How AI Can Become A Tastemaker
Can AI prediction transform hit-driven industries?
Last week, I talked about AI 2027, a very aggressive forecast about the future. This week let’s zoom in on one the ideas I mentioned that will happen well before super intelligence: a smart planning bot. In the far future, perhaps companies will be collections of billions of superhuman bots. In the meantime, though, smarter bots and synthetic users can change the way companies launch new products. The first versions of these tools are already coming to market, and by next year they’ll be shaping the way brands decide what to make.
The Innovation Treadmill
Think about a typical U.S. apparel label. How much are they depending on new SKUs for sales? I actually went down a bit of a rabbit hole to answer this question and pulled a bunch of data from Pyxis, our alt data platform that you might remember from my early posts here and here. What I found was interesting.1 A typical apparel brand like Abercrombie gets 45-50% of sales from new SKUs in a quarter. For more of a fast fashion brand like Forever 21, the number is 75-80%. I found one very old brand (that I won’t name) that is focused on a single iconic style and is getting only 5-10% of sales from new SKUs. It’s a huge range!
If you happen to work for or invest in that old iconic brand, you can probably stop reading this post. All other merchandisers, however, are on a treadmill and the only question is whether the speed is “fast” or “very fast.” They need to scramble for new silhouettes, colors, fabrics, and price points just to keep up. Other industries such as toys, video games, quick-serve restaurants, movies, etc. are also dependent on constant innovation just to stand still. Hit‑driven markets live or die on their next big idea.
The trouble, of course, is that most of those ideas miss. Executives must come up with many new things in the hope that a few will hit and make it all work out. Everyone accepts that this portfolio strategy is a necessary part of the business.2
How AI can help
Now picture a different feedback loop. A designer sketches a sneaker, presses “simulate,” and a cloud of synthetic customers—digital twins built from real transaction logs, survey panels, online comments, and TikTok feeds—swarm over the idea. Within minutes the model spits back a projected demand curve, a list of likely superfans, and five tweaks that would bump the conversion rate by ten points. The designer iterates, hits “approve,” and before long the factory is churning out a product that already has a waitlist.
This is not possible today, but pieces of it already are. AI design tools for fashion already exist. Cre[ai]tion, for example, makes it easy to turn a sketch into a real product.
“Sensing” tools try to predict what products or fashions will be popular by sifting through social media and survey data. WGSN, for example, uses a variety of data points from surveys and social to predict what will be on-trend for fashion. The missing element is that it can’t say what a company should do based on their specific brand positioning and customer preferences, and it can’t predict if a specific product you create will succeed.
Synthetic‑user panels are the next step, and startups sell these today. They infer preferences from scraped reviews, prior surveys, or clickstream data and then let a language model role‑play the respondents. They provide enough insight to be useful—and they improve every quarter as more data becomes available and models improve. The tools today can weigh in on whether a new idea will succeed, but they have limited ability to assess truly innovative concepts not covered in their samples. They are helpful but companies should not make decisions based solely on these tools.
Coming Soon
Now fast forward to late 2026. Your virtual panel holds the entire purchase history of your loyalty app, 3P panel data on sales, a decade of focus‑group transcripts and surveys, etc. If the bot hits a blind spot, it will launch a micro‑survey, collect 500 video responses from real humans, fold them into its model, and try again. The time between coming up with an idea and validating the idea with a statistically significant sample is now less than a day and one hundredth the cost of the human-powered survey or focus group you might commission today.
That’s already amazing, but the bots will do more than validate ideas – they will come up with new ones. Bots will suggest new products and designs. Human designers will still steer the brand voice and decide whether to make the products it recommends, but more and more ideas will start with a suggestion from a bot.
How this changes the business
If the bots can truly anticipate customer demands with high accuracy, we’re looking at fewer SKUs, higher success rates, and lower write‑downs. A mass‑market apparel chain might drop from twenty‑thousand new SKUs a year to five thousand with high confidence. A luxury house could trim its runway assortment and focus marketing on the looks the model predicts will become icons. DTC brands might explore smaller micro‑niches because they can validate demand without a six‑figure sample order.
But this will also mean a complete rethinking of the design process. Companies might move from a 30 step process that takes weeks to create a new SKU to a 10 step process (or fewer) that takes a day and has AI doing many of the tasks.
Bottom line
Again, we’re still 18 months from bots that dramatically change the innovation process. If you work at or invest in a company that is dependent on innovation, start experimenting now with synthetic respondents to see the current state of the technology. Then, partner with those vendors to see how you can get your historical data into the bots. Before you know it, they will go from a novelty to a useful tool to a thought partner that accelerates innovation and dramatically improves the number of hits you get.
I didn’t want to go too far afield, but I pulled a lot of data. So, if you are interested in diving into Pyxis data to understand these kinds of questions more deeply, reach out to me.
When I was in college, I built a machine-learning model that could predict with decent accuracy whether a movie would be critically successful. A couple of years later I mentioned this to an analyst at a big movie studio. He said that they have models that are pretty accurate, but a big consideration is supporting passion projects for important people. Then, they’ll agree to do the money-making pictures. So, analysis can only get you so far.