After traveling the spring user conference circuit, one topic that I find difficult to let go is the topic of “Artificial Intelligence ” (some want to call it “augmented ” intelligence), or AI. Every major tech vendor offers either AI-infused capabilities, or flat-out offers a platform for leveraging AI.

What’s the difference? Well, the answer is more complicated than that, because first you have to define what AI really is.

My take: AI is a collection of advanced technologies that are very good at defining patterns, and can get better at it over time – which makes them particularly good at unstructured data, like analyzing Tweets or categorizing images. The collection of technologies include Natural Language Processing (NLP), Machine Learning, a subset of that called Deep Learning, and also predictive analytics.

You can get into all kind of computer science really fast, once you start scratching the surface – into neural networks and game theory and different kinds of optimization algorithms. And the maturity of how to use these technologies is still very low. Our robot overlords aren’t coming for us yet, not when there is still a lot of human intervention that’s needed to make sure AI is coming to the right conclusions when its exposure to data is not infinite.

In fact, there are plenty of examples of AI going awry – if you have time to burn, you can try the creepy inspirational poster generator, or shop from the bot designing iPhone cases (the worst ones are no longer available on Amazon, apparently and sadly), or in case you forgot, there was Microsoft’s Tay, a Twitter bot that was supposed to learn from raw Twitter how to Tweet, and ended up having to be taken down for learning to be a Nazi instead. I’m sure there are more examples of these that I’ve missed (but do share in the comments below!).

The difference between a platform and an AI-infused solution comes down to which collection of technologies we’re talking about here. Are we talking about predictive analytics combined with machine learning? NLP combined with deep learning? And, even more importantly, are we talking about one AI “instance ” or many of them? An AI that is trained and tooled to implement personalization isn’t going to help you much in categorizing images. And we’re not really far enough down the road of what AI can do and how well it can do it to be able to say that an AI trained to do both personalization AND categorize images is going to be better at either one than two AI’s that are specialized.

The dilemma that is coming for retailers then becomes: do you invest in applications that have very specialized AI’s, or do you invest in platforms that have AI embedded in them? Or do you even need both?

I don’t have answers to this question, though it is clear that all the major tech players are drawing battle lines around some answers, with Einstein (Salesforce), Watson (IBM), Coleman (Infor), among others on the platform front, and a whole host of startups who offer specialized solutions at the other end of the spectrum. AI is still in its infancy, and in fact, it seems the industry is starting to move from Gen-1 AI tools like Watson, to a Gen-2 approach where apps are designed with AI embedded from the start, rather than bolted on or injected in along the way.

It really leaves retailers with a lot more questions as a result:

Data. Do you have enough data? I know that seems weird to ask, because a lot of retailers feel like they have too much data already. But when it comes to machine learning, you need a big enough starting data set for actual learning to occur. How many pictures of red dresses does it take before an AI can recognize a red dress when it sees it? How many pictures of red dresses do you have?

The tech providers painfully understand this challenge. It’s why they increasingly take the approach of a “trained ” AI out of the box, because pretty much only Amazon and Walmart might have enough data to feed the AI beast all by themselves, which leaves other retailers at a disadvantage (and I will note: this is what makes Amazon the scariest of all competitors for retailers – the data advantage, and how that data can be used – far more than price or convenience or Prime Now coverage).

It’s why cloud computing is getting more and more important to tech providers too, not just because of the advantages from a support perspective, but because of the opportunity to create data oceans, fed from billions and billions of client activities, where their AI’s can swim and learn. The future of technology in retail may not be a question of capabilities – it may be more a question of how much pooled data a company has access to.

It’s ironic for sure – retailers have long been paranoid about sharing data, particularly data about customers. But they may rapidly find themselves in a world where the tables have turned, and they will be falling over themselves to join up with a platform that has an enormous amount of data, and the AI skills to learn from it.

Workflow. I find myself using this term more and more lately. Back in the Services Oriented Architecture days, there was a lot of talk about workflow, as the tool for stringing together services into processes that support the business. All that talk is back, in part thanks to micro-services, which is really just SOA reduced to a more granular level (though not without creating new complexities in there).

But AI has a role here as well, and here is the dilemma for retailers. Cloud is an important push to be more disciplined about process – to not create inflexible, hard processes that can’t be changes just for the sake of “we’ve always done it that way “. Is there really any competitive advantage in how you process an invoice? AI will force this race to vanilla over custom code even faster than cloud. Because if AI can evolve a customer interaction, why can’t it be used to evolve a business process? If the process is enabled by configuring workflow, why does it have to a human who identifies that the process is drifting from the technology, and make the needed configurations to realign it?

I have a feeling this flexibility of process is going to become the new digital divide in retail: retailers who have invested and who have maintained the discipline around vanilla, configured implementations (vs. those who have hard-coded customizations) are going to be so much faster, and have the AI capabilities to identify when a process needs to change, as well as the capability to implement that change without a whole lot of angst (or human intervention) within the organization.

Sub-Optimization. With a new wave of “AI-first ” applications coming to market in the next 2-5 years, retailers will also have to face up to some long-held corporate myths. Retail has long held merchandising separate from supply chain, for example, and it does lead to sub-optimization, especially when you factor in distortions caused by misaligned metrics or competing priorities. But solving this problem requires throwing a century of merchandising process out the window, right at a time when retailers are struggling to figure out how to keep the store from getting thrown out the window in the rush of digital transformation.

It’s just way too much change. But that doesn’t mean the changes shouldn’t be made. And it most likely means that retailers who figure out how to survive those changes are going to redefine retail – yet again.

AI Is Not All Rainbows And Sunshine

I don’t want to leave you with the impression that I think AI is about to completely disrupt retail and everyone should run for the hills right now. But change is coming. On my radar scale, AI in retail feels about like the Internet in 2006. Everyone then thought it was going to take over the world, and that retail was going to be transformed by it. Eleven years later, we’re STILL talking about many of the exact same things. Transformation came, but it didn’t come overnight, even when we could see the seeds of that change and the direction it would take us.

AI seems to be in a very similar place. Transformation is coming. It will change the way technology works, and it will shift the value from closed, proprietary solutions to open, flexible ones. It’s coming at a time when retailers are still just waking up to the value of open and flexible – and aren’t entirely ready to give up the advantages they still see to closed and proprietary.

But we will likely still be talking about these very challenges – data, workflow, sub-optimization – ten years from now, when it really starts to hit retail hard. But don’t wait until then to care.