Clothing store network · 30+ locations · 2026
AI Cross-sell Stylist — a stylist that works the whole network 24/7
A good salesperson never says «here’s your dress, goodbye». They say: «have you seen this belt? It’s from the same collection». Good salespeople are rare, and they don’t work 24/7. AI became such a stylist — for all the stores at once.
// Business context
A network of 30+ clothing stores. Women’s collection + men’s + accessories. The key growth channel is repeat purchases and accessory cross-sell.
In a physical store, an experienced salesperson sees what you’re choosing and offers to round out the look — accessories, items from the same collection. They know what goes with what. Online there’s none of that: just a «similar products» block that usually shows ten more of the same dress.
// What hurt
- Consultation quality was a lottery. Stores with the best stylist-salespeople noticeably outperformed the rest on average order value — but you can’t «scale a stylist» by hiring.
- An online channel with no consultation. The buyer alone with the catalog — and cross-sell barely happens.
- The catalog was underused. The core assortment sells, while customers simply don’t see the relevant accessories.
- Cross-sell was random. The standard «recommended for you» shows products with no collection logic.
// How I thought about the task
I started not with «let’s add recommendations», but with a question: why do stores with the best salespeople sell more? The answer — a stylist who sees context. So the task isn’t «a recommendation algorithm», but «to reproduce expert advice» — and that’s a fundamentally different tool.
Most recommendation engines in retail are collaborative filtering: «whoever bought this also bought that». For clothing that works poorly: someone who bought a dress doesn’t need any bag. They need a bag that matches.
So the solution is an AI agent that understands styling logic: season, the collection’s color palette, drop, context. A vision model categorizes the catalog, and the stylist answers in chat.
The difference between «similar products» and «this belt from the new collection would go with this dress» is the difference between an algorithm and a stylist.
// What the stylist can do — four principles
Accessories for a specific item: bags, belts, jewelry, scarves — by styling logic, not statistics.
Items from the same collection match best — the stylist knows this and suggests within the drop.
Knows when to suggest and when to stay quiet. If a sale is on, it mentions discounts on relevant items.
Specific sizes or custom requests — escalation to a salesperson at the nearest store, with the dialog context.
// Stack
| Layer | Tools |
|---|---|
| AI brain | Claude with image vision (categorization + recommendations) |
| Catalog | The network’s product catalog API |
| Chat UI | On-site chat + the network’s messengers |
| Hosting | Cloudflare Workers (edge, low latency) |
| Storage | Sessions with limited TTL, no PII |
| Analytics | Owner dashboard: recommendations, dialog conversion |
// Result
- Cross-sell feels like a stylist’s advice, not «buy more» — recommendations within the collection, accounting for season and moment.
- The best salesperson’s expertise scaled across the whole network — 24/7, no weekends or shifts.
- Complex requests aren’t lost: escalation to a live salesperson with the full dialog context.
- Owners see in the dashboard which collections and pairings actually work.
The network’s specific commercial metrics are the client’s internal data, and I don’t publish them. Only what can be confirmed is published here. How it would look for your catalog — we can work out on discovery.
// What went wrong — lessons learned
The first versions were too pushy. The stylist «kept pressing» — on every question it offered several extra products, and customers felt the pressure. We adjusted the prompt: at most 1-2 relevant suggestions, and only when there’s clear context.
Categorization errors on new drops. The vision model confused evening and casual on certain fabrics. We added a human-in-the-loop: the collection manager validates the first SKUs of a new drop, then the AI takes over.
At first salespeople saw the AI as a threat. Over time it became the opposite: the stylist takes the «easy» questions, while salespeople focus on the complex clients the AI escalates to them.
// FAQ
How is the AI stylist different from Shopify’s «recommended products»? +
Standard recommended products runs on collaborative filtering — «people who bought this item also bought that one». The AI stylist understands context: it pairs a dress with a bag of the right size, a coat with a scarf from the same collection, accounting for seasonality and color. It’s expert advice, not statistics.
How much does it cost to implement an AI stylist? +
It depends on catalog size, channels, and depth of integration — the Grow2.ai team gives an estimate after discovery. The strategic part — whether the stylist gives leverage to your specific assortment, and where — we work through together on an advisory session.
How does the AI know about collections, colors, seasonality? +
We connect to the product catalog via the store’s API. We categorize products with a vision model (Claude with image input). Styling rules are written together with the collection manager — like brand guidelines.
What about products the AI isn’t «trained» on yet? +
We set up a graceful fallback: if confidence is low, the AI honestly says «I don’t have a better recommendation for this pair». Silence is better than a bad recommendation that erodes trust.
And what about customer data — is it GDPR-compliant? +
Yes. The AI stylist does not store PII (personal data). It works with anonymized session_id; interaction history has a limited retention period. No sharing with third parties.
Case author
Andrew Maryasov
AI consultant, founder of Auspex (CRM automation) and Grow2.ai (AI agents + community). 25+ years in business automation: accounting → CRM → AI. I build AI strategy for owners and teams; implementation runs through my own brands.
// Other cases
// Conclusion
The win came not from the model but from reframing the task: not «a recommendation algorithm», but «to reproduce a stylist’s expert advice». When you correctly name what is actually needed, the choice of tool becomes obvious.
Implemented via Grow2.ai — AI agents and recommendation systems. I build the strategy personally; the implementation itself is run by the Grow2 team with my active involvement.
Are you in retail / e-commerce — with cross-sell left to chance?
Let’s start with a conversation: where the lever is hidden in your catalog. We build the strategy together; implementation runs through Grow2.
Personal advisory →