AI That Styles 4.2M Clients
61% Better on the First Box
Agix built the AI Stylist Copilot for Stitch Fix, a pre-selection engine that narrows 10,000+ inventory items to a curated shortlist for each client, slashing stylist time per box by 58% while increasing keep rates and client lifetime value.
America's leading personal styling service, where human taste meets algorithmic scale.
Stitch Fix is the pioneering personal styling platform that combines the art of human stylists with the precision of data science. Founded in 2011, the company serves over 4.2 million active clients across the US and UK, each receiving curated "Fixes" of clothing and accessories selected specifically for their size, style, and budget.
As the client base grew past four million and inventory expanded to 10,000+ active SKUs, the core operation hit a scaling wall: each stylist manually reviewed hundreds of items per Fix, spending the bulk of their time on inventory search rather than the high-value judgment of style curation. Keep rates for new clients, the critical metric for long-term retention, lagged because first-fix personalization wasn't sharp enough. Agix was brought in to build the AI layer that makes every stylist faster, smarter, and more personal.
Human stylists can't manually search 10,000 SKUs for every client, and keep rates suffer when they try.
The Stitch Fix model is brilliant in concept: a human stylist who knows you personally selects five items you'll love. The problem is that "knowing you personally" requires reviewing thousands of options to find the five best matches, and at 4.2 million clients, that search cost is unsustainable. Stylists were spending most of their time on inventory triage, not style judgment. First-fix keep rates paid the price.
End-to-end AI. From client onboarding to continuous learning.
Style discovery, fix curation, inventory sync, feedback loops, and business intelligence, all connected through a single AI platform that gets smarter with every interaction.
AI Stylist Copilot: pre-selection, personalization, and a learning feedback loop.
Agix built six integrated AI systems that work together, giving every stylist a 20-item AI shortlist instead of 10,000 raw SKUs, while learning from every kept and returned item to sharpen future Fixes.
AI Stylist Copilot, Pre-Selection Engine
Narrows 10,000+ active inventory items to a curated shortlist of 20 for each Fix, ranked by predicted match score across size, style, fit profile, color preference, and price sensitivity. Stylists review the shortlist and make final selections, spending their cognitive energy on judgment rather than search. The system achieves a 95% match accuracy on shortlisted items, meaning 19 of 20 items on the list are genuinely viable candidates.
Client Insights Platform
A 360° view of each client's preferences, purchase and keep history, style quiz responses, and behavioral signals, synthesized into a living style profile that updates after every Fix. For new clients with limited history, the platform uses quiz data, stated preferences, and cohort-level signals from similar clients to bootstrap personalization from day one instead of waiting for enough history to accumulate.
Fit Match & Style Discovery AI
Combines size measurements, past fit feedback ("too tight in the shoulders," "runs large"), and body proportion signals to predict fit likelihood at the item level, not just the size level. A client who wears a size 8 but has left consistent feedback about certain cuts running small gets recommendations adjusted accordingly. Fit Match scores surface in the stylist's shortlist alongside style match scores, letting stylists make informed tradeoffs between fit confidence and style novelty.
Inventory Optimization & Demand Forecasting
Forecasts demand at the SKU level by client segment, factoring seasonal trends, return rates, and keep velocity by style cohort. Items with high predicted keep rates for a given client segment are prioritized in the shortlist engine, ensuring the AI recommends items that are actually in stock and likely to perform well, not just stylistically aligned. This reduced excess inventory in low-keep categories by 24% within the first two quarters post-launch.
Returns & Feedback Learning AI
Extracts structured signals from every kept and returned item, not just a keep/return binary but the free-text feedback, star ratings, and item-level notes clients provide. A client who returns a blazer noting "too formal for my lifestyle" won't see formal blazers in the shortlist for their next Fix. A client who keeps a relaxed-fit denim consistently gets more relaxed-fit denim in future shortlists. The model updates client profiles after every Fix, compounding personalization quality over time.
Performance Analytics & Merchandising Intelligence
Real-time dashboards for stylist performance (keep rate by stylist, Fix throughput, feedback quality), merchandising insights (keep rate by brand, style category, and price tier), and trend signals (emerging preferences across client cohorts). Buyers use the merchandising intelligence to inform assortment decisions, reducing investment in style categories with structurally low keep rates while increasing depth in high-performing categories.
AI pre-selection. Faster styling. Better personalization.
The Copilot gives every stylist an AI-curated shortlist of 20 items, ranked by match score, annotated with fit confidence and style profile alignment, so they spend their time on the judgment calls that actually require human expertise.
Better keeps. Fewer returns. Stylists who spend time where it matters.
Measured against pre-deployment baselines across Stitch Fix's active client base post-launch.
Before the Copilot, I'd spend 40 minutes just filtering inventory before I could even start thinking about what the client actually wants. Now the AI hands me 20 great options and I spend my time on the part I'm actually good at, making the judgment call on what this specific person will love.
AI that made stylists better, not redundant, and unlocked a personalization flywheel.
The Copilot succeeded because it attacked the right bottleneck. The problem was never that human stylists lacked taste, it was that inventory search consumed so much of their time that taste barely had a chance to operate. By eliminating 70%+ of unproductive search time, the AI gave stylists the space to do the work they're uniquely good at.
The feedback loop is what compounds the value. Every kept and returned item teaches the model something specific, not a generic "this client likes casual" but "this client returns blazers but keeps relaxed blazers in navy or olive." That granularity is only achievable through a structured AI learning layer, and it's what separates a one-time accuracy improvement from a continuously improving system.
Eliminating search unlocked styling capacity
The inventory pre-selection step was pure overhead, no customer value, just necessary friction. Eliminating it through AI didn't change what stylists do; it gave them the time to do it properly. Throughput per stylist increased dramatically without any loss in Fix quality.
First-fix quality is the retention lever
The first Fix is the one that determines whether a new client stays or churns. Cold-start personalization using quiz data and cohort signals, rather than waiting for real purchase history, meaningfully improved first-fix keep rates, starting the personalization flywheel earlier for every new client.
Granular feedback beats binary keep/return
Structured extraction of free-text feedback, not just a keep/return signal, produced learning signal 10x richer than the raw binary outcome. "Too formal" and "wrong fit in shoulders" are different signals that produce different future recommendations. That granularity compounds into dramatically better personalization over time.
Inventory intelligence closed the buying loop
Feeding keep-rate analytics back to the buying team changed what merchandise was purchased, not just how it was recommended. Style categories with structurally low keep rates got less investment; high-performing categories got more depth. The AI improved both supply and demand sides simultaneously.
What this system doesn't solve.
Personalization AI compounds with history. It has real constraints in the early stages of a client relationship and at the edges of taste exploration.
New Clients Have Limited Personalization Signal
The system is most powerful for returning clients with multiple Fixes of history. New clients rely on quiz data and cohort signals, which bootstrap personalization but don't match the accuracy achievable after 3–5 Fixes. First-fix performance improved substantially with the AI, but it remains the weakest point in the personalization curve. The model gets meaningfully better after each Fix.
Style Exploration Is Hard to Model
When a client wants to explore a new style direction, trying maximalist after years of minimalist, or workwear after exclusively casual, the model initially resists based on history. Stylists can override the shortlist with exploration-focused picks, but the AI's prior-based shortlist is less useful in these moments. The model doesn't yet have a reliable signal for "this client wants to change."
Fit Prediction Has Physical Limits
Fit match scores improve keep rates but don't eliminate fit returns entirely. Body proportions and garment construction interact in ways that are difficult to capture from measurements and feedback text alone, particularly for structured garments like tailored blazers or fitted trousers. Fit prediction is significantly better than chance but not yet close to perfect, especially for new brands without established fit history in the system.
Requires Clients to Provide Feedback
The learning loop quality is directly proportional to the quality of client feedback. Clients who return items without leaving detailed notes, or who leave only the minimum required response, generate weaker learning signal. The AI can infer some things from patterns across clients, but item-level personalization improvement is slower for clients who don't engage with the feedback form. Feedback completeness is an ongoing UX challenge.
Is AI Stylist Copilot the right build for your business?
What powers this system.
AI Predictive Analytics
Client preference modelling, keep rate prediction, and cohort-level personalization signals, the data layer that powers shortlist accuracy and cold-start recommendations for new clients.
AI Process Automation
Inventory pre-selection, workflow orchestration, and feedback extraction automation, eliminating the manual overhead that limits stylist throughput at scale.
Conversational AI
Personalized client communication at every touchpoint, from Fix delivery notifications to style check-ins, driving engagement and richer preference signals between Fixes.
Retail AI Systems
AI purpose-built for fashion and specialty retail, from personalized curation and demand forecasting to returns reduction and assortment intelligence across large catalogs.
Demand Forecasting AI
SKU-level demand prediction by client segment and season, ensuring the shortlist engine recommends items that are in stock and aligned with proven keep-rate performance by category.
Stylist Performance Intelligence
Keep rate dashboards by stylist and Fix type, feedback quality tracking, and merchandising analytics, giving operations teams the visibility to coach stylists and optimize buying decisions.
Common questions about AI-powered personalization for retail curation.
The AI is a copilot, not a replacement. Stylists still make the final selection for every Fix, the AI handles inventory pre-selection, surfacing a ranked shortlist of 20 viable candidates from 10,000+ SKUs. The stylist reviews the shortlist, makes judgment calls about which five to send, and writes the personal note. Human taste and client relationships remain central to the Stitch Fix model, the AI removes the search overhead that was consuming stylist time.
New clients complete a style quiz that captures size, stated preferences, lifestyle context, price sensitivity, and style direction. The model uses quiz responses combined with cohort-level signals from clients with similar profiles to generate a meaningful shortlist from day one. Cold-start performance is weaker than for returning clients with multiple Fixes of history, but substantially better than pure random selection. After 2–3 Fixes, the model has enough real behavioral signal to move past cohort inference into genuine individual personalization.
Yes, stylists can always search inventory and select items outside the AI shortlist. Override behavior is tracked and fed back into model improvement: when a stylist substitutes an item outside the shortlist and it gets kept, that's a learning signal for what the model missed. Stylists who override frequently in specific style categories help identify blind spots in the model. The system is designed for stylists to improve it, not just use it.
Feedback is extracted at the item level, not just a keep/return binary but the free-text notes, star ratings, and specific attribute feedback (fit, style, quality, price). A natural language processing layer structures the free text into preference signals: "too boxy" updates fit preference signals; "love the color but not my style" updates color vs. style weighting separately. Client profiles are updated after every Fix, so the shortlist for each subsequent Fix reflects everything learned from all previous interactions.
Yes, the core pattern (large inventory + human curators + client feedback loop) applies broadly across subscription box and curation models in beauty, home goods, food, specialty retail, and beyond. The specific models need to be trained on domain-appropriate data, but the architecture, pre-selection engine, client insights platform, feedback learning loop, and performance analytics, transfers with configuration adjustments. Businesses with 500+ SKUs and a meaningful feedback infrastructure are typically strong candidates for this approach.
Ready to build AI that makes every curator faster and every recommendation sharper?
Most projects go from kickoff to deployed AI system in 8–16 weeks. Let's talk about what a pre-selection engine and personalization learning loop could do for your keep rates and stylist capacity.