Any job that can be automated, will be.
Like it or not, this rule of thumb largely conforms to labor trends. Just look at the manufacturing sector. In the U.S. alone, as many as five million factory jobs were lost to automation between 2000 and 2010.
According to some reports, knowledge workers may face even fiercer competition from automation than their manual-laboring counterparts. No wonder white-collar workers are worried. Consider the superhuman volume of data that AI can process and funnel into almost instantaneous decision-making.
However, most executives would argue there’s much more to making decisions than just the numbers. A smart, experienced human has gut instinct and familiarity with how the real world operates. You can’t capture that on an Excel spreadsheet.
Tarun Kushwaha, a professor of marketing at the George Mason University School of Business, recently ran an experiment that pitted the brainpower of actual human executives against trained algorithms. Together with co-author Saravanan Kesavan, a professor of operations at UNC Kenan-Flagler Business School, aimed to discover whether humans possessed decision-making advantages over algorithms – and if so, what conditions or circumstances brought those advantages out.
The professors collaborated with an automobile replacement parts retailer to launch a randomized controlled trial (RCT) centered on the corporate buyers responsible for stocking the shelves for affiliated stores across the United States. These are tough selections. Hundreds of automobile models are currently available in the U.S., each one containing approximately 30,000 parts, most of which are non-transferrable to other models. Multiply these numbers by hundreds of store locations, and you have a dizzying array of possible combinations numbering in the billions.
Moreover, auto parts differ from most other retail industries in that customers won’t wait for back orders. Navigating most places in the U.S. is next to impossible without a functioning car, so the demand for a replacement part must be fulfilled right away or the customer will look elsewhere. And with 30,000 parts that could potentially need replacement, demand is spread incredibly thin – annual inventory turnover for this industry is close to one. In a low-turnover, high-margin business like this one, everything hinges on demand prediction.
The company decided to supplement the buyers’ intuition with some algorithmic firepower. They created an AI tool to flag underperforming products for removal so that they would no longer occupy valuable shelf space. The algorithm uses local car registries, prevailing weather conditions and detailed product histories (among other resources) to generate demand forecasts that are highly targeted both by item and by store. However, the gut-trusting buyers rejected the algorithm’s choices more than half the time, opting to hang onto the flagged products rather than remove them.
For obvious reasons, the retailer wanted to know whether buyers were vetoing the AI recommendations out of sheer pride, or because they knew something the robots didn’t. Kushwaha and Kesavan devised an experiment whereby the human overrides were followed in some store locations but ignored in a random selection of other stores. The trial lasted 12 months and involved more than 30,000 discrete products.
Taken as a whole, the results appeared to be bad news for the buyers. The stores that honored the human overrides were 5.77% less profitable than those that obeyed algorithmic recommendations to the letter. However, that’s not the end of the story. The professors also looked at growth-stage (i.e. new to market), mature-stage, and decline-stage products as isolated categories. They found that the algorithmic advantage was confined to the latter two stages. For growth-stage products, the buyer overrides drove more than 23% greater profitability. The simple reason for the overall 5.77% difference was that the vast majority of car parts were either mature- or decline-stage.
The professors theorized that AI’s vaunted data processing capabilities fell flat when faced with new products, which by definition have no past performance record on which to base decisions. By contrast, buyers can tap their networks for priceless clues. This is particularly important in the auto parts industry, because replacement parts are handled through car dealerships, rather than external retailers, during the warranty period. The dealers, in turn, source parts directly from the original manufacturer (Bosch, Johnson Controls, etc.). Auto parts suppliers, therefore, are rich sources of information about buying activity early in the product life cycle. Algorithms cannot get at this information independently. But buyers can. It’s as easy as asking your buddy at Bosch out for a drink. Through interviews conducted as part of the research process, the professors learned that these info-sharing conversations between buyers and suppliers happen all the time.
Armed with this knowledge, Kushwaha and Kesavan are working with the retailer on further experiments that they hope will help define the ideal way for humans and AI to make joint decisions.
Needless to say, every company and context is different. What works for the U.S. auto parts industry today may not for another time, place or sector. But the general principle seems universally sound: Smart humans have private information that should be extracted and added to the data-driven decision making that algorithms perform much better. For this retailer, that may mean the buyers choose which products to stock, and the algorithm draws upon their selections to allocate store inventory. This would leverage buyers’ private information about the likely market performance of products. It would also radically simplify the choices presented to buyers, freeing up time for activities that they – not the algorithm – do best, such as supplier negotiations, scouting for promising new products, or drinking in more precious information at happy hour..
Source: Saravanan Kesavan, Tarun Kushwaha (2020). Field Experiment on the Profit Implications of Merchants’ Discretionary Power to Override Data-Driven Decision-Making Tools.