We’ve talked quite a bit on this blog about recommender systems; so much, in fact, that it’s beginning to border dangerously on an obsession. And who can blame us? Recommenders are fascinating creatures, being some of the simplest AI systems capable of making meaningful predictions based on data. If you’re curious about how they work, I recommend checking out our previous blogs, but for today I’ll be talking less about how they work, and more about my take on one of the cleverest uses for them I’ve seen in a while.

The story starts with word of mouth; I heard from a friend about a company that sells posters printed on a durable sheet of metal. I had a new apartment and several blank walls to fill, so I was a prime target for indiscriminate ads. I idly clicked over to their site and began browsing the merchandise.

While I browsed, the first step of their AI system spun up. Although they had little to go on at the start, a quick survey primed their recommender with my preferences, and they began displaying recommendations immediately. Meanwhile, I struggled through an extremely clunky search experience, and after a few minutes of browsing I gave up and moved on. Customer lost, right?

Wrong; because the next day, I started noticing their ads popping up around my web content. That’s nothing terribly special these days; I had demonstrated that I was a potential customer, and that made me worth the ad space. What was unique was the ad content: instead of just a generic logo, they were advertising prints related to the ones I had viewed before, alongside the ones I had viewed previously. And so, intrigued by the new relevant content, I clicked through and was convinced to register an account. And after fighting through the search features again, I left, was greeted by more recommender-based ads the next day, and the process repeated. In the end, I had a cart full of more prints than I care to admit, and all of this despite a limited art selection, and one of the least impressive systems I’ve seen in a digital-native operation.

So, victory from the jaws of defeat! A lost customer turned into a very profitable sale thanks to the union of targeted advertising and recommender systems. Alas, I have a confession to make; at this very moment all of the prints are still sitting in my cart for one very simple reason: in the few weeks this process took, I kept watch of their sale prices and realized that the current offers would be a bad buy. And so here I sit an unclosed deal, despite the best efforts of one of the cleverest and best-integrated AI systems I have seen all year.

The bottom line – no AI solutions exist in isolation. They exist within complex systems, systems designed around humans. And no matter how good the model is, no matter how innovative the strategy, if the system isn’t designed for humans it can easily end up high and dry, through no fault of the AI.

Chris Brace is a Data Analyst at Pandata.