Earlier in my career, I was lucky enough to work with what seemed to me at the time to be a chaotic fashion retailer, owned and operated by a Mr. H.
Mr. H was the quintessential fashion entrepreneur; he instinctively knew the target customer and had a success rate with picking the right products that was so far ahead of statistical probability that he deserves to be accredited with a "golden touch" award.
The product was disposable extreme fashion, with styling and designs that were influenced by designer brands but leaned toward the edgier side. In fact, this company had a large sales bump for Halloween when the conventional kids bought the low-priced extreme styles for party wear.
New product floor sets were introduced all the time. The company had in-house design and manufacturing for small production runs and ran a comprehensive testing process that went like this.
A number of new styles were created, from which a few were selected to be produced through the in-house factory. Typically, the factory produced from 10 to 20 pieces. Even if the cost of making these sample runs was high, they were to be retailed at costs that would be achievable if purchased (and flown) from countries with lower manufacturing costs in large quantities. So, these sample runs cost, say, $60 a piece, while the pieces might ultimately retail for $29.99
Mr. H then selected a few stores to test these samples in. He knew where to test them. Stores received four to six units each to display in their stores over the weekend. On Monday when the merchandise reports were run, he'd review the sales and make multimillion dollar purchase decisions. If the stores sold through 3 of the 4 units, sizable purchases were booked offshore for items to be produced at economic values and flown to the company distribution center. It should be noted that in this type of disposable fashion product, if the customer didn't like the item, the markdown to clear wasn’t 10 to 30 percent; if it was wrong, the cost of the mistake was big — more like 50 to 70 percent off.
At the time, with my working knowledge of statistics, I thought that taking this kind of risk on such small samples was complete craziness, but as noted, Mr. H's batting average was way above the proverbial coin toss. The process wasn't perfect, mind you. When one big bet didn't work on an item that Mr. H was certain would be a runner, we had to improve the methodology. Seems that store employees with their big employee discount had done the initial purchasing of the samples, and contaminated the results. So a two-week ban on employee purchasing was introduced and that helped keep the methodology valid.
Since my days of running planning and allocation with Mr. H, I've worked with hundreds of other retail companies. While I have met some superb merchants, I rarely see any of Mr. H's caliber and few who exercise this type of test-and-learn approach as a vehicle for reducing product selection risk. I label a more extended version of this process “Moneyball retail”; that is, the application of applied mathematics to retail. But as noted, few companies practice this. (One notable exception was a company started by some math graduates who had never worked in retail before and didn't know any better.)
Since those days, retailing has actually become more complicated, mostly by the requirement to tailor assortments to local market needs, which requires buying (and testing) many more SKUs and creating separate offers at least by channel and ideally by cluster. When Mr. H decided that a test was going to be a runner, it was distributed chain-wide, not on a localized basis.
Finding more Mr. Hs isn't likely, and even his golden touch would have trouble keeping up with today's product localization needs. By reducing risk and doing a much better job with the need for more in-store items, better differentiated and localized assortments can be supported by AI-based technology. Sticking with fast fashion as an example, let's break down how AI can support that golden touch.
There are always candidate products to be evaluated, and this is true for all types of retailers, whether specializing in apparel or consumer goods. New product introduction, or NPI, is something all retailers have to be great at. Competent testing will help ensure that the best products are selected and deployed to the best distribution channels and/or stores.
With a candidate product in mind, the first golden decision Mr. H made was the selection of which stores would provide valid test results (if you consider sales of 2 or 3 units a valid measure). Mr H knew which stores served as bellwethers, something a trained AI model can also do. For some types of products, certain stores are going to be better early and reliable predictors of potential performance. Given today's need to localize, it's just as important to also test less likely stores (or web pages) for sales success as well as finding some losers to completely validate the sample. Doing so shows that your model is picking winning stores for the right reasons — if the stores it expects to be losers fail to lose, then the model needs to be tweaked.
The AI model needs to make these golden store decisions based on not just which stores are reliable predictors of sales, but which stores are rapid and reliable predictors — again, a perfectly feasible outcome of a properly trained and deployed AI model.
The next golden decision our AI model needs to make is to forecast the best distribution for that product as well as the right quantity to purchase based on the rapid sales read of the bellwether stores. If it sold in store X, what other stores will perform like store S and in what quantities? While this is a bit trickier, it's also completely within the realm of possibility of a properly set and well-trained AI model. Consider that the model doesn't have to be perfect. It just has to make better decisions than a good merchant, aspiring to the batting average of Mr. H.
For a lot of retailers who do not care to understand how AI models really work, this NPI methodology hopefully provides some insights into the potential that properly delivered AI can offer. Finding more Mr. Hs is not likely, but introducing capabilities as described is.
AI doesn't replace retailers; in fact, it needs their domain skills to be successful. Such skills are essential to the creation and ongoing management of the model. That outlier induced by employee purchases described above? That’s the kind of bias that may go completely undetected by an AI model or the data scientist managing the training. But it was the first thing that Mr. H. considered when it came time to tune up his methodology.
And lastly when Mr. H. came in on a Monday morning, he knew exactly what items had to be reviewed, probably having never even made a written list. For too many retailers, test and learn is really test and churn. They often miss the key organizational, orchestration and analytical components to support what should be a methodological approach, particularly with the number of candidate products and candidate locations that deserve proper tests. That's not really an AI problem, but for many, it is a fundamental requirement to properly manage this NPI test and learn capability along the way to becoming a successful moneyball retailer.
If you'd like more information about establishing a valid test-and-learn approach for NPI or would just like to compare nostalgic experiences with golden touch merchants, contact me here.