Source: Stitch FIx
How much do e-tail algorithms need humans?
“Data-driven” is one of the retail world’s biggest buzzwords. But as tech companies bring on experts to build out personalization solutions, it’s raising the question of how much of a human touch needs to remain to give customers the personal feel they’re looking for.
As Amazon.com has pursued its foray into apparel, it has hired people with backgrounds in fashion to train the e-tailer’s algorithms on what kind of recommendations customers are looking for, according to the Washington Post. Describing the workday of a person who held such a job, the Post describes a fashion school graduate making split-second decisions between photos of two outfits and voting on which was more fashionable. The idea of bringing on a large number of stylists with varied tastes to train the algorithm was done with an eye toward creating a “universally trendy master stylist.”
And yet similar attempts by Amazon to automate taste making haven’t necessarily gone well.
Amazon Books, for instance, curates its entire assortment based on what’s popular on the website and other data and uses localized information to inform a featured section in the store. In the e-tail giant’s perfect world this would lead to an experience that would meet a customer’s needs. But reviews of Amazon Books locations have been mixed. For instance, an account of a first interaction with the stores published in Quartz last year disparaged the New York City location as being impersonal and sucking the joy out of the bookstore experience.
Pursuing a more human but still data-driven model, Stitch Fix employs 3,700 remote stylists that make choices based on trend data and automated suggestions as well as personal data about the customer, the Post reported. The company says that the model improves both the stylists’ work and the algorithms over time. The retailer has created 16 private label brands based on what it has managed to learn.
Others argue, however, that this hybrid human/data model doesn’t provide the real personal touch customers are looking from a personal stylist.
Discussion Questions
DISCUSSION QUESTIONS: Where do you come down on the data/human split when it comes to retailers making product recommendations? Do you think a hybrid model is superior to ones that rely solely on data or humans?
Poll
BrainTrust
Michael La Kier
Vice President, Brand Development - IGA, Inc.
Cynthia Holcomb
Founder | CEO, Female Brain Ai & Prefeye - Preference Science Technologies Inc.
In very general terms, I believe a hybrid model using both data and humans will deliver the best results for product recommendations. That said, the mix could vary significantly by product category. The ability to discover new insights from massive amounts of data is undeniable, however, the effectiveness of the product recommendations will be impacted by the quality of the data, the effectiveness of the algorithms and how the recommendations are executed or delivered to the shopper.
Retail has and always been about the blending of the art and sciences. Algorithms, machine learning, predictive analytics, artificial intelligence and intelligent forecasts are all tools available to an actual human with which they can drive their own recommendations. A hybrid model of analytics along with a human touch will help drive a personalized and enriching in-store, as well as digital, experience.
Outside of pure analytics and algorithms there is something known as intuition, which helps the retail brand ambassadors stand out in the eye of the customer. Making informed recommendations, based on facts and intuition, is the winning combination. While consumers may enjoy the convenience of digital commerce, the in-store experience is a refuge from all the algorithms and the art of retail.
It is not “personal,“ it is a guess based on a whole lot of other guesses. Watson works in healthcare so well because there are black and white cases it can compare. Taking a bunch of guesses based on people’s opinions and thinking you have facts to work off seems to be remarkably un-advanced.
As the old Rick Springfield song goes, “we all need a human touch.” After all, if we want to be data-driven as an industry, we need to set the destination for data to reach and achieve on our behalf.
Humans versus technology is an argument that has been going on for decades and will continue for decades to come. The answer is you need both, but you have to have a healthy balance. Technology cannot replace a well-trained human who knows how to successfully interact with a consumer, making them feel special, being able to answer questions and make recommendations as of today. Perhaps someday in the future we may get there, but I don’t see it happening any time soon. However, technology can provide speed and for simple functions an easy way to get what we want or need to have done. Indeed, scheduling a phone call or appointment online, booking travel, making a purchase and so many other tasks have been made simple and easy by technology. I don’t have to call anyone, be on hold and run the risk of speaking with someone rude. However, when I’m in a store or on a phone call and need assistance, having to deal only with technology is usually frustrating.
What businesses and retailers have to understand when investing in technology is to make sure that what they are spending on is what the customer will want and will respond to favorably and not invest in something just because they see an opportunity to save money on payroll. That type of thinking will work against you. Technology has a significant place in the world today, and it will keep expanding, but it is essential to know that the simple things like a smile or a genuine “thank you” is what makes customers feel special and no technology can replace that.
The quick answer is that the goal should be about 90 percent data/AI and 10 percent human, with the humans only getting involved with initial concepts and design, data issue resolutions, context confirmation and exceptions. Right now, the “seat of the pants” analysts still hold sway at many retailers, so the division is closer to 50-50. More and more of these folks are retiring every day and the new managers who grew up digitally informed are taking over, so the industry is heading in the right direction.
Humans are hard-wired to believe that we have some inherent advantage over algorithms, in the same way we were certain the sun revolved around the earth until science forced (most of) us to believe otherwise.
The basic product recommendation engine on Amazon (which is not a hybrid) drives 35 percent of Amazon’s revenue. Netflix and Spotify’s non-human curation have won over millions of consumers. Amazon is rapidly replacing merchants with data scientists in its “hands off the wheel program” and driving greater profitability AND customer satisfaction simultaneously. Digital native vertical brands like Rockbox have already disclosed that their algorithms substantially outperform their human stylists.
Conflating the pros and cons of Amazon Book retail stores with the long-term success of data-driven curation/merchandising makes no sense. Is the failure of the Amazon Fire phone also proof that data based merchandising doesn’t work?
I know it’s not a popular opinion (and in many cases goes directly against our cognitive biases) but data-based curation and merchandising is already better than humans in some cases, and its lead is only going to get bigger. The open question is simply the timing. I certainly wouldn’t turn my cancer diagnoses over to Watson today, but I’m pretty confident that my three-year-old will enjoy a lifetime of much better healthcare as a result of data-driven decisions.
We need to be cautious about believing the myth that algorithms become personal just because a person helped build the list of recommendations.
This is a classic tech error. And the truth is we need to treat tech as tech – and not expect it to be personal or be able to be personal. It is what it is – retailers should use it for what it does well.
Our experience is that personalization without a human touch doesn’t really connect with the customer and doesn’t motivate them. When you look at the results, you can see the data support for why something was recommended, but the reason why it was selected makes sense to the computer, not the customer.
The crucial time when humans need to be involved is BEFORE the algorithm does its magic. Algorithms work based on historical data. The key role for humans is to organize the historical data and identify characteristics, features or relationships that matter to customers, so that the resulting data analysis reflects that relationship.
No two consumers are identical, the same can be said with their preferences when it comes to selecting a product. E-tail algorithms can help but should be used as a tool rather than the complete solution. Consumers are constantly researching new trends, brands and products and the last thing they want to see is an outdated product popping up on their pages constantly.
Humans are inherently biased to their own individual preferences and viewpoints. Data interpretation itself is subjective and open to biased interpretations by those hired to interpret the data. AI/ML is subject to the biases of both data scientists and computer scientists behind the reasoning designed to search for a solution. Fashion, for instance, is subject to individual fit, look and feel preferences. A product can be fashionable but not preferentially wearable to an individual. A hybrid model itself can be subjective, as pointed out by the outcomes listed in this article.
The only way to return individually, personal human recommendations is to build deep learning algorithms based on expert knowledge of a category, like fashion, imbued with the nuances and subtleties apparel experts know stripped of human bias. Filtering Big Data through this system results in individually personalized outcomes.
Agreed on the general consensus that hybrid models are the best solution. I’m particularly in line with Ken and Seth, who both point to the danger (or irrelevance) of algorithmic input when recommendations generated are purely based on historical patterns.
Stale assortments and boring product recommendations are a symptom of over-reliance on automation and backward-looking merchandising strategies. Paradoxically, as the retail industry struggles, buyers often become more conservative and less willing to take risks based on gut when predicting what customers will want. And companies become more reliant on computer-generated “bets” because they feel fully validated by data.
Human input is crucial, if for no other reason than to inspire customers with unexpected recommendations that push them out of their comfort zone when it comes to personal style and consumption. This is part of the value we look for as customers when shopping. A bit of surprise and delight builds trust and creates a foundation for relationship with a brand, and can be difficult to nail without human perspective.
Putting so many assumptions together to give me an answer/advice seems to be sorting my tastes into a category of others similar to me and suggesting their choices that seem to make sense for me, on average — which in turn kind of makes me fit into that average. That’s personalization?
The two “unassailable” advantages of online selling are the “infinite” long tail — the “everything store” — and algorithmic selling. (The advantages of brick-and-mortar are immediacy and the 360 experience.) Not surprisingly, as Amazon moved into brick-and-mortar, they began by trying to apply their algorithmic genius into their brick-and-mortar stores. As others have noted here, their beginning algorithmic approach was less than totally satisfactory. But Amazon is inherently a LEARNING organization — as contrasted with the 100-year-old brick-and-mortar logistics selling model.
The industry has yet to come to grips with the loss of “active selling” skills 100 years ago. The new passive selling, aka SELF service was such a massive success, delivering billions in profits for retailers, and MASSIVE attendant benefits for shoppers, the original self-service mantra became deeply embedded into the industry’s psyche — “Pile it high; and let it fly!” So much that the personal selling skills associated with salesmen were ignored, with the non-personal selling of a brick-and-mortar store, what I call PASSIVE selling, becoming mistakenly thought of as “selling.”
With the advent of Amazon, the industry was faced with a genuine approximation of ACTIVE selling, embedded in Amazon’s algorithmic selling. (Remember, true selling is always about something that happens in the mind of the shopper. See: Selling Like Amazon… in Bricks & Mortar Stores! – October 25, 2013. THIS is the true challenge of clicks for bricks!
Thinking that pasting online operations onto a massive brick-and-mortar passive selling operation is an adequate response to Amazon is incredibly superficial. Adopting long vanished (gone 100 years) personal selling techniques within the brick-and-mortar store apparently is not going to gain major traction until Amazon works out their bugs and continues their voracious growth — they’ve hardly yet begun! 😉
Sometimes we forget that all algorithms have a starting point. All learning systems are affected immensely by the initial inputs. Garbage in, garbage out is what happens when a learning system takes inputs that are not realistic or poorly defines the reality of an industry. We can see some of that with historically based algorithms.
Lastly, humans will always need to tweak the system. Why? Because the systems are developed to cater to humans, who are always changing. A perfect example was how quickly fidget spinners came on the market to build over $200mm of spend in just a few months. The algorithms were too late to catch it, and the opportunity was lost. By the time retailers realized the fad had started to diminish, there were millions of backorders for fidget spinners that end up being liquidated due to limited demand.
People will be needed to put in the initial assumptions and continue to monitor algorithms to make them more efficient. They might get by without it for generic items, but the hybrid model will beat out the algorithm by itself every time.
I like to think of AI as “Augmented Intelligence,” rather than “artificial.” Outcomes are greater with machines AND humans versus machines OR humans.
Hybrid is the best practice almost universally. Computers need the intelligence and experience of subject matter experts to create effective algorithms that are relevant for consumers. Relying on just computer-based algorithms may miss some subtle nuances that result in non-relevant recommendations that could potentially alienate customers. Infusing the insights of your best sales associates and designers into automated recommendations can turn B, C and D level associates into A associates. Taking expert knowledge, customer context and real-time product information creates a combination that is superior to a one dimensional algorithmic formula.
Obviously it depends entirely on how sophisticated the “data” is. The worst human is probably better than the worst algorithm, but the best human is probably inferior to the best ‘bot … or soon will be.
For decades, the market has been leading consumers and suppliers to believe that Information Technology (IT) will overtake the human race in intangible productivity and self reliance. It is commonplace to point out advancements in manufacturing, design, communications, transportation agriculture and much more. The market never discusses the errors that exist in our mathematics and in the design limitations of the things we build.
IT, with all of the speed and accuracy it offers, is and will remain a tool for far into the future. The tool will make life easier and allow for far reaching solutions sooner than ever anticipated but it is only a tool with no ability for creative abstract thought.
Staying tuned into the consumer with direct access and exchange of ideas will continue to give the best results. Managers at all levels need more time on the floors and less time aimlessly reflecting on the significance of retrospective reports. Remember that at best, computer reports are about what already happened and thus worth no more than yesterday’s papers.
As a general rule, I think the hybrid model — especially with current technology — makes for a better experience. There are of of course edge cases where relying more heavily on one or the other is better economically, or is an integral part of the business or product category. A great example of mixing these two for a quality experience is Stitch Fix. The algorithms, data and “machine learning” help speed, sort and optimize, but over 2000 (last I checked) stylists actually choose the clothes and write the personal notes.
There is no “one fits all” answer. It would be different if you’re a brick and mortar, click and mortar, online retailer, customer base and the type of merchandise offered.
Retailers need a combination of both humans and data. A high-end specialty brick and mortar customer is probably looking for a much more personalized recommendation rather than an algorithm. On the other hand, for an online customer the experience is more transactional (quick and cheap) and an algorithm works for that scenario. Algorithms are based on history and not everything a customer purchases is for themselves (gifts for example), and that tends to throw off the recommendations.