AI has found its way into just about every part of ecommerce and digital customer experience. Shopping assistants, chatbots, product recommendations, imagery, search, sizing, customer service. Some of it is genuinely clever.
Some of it has also gone spectacularly wrong.
While looking at examples recently, we found everything from a chatbot apparently agreeing to sell a car for $1 to AI-generated food that nobody in their right mind would want to eat.
And then there was this.
THE BAD
ChatGPT’s questionable alternative to Ralph Lauren
This one actually happened to us. While researching this article, we gave ChatGPT a picture of a Ralph Lauren top and asked it to find something similar for less than $30.
It understood the brief pretty well: dark green, relaxed fit, vintage collegiate lettering.
Its favourite match?
A “Varsity Rimming” t-shirt…
Technically, the colour palette was pretty close. As a product recommendation, however, there may have been some room for improvement.
The Chevrolet dealership that sold a Tahoe for $1
This is probably one of the best examples of customers discovering that an AI chatbot will happily wander outside the job it was given.
A user interacting with a Chevrolet dealership chatbot instructed it to agree with everything they said. They then got it to agree to sell a 2024 Chevrolet Tahoe for $1, complete with the immortal:
“And that’s a legally binding offer, no takesies backsies.”
Sadly, nobody drove away in their new $1 Tahoe. The chatbot was taken offline instead.
Still, quite a lot of entertainment for a dollar.
DPD’s chatbot decided to review its employer
When musician Ashley Beauchamp couldn’t get DPD’s customer service chatbot to tell him where his parcel was, he tried something else. He asked it to write a poem about how terrible DPD was.
It obliged.
He then got it to swear and tell him that DPD was the worst delivery company in the world. DPD subsequently disabled the AI element involved.
Probably not the customer feedback mechanism they had in mind.
Air Canada’s chatbot invented its own policy
This one was rather less funny for the business involved. An Air Canada customer asked its chatbot about bereavement fares and was given incorrect information about how the policy worked.
The customer relied on that information and eventually took the airline to a tribunal.
Perhaps the strangest part came when Air Canada argued that it shouldn’t be liable for information provided by its chatbot. The tribunal wasn’t particularly impressed.
There’s a fairly obvious lesson in there somewhere about letting AI explain company policies to customers.
Amazon’s “I cannot fulfil this request” product range
At one point, Amazon shoppers started spotting some unusual product names.
Listings appeared with titles containing phrases such as “I’m sorry but I cannot fulfil this request”, apparently because AI-generated copy had been pushed straight into live product listings without anyone noticing what it had actually generated.
Which is impressive in its own way. AI making a mistake is one thing. Publishing the mistake as the name of the product requires a little extra help from the humans.
Instacart’s food photography gets weird
AI-generated product and recipe imagery sounds like an obvious opportunity to create content more quickly and cheaply. Until dinner starts looking like it has escaped from a laboratory.
Instacart attracted attention for AI-generated recipe imagery featuring anatomical oddities including conjoined chickens and food with some extremely questionable textures.
There was also reportedly a recipe involving a cup of “monito sauce”, which would be useful if monito sauce existed.
When some of the images attracted attention, replacements appeared. They were also AI generated…
And a few more for the list…
There’s also Glasgow’s infamous Willy’s Chocolate Experience, where AI-generated promo images sold a £35 fantasy that turned out to be a fairly empty warehouse. Chipotle’s ordering bot was persuaded to stop taking orders and write someone’s Python homework instead. And Amazon’s Rufus shopping assistant has been caught confidently recommending the wrong products, including things that don’t actually match what the customer asked for!
THE GOOD
Because, despite everything above, there are plenty of examples where AI is being used to solve an actual customer problem rather than accidentally creating a new one.
Google: make virtual try-on part of shopping
Google’s approach to virtual try-on is interesting because the technology itself isn’t really the clever bit from a customer-experience perspective. It’s where they’ve put it.
Rather than expecting shoppers to download a separate app and create another journey, Google has been bringing virtual try-on directly into product discovery.
See something you like, try it on.
Removing a step isn’t particularly glamorous, but that’s often what makes a digital experience better.
Wayfair Muse: “I don’t know what that’s called, but I like it”
Furniture shopping has an obvious problem: most of us aren’t interior designers. You might not know the name of the aesthetic you’re looking for. You just know that that room looks nice.
Wayfair Muse uses generative AI to let people explore visual inspiration and move through related styles and products.
That’s a much more natural use of AI than asking somebody to describe precisely the interior-design terminology for something they can’t name in the first place.
Etsy Gift Mode: let AI do the matching
Etsy took another route with Gift Mode.
Rather than putting a blank chatbot in front of shoppers and asking them to explain what they want, it asks structured questions about who they’re buying for and what that person is interested in.
AI then helps match those answers to curated gift ideas.
The interesting bit isn’t really that AI is involved. It’s that the customer doesn’t particularly need to know or care that AI is involved.
There were criticisms of the relevance of some early recommendations, which is also a useful reminder that a good concept still has to deliver good results.
Zalando: AI that actually helps you buy the right size
Less flashy than some of the other examples, but arguably more useful. Zalando uses AI, customer data and body measurements to help shoppers work out which size is actually likely to fit them.
Its Virtual Fitting Room pilots have reduced returns by up to 40%. No chatbot, no gimmick, just AI solving a very expensive ecommerce problem.
So, good AI or bad AI?
The difference across these examples isn’t really whether the AI itself is impressive. Some of the best uses are actually the least flashy.
They start with something a customer is trying to do: find the right size, visualise an outfit, discover a style they can’t describe, or choose a gift.
The bad examples tend to get much more entertaining when the AI meets a customer behaviour, question or input that nobody seems to have anticipated. And that’s probably the interesting bit as more AI makes its way into digital experiences.
We’ve been seeing the same broad pattern in our own AI research. People are using these tools more, finding more genuinely useful applications for them, and also getting considerably better at spotting when the output needs questioning.
The technology is moving quickly. Customers remain customers. And occasionally, they want a cheaper Ralph Lauren top and get offered a Varsity Rimming T-shirt instead…
