Can AI recommend a comfortable bra? The evidence does not yet connect body scans to fewer returns
A new review finds useful work on body shape, pressure and wearer experience, but no validated end-to-end system linking those components to comfort or business outcomes. Detailed body data also create privacy obligations that a fit score cannot solve.
The short version
- What happened
- Published work measures different things: morphology classification, garment mechanics, short wear tests and longer wearer experience. Those outputs cannot be substituted for one another.
- Who it affects
- Shoppers may welcome less trial and error, especially when physical stores offer limited sizing. A system that asks for intimate body data should also provide a non-scanning route, clear deletion and reuse controls, and evidence that its recommendation reflects comfort rather than only geometry or sales conversion.
- What remains unproven
- Critical narrative review
Keep in mind
Critical narrative review
Purposive rather than systematic sampling
No fixed study denominator
No duplicate screening
No complete risk-of-bias assessment
No meta-analysis
Several direct studies available only through abstracts
No validated end-to-end retail system or causal business outcome
Source analysis · 1 linked source. Primary analysed source date: 28 October 2026.
Read the full report · 4 min The evidence, methods, implications and limitations
Body shape, pressure and comfort are different targets
The review separates four concepts that retail systems often collapse: a size label, physical fit, perceived comfort and appearance. An algorithm can classify a body shape or estimate a measurement without knowing how one brand's garment will move, press or feel during a particular activity. Pressure sensors and simulations can explain mechanics, but they cannot replace a person's account of comfort.
The reviewed direct evidence includes 3D morphology studies of 201 female college students and 150 individuals, a ten-person running study, a seven-person pad comparison and a 14-week military field study that enrolled 93 recruits. Each contributes one part of the chain. None alone establishes that an AI recommendation works across brands, bodies, activities and sustained wear.[1]
A recommendation needs a product and a real outcome
A retail claim must specify what improves. Measurement accuracy needs reference measurements and errors in physical units. Product ranking needs real alternatives and defined fit labels. A comfort claim needs repeated wearer reports under stated conditions. A claim about fewer returns needs a live comparator, reason-coded outcomes, follow-up and a design capable of separating the service from price, availability and return costs.
The review found component-level progress rather than an empty field. Its central gap is integration: body representation must connect to a particular garment; geometry or pressure must connect to a wearer's experience; and the recommendation must connect to an attributable outcome. A purchase or retained item is not automatically evidence of comfort because customers may face cost, inconvenience or limited alternatives.[1]
The distinction that matters
One retail claim, four separate evidence steps
A body scan is only the first link in a fit-and-comfort decision.
Measure
Estimate dimensions or morphology
Report physical error and uncertainty
Match
Rank real garments in one product range
Test product-specific fit labels
Prove
Follow comfort and return outcomes
Compare with usual shopping and record privacy effects
Sensitive body records add a separate test
Detailed scans or derived body features may remain sensitive even after names and faces are removed. Linking them to an account, transaction or outside image can increase identification and inference risk. The authors propose collecting only data needed for the service, separating identity from measurements, limiting raw-image retention, controlling reuse and vendor access, and keeping a manual-measurement route available.
Those proposals are a governance framework, not evidence that a deployed service is private, accurate or profitable. The review itself used purposive sampling across five databases and gives no fixed denominator; it did not perform duplicate screening, a full risk-of-bias assessment or a meta-analysis. A controlled, multi-brand field evaluation with diverse participants, sustained comfort measures and documented privacy controls would materially strengthen the case.[1]