AI for beauty brands: closing the online shade-matching gap.

The problem

The single biggest gap between shopping beauty online and in a store is that a customer can't test a foundation shade or formula against their own skin before buying, so guesswork drives returns, one-star "wrong shade" reviews, and customers who quietly never try the brand again.

How it's done manually

Static shade charts and generic "fair, medium, deep" labels that don't account for undertone, and can't account for how a formula behaves on different skin types.

The AI solution

A shade-and-formula-matching model uses uploaded photos, past purchase and return data, and stated skin-type inputs to recommend the closest shade and formula match across the catalog, including cross-brand shade equivalence when a customer names a shade she already wears.

Example workflow

A customer enters a shade she wears from another brand; the model maps undertone and depth against the catalog's shade range and recommends the closest match, along with a formula suited to her stated skin type.

Business impact

Fewer "wrong shade" returns and reviews, and more first-purchase confidence in a category where getting it wrong once often means the customer doesn't try that brand again.

Estimated ROI

Brands with a wide shade range, twenty or more shades, typically see more return reduction here than a brand with four to six shades, since the matching problem is smaller when there's less range to get wrong.

Common mistakes

Launching shade-matching without accounting for undertone, only sorting by depth, when two shades can be the same depth and completely wrong undertones for a given customer.

Best practices

Capture undertone and finish as structured data per shade, not just a marketing name, since the model can only match on data that actually exists.

The problem

A customer buying one serum still has to guess at the cleanser, moisturizer, and SPF that actually work with it and with her skin type, and most beauty sites only recommend the next single item, not a coherent routine.

How it's done manually

A merchandiser manually curates "complete your routine" bundles that are the same for every customer, regardless of skin type or actual purchase history.

The AI solution

A recommendation model builds a routine, not just a next item, from quiz answers, skin type, and purchase history, adjusting recommendations as a customer's history grows.

Example workflow

A customer completes a skin quiz flagging combination skin and sensitivity; the system recommends a full routine set built for that profile, and updates the recommendation on repeat visits based on what she's already bought.

Business impact

Higher units per order from routine-based buying, and returning customers get recommendations that reflect what they've already built rather than starting over.

Estimated ROI

The upside shows up in average order value and repeat-purchase rate on routine-driven categories, skincare especially; a brand selling mostly single-use or gift-oriented products has less of this dynamic to capture.

Common mistakes

Recommending a routine based only on the last item purchased instead of the full quiz and purchase history, which produces generic pairings that ignore stated skin concerns.

Best practices

Refresh the underlying quiz and routine logic periodically, since skin concerns and product formulations both change and a routine model trained once gets stale.

The problem

Beauty customers increasingly shop by exclusion, sulfate-free, fragrance-free, non-comedogenic, and a catalog tagged only by product type and price has no way to answer that search.

How it's done manually

Ingredient information sits in a paragraph on the product page, unsearchable and unfilterable, so a customer has to open and read every product to check.

The AI solution

An AI-assisted tagging layer reads ingredient lists across the catalog and generates structured, filterable attributes, free-from claims, key actives, comedogenic rating, kept current as formulations change.

Example workflow

A customer filters for "fragrance-free" and "non-comedogenic"; the system returns an accurate filtered set because ingredient data was tagged at the formulation level, not guessed from the product name.

Business impact

Customers with specific sensitivities or preferences can actually find what they need, instead of bouncing to a competitor whose site supports the filter.

Estimated ROI

The impact concentrates on categories with real sensitivity-driven shopping, skincare and haircare; a decorative-only category like nail polish sees less demand for this kind of filtering.

Common mistakes

Auto-generating ingredient tags once and never re-running the process when a formulation changes, so the filter quietly becomes inaccurate.

Best practices

Re-tag ingredient attributes whenever a formulation update happens, and treat this as an ongoing data process, not a one-time project.

The problem

A one-size-fits-all subscription box either over-delivers products a customer will never use or under-delivers on what would actually keep her subscribing, and manually curating a personalized box for every subscriber doesn't scale past a small list.

How it's done manually

Someone manually assembles box variants for a few customer segments, or the same box ships to everyone regardless of stated preferences.

The AI solution

A personalization model assigns each subscriber's box contents from quiz data, past ratings, and purchase history, at a scale no manual curation process could match.

Example workflow

A subscriber rates her previous box's items; the model adjusts her next box's product mix based on those ratings plus her skin-type profile, without a human manually reassigning her segment.

Business impact

Lower churn from boxes that feel personalized rather than generic, and less manual curation overhead as the subscriber list grows.

Estimated ROI

The clearest signal to track is churn rate on personalized versus non-personalized cohorts; brands running a single universal box have more room to gain here than ones already segmenting manually.

Common mistakes

Personalizing at signup only and never updating the model as ratings and purchase history accumulate, so the box stops improving after month one.

Best practices

Feed post-box ratings back into the model continuously, and pilot on a subset of subscribers before rolling personalization out to the full list.