Case study · Generative AI in marketing
StyleMod, a fashion retailer with about 4 million email subscribers, piloted generative AI to personalize its marketing. The numbers went up, and so did the complaints. Our group looked at where the AI created value, where it destroyed value, and how to keep the gains without the guesswork. The demo above is a working prototype of our recommendation: personalize with customers, not about them.
Value created
When each customer gets a message tailored to them, engagement goes up, and tailored messaging leads to higher retention.
The model withheld discounts from people who would have bought anyway, so StyleMod wasn't giving away margin or training customers to wait for deals.
Training an internal model meant less effort spent writing copy by hand.
Threats to value
The complaints were few next to 4 million subscribers, but they went viral, and StyleMod's brand depends on trust and inclusivity. One message like these can do more damage than any gain in click-through.
Recommendation
A better algorithm still guesses. Letting customers state their preferences removes the guessing. We'd keep the system with three guardrails, and add a short, optional style quiz backed by first-party behavior, so customers who skip it still get relevant messages.
Personalize from what customers buy and tell us, not from their browsing elsewhere.
A hard rule against inferring pregnancy, health or loss, whatever the model predicts.
Customers choose what to share, and can edit or delete it at any time.
The CFO's expansion plan
Built on first-party data and low risk.
Lower risk when the AI works within human-set spending limits and safety rules.
Chatbots talk directly to customers, so start with narrow tasks like FAQs and order tracking, and route sensitive topics to a person.
Personalized prices could make customers feel they're being treated unfairly.
From slides to prototype
We built the demo to show the recommendation working rather than just describing it. A rules-based engine stands in for the model, and comments in the source mark where a real generative model would plug in.
People and measurement
Their role shifts to training the AI, polishing high-profile campaigns and refining the brand voice, work that still needs human judgment.
The plan still saves on copywriting and scales the low-risk uses, while avoiding the cost of lost trust.
Track engagement and conversion alongside unsubscribes, complaints and repeat purchases, with a holdout group that gets no AI messaging to confirm the lift.
Q&A
A success on results, with a real trust risk that needs fixing.
The results are consistent with the AI driving it, and we'd confirm with a holdout group that doesn't get AI messaging.
A better algorithm still guesses. Letting customers state their preferences removes the guessing.
Most won't, so we'd keep it short and optional, and first-party behavior would fill the gaps. That way low participation doesn't erase the gains.
It's different because customers choose what to share and can edit or delete it.
Some don't, but the complaints show others do, and lost trust is hard to rebuild.
See it for yourself. Answer five questions and get three outfits built only from what you chose.
Try the demo