The State Of AI In Retail

If you judged by the number of times the words “artificial intelligence” were used at NRF’s Big Show 2019, you would think that it is a mature capability, well on its way to being rolled out across every retail enterprise.

The reality is a little different. Gartner reports that only 2% of all enterprises (not just retail) have deployed AI and only another 24% are “experimenting” in the short term. However, it’s clear that retail lags behind only financial services in terms of being at the forefront of AI investment – AI startups that are focused on retail racked up more than 2x the market value as all the rest of startups focused on retail combined.

But what does “AI” really mean when it comes to startups, or capabilities, or even specifically retail applications? When you look at the different types of AI out there, it becomes clear that not all AI is equal.


Classifying AI

McKinsey identifies three types of AI: classification types, prediction types and generation types.

Classification generally focuses on Natural Language Processing (NLP) or Computer Vision. AI in this context identifies words or images and classifies them. Tweets can be classified according to consumer sentiment, for example, or product images can be used to identify attributes, sometimes as simple as “short sleeved” or as complex as “floral print.” The value of AI in this context is really in providing details about unstructured information. But to really get value out of it, those details have to be used in order to make new decisions – and that requires either a human or a machine to transform those details into an action.

Prediction is really about forecasting – predicting the most likely next action. Prediction drives everything from personalization to route optimization to everything in between. It takes a different set of AI tools – focusing more on granularity of data, and applying AI in a more “internal” way, where it is looking at traditional prediction models and fine-tuning them to make them better. Good AI-driven prediction separates data that is noise from data that actually contributes to a better result, and identifies the best models to use even as actual behavior and results changes over time.

The last type of AI is generation. Chatbots are the most recognizable application of generation AI – sort of. Most chatbots are built as a broad set of canned responses, with the AI being applied in a classification mode to identify what the user is saying, and then marshal the appropriate canned response. This is how you end up stuck in chatbot loops of “I’m sorry, I don’t understand.”

More sophisticated versions of generation AI do things like take the ability to identify dogs in pictures, and use that knowledge to generate pictures of dogs. A recent addition to this type of AI is a tool that can take a recipe and generate an image of what the finished product looks like. The developers note that it tends to do better with soups than complex plated meals (I wonder why), but it demonstrates a particularly complex set of AI capabilities: the ability to parse the written recipe, translate the combination of the ingredients and the cooking instructions into an expected output, and combine the multiple outputs into a generated image.

For retail, I would argue that prediction is by far the most valuable type of AI that can be applied to a business problem. However, most of the activity in AI in retail is focused on NLP and computer vision – not on the much harder (and more valuable) problem of forecasting. And most of the NLP and computer vision work is focused on classification, rather than generation.

That’s not to say that NLP and computer vision aren’t valuable, just that their value is limited. With forecasting, you have an opportunity to decide what to buy, how much to buy, where to put it, and how to price it, according to the customers you’re targeting. With NLP and computer vision, the way that we can apply them today, you’re doing your best to make the most out of the decisions you’ve already made – trying to get consumers to buy.

That’s just one place where the hype about AI doesn’t really match up to reality. It’s not the only one. Here are five more.

Not Really an Application

When you break down AI solutions, they end up being a lot of “regular” software surrounding very small, specific AI capabilities. 8 by Yoox got a lot of credit as an assortment “designed by AI”. But when you dig beneath the covers, Yoox was assembling images from designated sources – influencers in high-fashion cities – and classifying them into trends that humans turned into a clothing collection. While that’s impressive, it’s many steps short of taking those trends, identifying the right number of items to have in the line, then filling in those item placeholders with real designs.

Solving the Easy Problems First

This isn’t really a criticism of AI, but it’s important to note that where AI is getting applied are relatively “easy” problems to solve, compared to all the problems in retail. In forecasting, all the AI progress is being made in grocery replenishment – a corner of forecasting that is already data-rich, and generally overflowing with inventory (especially compared to the limited runs and short seasons found in fashion). So while forecasting use-cases deliver value, it’s currently in a pretty small slice of the total opportunity. Almost every AI application is like that, at least in retail.

AI’s Black Box Problem

There are two sides to the black box problem in AI, but they are both related. Employees who have to use the recommendations that come out of an AI-driven analysis need to be able to trust the results. And if they can’t see or understand how the inputs translated into a result, they’re not going to have a comfort level with those results, especially if they’re counter-intuitive to what employees already “know” to be true (whether actually true or not). I’ve encountered a few examples of adoption problems coming out of AI projects already where employees have outright rejected AI recommendations, and it has put projects at risk. Telling people “Just do what I told you to” isn’t going to solve that problem.

The other side of the black box problem is more fundamental: how do you keep the AI from learning things it should not? Apparently, AIs can learn to collude and discriminate fairly easily, and without someone monitoring the conclusions that AI’s learn, a retailer relying on an algorithm that has learned the wrong thing could find itself in hot water with regulators very quickly. And the moment employees catch wind that the AI is “wrong”, then your AI project is sunk.

Ethical Dangers for Retail AI

AI developers and researchers are grappling with how to make AI ethical – how to expose the hidden assumption about the world that humans have that are so ingrained that we don’t think about them until it’s too late. You know, simple things like “don’t eat people,” or that price discrimination based on income level or ethnicity is wrong.

But retail has a special blind spot when it comes to managing the ethics of a technology, as we have seen over and over again with consumer privacy. Let’s just take mobile phone tracking as one example. As retailers first tried to roll out consumer in-store tracking using things like mobile phone sniffers, there was a lot of finger-pointing between retailers, who said “we rely on the technology vendors to manage the privacy impacts” and vendors who said “we rely on the retailers to implement the technology in a way that meets their customers’ expectations for privacy.” We’re easily headed down this road again with AI.

Avoiding Hammers and Nails Problems

Retailers aren’t the only ones who can fall prey to the problem of having a shiny new hammer and thus seeing every problem as needing nails. Technologists in general are easy victims to this mindset. AI isn’t going to be the right solution for everything. And while you can definitely cast AI in terms of “making prediction cheaper” to the point where you can use cameras to predict human movement and embed that in a driverless car, for example, that doesn’t mean that all prediction is valuable or can be applied in a useful way. People are remarkably good at spotting patterns and we still have not unraveled what goes into an intuitive leap. And there are plenty of places in retail where the combination of art and science is far more powerful than either one alone.

AI today, and for the near future, is excellent at shifting through volumes of data that humans can’t absorb, to find the patterns that people can’t see. It is also excellent at applying that kind of analysis to a much larger number of problems than humans can. But it still needs to be applied judiciously – to not lose the art, and to make sure that we’re prioritizing problems that actually need solving.

The Bottom Line

The hype around AI does not match the reality of what’s actually going on in retail. That’s not to imply that AI is not worth pursuing, or even to suggest that it will fail. Rather, in order for AI to be successful in retail, these issues that exist need to be addressed. Otherwise, we’ll see a lot of talk, as we did at NRF, and not a lot of action.

Read Source Article: Forbes

In Collaboration with HuntertechGlobal

© copyright 2017 All Rights Reserved.

A Product of HunterTech Ventures