In order to understand the coming AI adoption problem in retail, you first need a little history. A history of price optimization.
Climbing the Markdown Hill
In the early 2000’s a new technology hit the retail market, called price optimization. The first use-case to be adopted focused on markdown optimization, or pricing inventory near the end of its life to clear out as fast as possible at the greatest margin possible. Coming off of the Internet bubble bursting, retailers had big problems with too much inventory, and markdown optimization was their savior – helping them clear out overstocked items without taking an entire bath on margin in the process.
Markdown optimization produced some counter-intuitive results, and was initially resisted, but as its value became proven, more and more retailers with short lifecycle products found themselves in a position of a market expectation that they would have markdown optimization to protect themselves from bad product purchase decisions.
The counter-intuitive part came in two ways. One, it turned conventional wisdom about markdown cadence on its head. Retailers historically waited as long as possible before marking a product down, and then marked it down steeply in a short wave of cuts that came every week or every two weeks at the end of the product’s life. What markdown optimization did (and still does today) is recommend shallower cuts earlier. If you have a 12-week lifecycle for a product, instead of waiting until week 8 to cut prices by 40%, and then 50%, 60%, and 75%, you start cutting prices by 25% in week 6, for example, and then follow a shallower set of cuts until the last hard mark before pulling whatever product is left from the store.
If, by week 6, inventory just is not moving, there’s no sense in keeping it at regular price – full margin on $0 in sales is $0. Making a shallow cut gets product moving again, and sells more of it at 25% off, avoiding a greater amount of sales at 40% off.
The second thing markdown optimization did was take the cadence to a more granular level. Instead of moving all product at the same cadence across all stores, retailers could apply different timing to different locations, and even to different SKUs within the same line. There’s no sense marking down bikinis in September in Miami, even when it makes a lot of sense to start doing that in Chicago. But additionally, markdown optimization could recommend marking down specific colors within the same SKU – markdown the white and yellow now, but keep selling the red and blue at full price, for example.
Retailers initially resisted this idea, on the grounds that consumers wouldn’t tolerate paying full price for the red when the white is sitting on the clearance rack. Turns out, that wasn’t so much of an issue – it’s a pretty easy explanation of “the white and yellow aren’t selling, that’s why they’re marked down but the red is not.” And in the end, allowing a greater granularity in how and when hard marks are taken both captures more margin and is tolerated by consumers pretty well.
Omnichannel fulfillment has created some disruption here since the early days of markdown optimization (why not sell the bikini online in September at full price, no matter where the customer may be, and just ship it from Chicago?), but other than that, the business case for markdown optimization is pretty cut-and-dried, and there is little resistance among merchandising organizations about using it to get themselves out of embarrassing purchases that aren’t moving according to plan.
If you think this is the parallel for adoption in AI – after all, anything that contains the word “optimization” is just a machine learning feedback look away from becoming AI – think again.
Hitting The Wall: Base Price
Markdown optimization is not the full story of price optimization. Markdown is there to get rid of bad buys with the least amount of pain. But it’s only needed if you make bad buys in the first place. Or, at the very least, it’s only there when sell-through is not matching the plan.
Once most of the retailers who were going to buy markdown optimization had bought it, the vendors had to move upstream into other areas of pricing. There are two: promotion optimization, and initial or base price optimization. In the interest of time, let’s look only at base price. And it is here that price optimization hit a wall – an adoption wall.
Base price optimization is exactly what it sounds like: what should be the initial price of a product (when looking at short lifecycle merchandise) or what should the “regular” price of a product be, for items that are longer lifecycle, with multiple replenishment phases (like grocery)? Base price has had the greatest adoption in grocery, at least in the US, mostly thanks to Walmart – not because Walmart adopted it, but because grocers who were competing with Walmart needed to take a more nuanced approach to price in order to survive Walmart’s relentlessly low prices.
It generally worked like this. Grocer A competes in a region of the United States, but does not have a 100% overlap in locations that directly face Walmart in an immediate shopping radius. If Grocer A cuts all prices to match Walmart, they go out of business because they don’t have the scale or operating margin be able to match Walmart for long. But if they take a more granular – localized – approach to pricing, they can differentiate pricing, usually into three tiers: locations that face no competition (price what the market will bear), locations that face non-Walmart competition (be aggressive on key items that build a consumer’s price image of the retailer, and make it up on the rest of the basket), and locations that directly face Walmart as competition (a more aggressive promotional stance, combined with price matching for those key items that contribute to price image). Effectively, the locations that face no competition become margin subsidizers of the locations that face stiff competition.
But even in grocery, which had an easy, compelling business case for base price optimization, there was push back. I vividly remember attending a price optimization conference in fall of 2008 – it was the same week that the stock market crashed, which is why it sticks so permanently in my mind – and one of the speakers was from a general merchandise retailer who explained the lengths she had to go to in order to get her company’s stores to implement some of the base price recommendations that were coming out of the optimization. She was in a bit of a unique position, because stores didn’t have to accept price recommendations, but the example still illustrates the problem.
In her example, she talked about the price of ice – you know, the big freezers at the front of the store, which you bust into pretty much only when you’re throwing a party and you need 10 lbs of ice chips for the coolers so that you can store drinks there. No one looks at the price of ice. More often than not, consumers have shopped the store, remember at the last minute they need ice, and request that the store associate ringing their sale just add a couple bags, which they’ll pick up on the way out.
For this particular retailer, they had gone through implementing base price optimization but were getting very tepid results, and they had a hard time figuring out why. Finally, they just sent team members out to stores to physically observe what was going on. Turns out, none of the stores were implementing the new prices. When they asked store managers why they were ignoring the new prices, the store managers said that the new prices were unreasonable and there was no way they were going to implement them. The store managers pointed out that their own bonuses were on the line, and they weren’t going to take a hit just so someone else’s team could look good.
When the team asked what was so unreasonable about the prices, the store managers pointed to a lot of different products, but one that was 100% common across all of the store managers interviewed was ice. The retailer generally sold ice for $1.27, and the price optimization had recommended jacking the price up to $2.99, way more than double the old price. “If we put that price up, we’ll never sell ice again” was the general response. Additionally, the store managers took this one example as proof that the optimization was just wrong, and that it must be wrong across the board. Thus, nearly 0% compliance on the base price recommendations in stores.
The project manager and event speaker, exasperated, finally arranged for one store manager’s bonus to be guaranteed, so long as there was 100% compliance on implementing prices coming out of the base price optimization. At the core of any price optimization tool is an assessment of elasticity of demand – think Economics 101, where the price of gasoline is very inelastic (you need it, pretty much no matter what) while the price of, say, chocolate cake is very elastic (if the price goes up, you can easily opt to do without cake).
Turns out, the price of ice is very inelastic. Even at $3 a bag in 2008, it was inexpensive enough that no one blinks at the price. It’s a convenience item, which means you kind of want to buy it at the place where you buy your groceries so that you don’t have to make an additional trip just to get ice. And it’s purchased infrequently enough that people generally don’t have an anchor price or set expectation for how much ice should cost.
The store manager grudgingly raised the price of ice (along with all the other price recommendations), and found… no impact on the purchase of ice. Demand stayed the same, and the retailer made a much higher margin on ice. And the world didn’t end with all the other price recommendations, either (some increases, some decreases). In fact, the store manager got his 100% guaranteed bonus, but ended up leaving money on the table because the store outperformed even higher-end expectations during the test period.
With that store manager as an ally, the price optimization project manager was able to convince other store managers to buy in, and eventually got the project rolled out as intended, although more than six months behind schedule because of having to fight a ground war store manager by store manager to convince them to adopt the new price recommendations.
Explaining price elasticity to store managers was not enough to convince them to adopt. It was only after the team convinced one store manager to try it that they had enough leverage to begin to convince other store managers to try it. But even then, it was really the combination of both parts – the evidence that it worked, combined with an explanation of why it worked – that really moved the needle on adoption.
This is the lesson for artificial intelligence: you have to both convince someone to take a header on a ‘wing and a prayer’ and then explain to them why it worked differently than expected. And that is exactly the biggest challenge for AI today – most AI solutions do a very poor job of explaining why it should work differently than their expectations. Think of it as the Kung Fu Hustle of retail: when faced with people who have spent their entire lives being good at the thing they base their reputations on (whatever it is that AI optimizes), sometimes you need to show them that there is something new to learn that they never expected in order to get them to be your ally instead of your enemy.
The Bottom Line
The difference between price optimization and AI, at least around where it can be applied within merchandise planning, is that price optimization promised to make existing jobs easier – and, in fact, created whole new pricing departments within the merchandising organization. With AI and merchandise planning, the promise is a threat – no new merchandising jobs (maybe some in IT), and potentially doing away with existing jobs as it automates more and more of the merchandise planning process. If you can’t promise people some intrinsic benefit out of that, like a whole new opportunity to learn, just convincing some people to be guinea pigs for a new process isn’t going to be enough.
And the one place where AI is the weakest today, is in teaching people the why behind what an AI recommends. Black box AI should not be trusted – we don’t have enough controls in place today to make sure that it doesn’t go off the rails in what it learns. But worse, when it doesn’t educate the people using it on what it learns, it keeps a critical benefit off the table – one that would help overcome the barrier of user adoption.
In price optimization, base price made some traction in grocery, but it took a lot longer outside of that vertical, and it’s only through extensive efforts of companies like First Insight, which exposes the “why” behind its price recommendations, that adoption is coming. Promotion optimization is a whole other mess, which lags even base price – and adoption challenges there are not limited just to fashion. AI needs to learn the lessons of price optimization, and address them. Retail is an early adopter of AI, but that alone does not guarantee success.
Nikki Baird is a vice president of retail innovation at Aptos, a retail enterprise solution provider. Her opinions are her own.