers have availed themselves of advances in big data analytics by creating more fine-grained buckets. Instead of tracking sales of a given dress at a certain location, they’re tracking what demand for a specific size and color of the dress at a certain location. Not all retailers are able to sufficiently track and predict sales at this granular level, although some retailers, like Walmart, are investing aggressively.
The breakthrough will come when smaller, non-Walmart-sized retailers can leverage AI technology to more accurately predict demand and lower those large error rates, without exposing themselves to escalating shipping costs associated with the omni-channel appraoch. Baird ultimately is bullish on AI’s potential to deliver the goods.
“The amazingly beneficial part of machine learning is the fact that it can look at that most granular level and say ‘I was wrong this time. Now I’m going to change it so I’m not as wrong. Here’s what I learned from that. Here’s how I’m going to apply that learning.’ That’s happening at a speed and level of granularity that people just can’t do. That’s where machine learning is really driving the value on the forecasting side of retail today.”
Demand forecasting among large retailers is largely conducted the same way that it has been for the past 20 years, Baird says. However, there are requests for proposals (RFPs) from retailers for more AI-style approaches, which she expects will result in proof of concepts (POCs) by the end of 2019 and 2020.
“They’re definitely thinking about it and looking to implement it,” she says. “I can say there is some really interesting activity going on and a lot of retailers are looking how to apply AI in merchandizing and merchandize planning – the forecasting part of the world.”