How do you measure the ROI of an in-store AI stylist?

Compare conversion, average transaction value, units per transaction, full-price sell-through and kept rate in pilot stores against matched control stores over the same weeks, after an 8 to 12 week baseline, then convert the difference into gross margin and subtract the full cost.

What sales lift should a brand expect from AI?

Modest and uneven. Randomised field experiments in online retail found effects from none to +16.3% depending on the workflow, Zalando's size advice cut size-related returns 4.5% relative, and Stanford's AI Index found most reported marketing and sales revenue gains under 5%.

How do you measure the ROI of AI in fashion design?

Track brief-to-sample time, samples per adopted style, cost and time per approved asset, first-time approval rate, and the sell-through and markdown of AI-assisted lines against comparable lines in the same season.

Why are vendor AI sales figures often higher than pilot results?

They often compare shoppers who chose to use a tool with those who did not, or count sales broadly influenced by AI over long attribution windows. Amazon's Rufus figures and Salesforce's holiday figures both use such methods.

What is a control store?

A store as similar as possible to a pilot store, matched on sales and traffic trend, format, customer mix and assortment, that does not get the AI. The effect is how much pilot stores changed minus how much control stores changed over the same weeks.

How to measure the ROI of in-store and design AI

To measure the return on in-store and design AI, compare what changed where the AI ran against what changed where it did not, over the same weeks, and subtract the full cost. For store AI, that means conversion, average transaction value, units per transaction, full-price sell-through and returns in pilot stores against matched control stores, with a baseline before launch. For design and content AI, it means cost and time per asset or style, the share of drafts approved, and how AI-assisted lines sell against comparable lines. Expect modest effects: the controlled studies in our research found single-digit to mid-teens gains, and the large figures in vendor reports usually come from broad attribution or from comparing shoppers who chose to use a tool with those who did not.

This is the final part of our fashion AI research series. It completes the implementation guides after the 90-day pilot plan and the data and governance guide, and builds on our earlier article on how to measure an in-store AI pilot.

What effect sizes should a brand expect?

Start with what controlled studies have actually found, because it sets sensible targets and a sensible test size.

Study	Design	Effect
Fang et al., online retail field experiments	Randomised, millions of users, seven workflows	Sales from no effect to +16.3% by workflow; returns did not worsen
Zalando and TU Dortmund, size advice	A/B test, 132,000+ customers per group	Size-related returns down 4.5% relative
Amazon, size recommendation (RecSys 2017)	Offline test, then live A/B	17 to 21% better offline; +0.49% fit transactions live
Tan, Chandukala and Reddy, Journal of Marketing	Sales data before and after AR try-on	Effect larger for less popular brands, pricier items and new customers
Stanford AI Index 2025	Compiled company surveys	Most reported marketing and sales revenue gains under 5%

Two lessons stand out. Offline or lab gains shrink in live use, as Amazon's size paper showed; so judge only on live results. And the effect is uneven: the AR study found the biggest gains for less familiar products and new customers, so split your results rather than reading one average.

Which metrics measure in-store AI?

Use three layers, each explaining the next.

Layer	Metric	Definition
Use	Engagement share	Shoppers using the screen ÷ store traffic
Use	Try-ons and questions per day	Per store per trading day
Visit	Hand-off rate	Sessions ending in a hold, pay link, cart or staff call
Visit	Asked for, not stocked	Requests the store could not meet, by theme
Visit	Attach and step-up	Second pieces added; higher-priced alternatives chosen
Sales	Conversion	Transactions ÷ traffic, pilot against control
Sales	ATV and UPT	Average transaction value; units per transaction
Sales	Full-price sell-through	Share of tried pieces sold before markdown
After	Kept rate	Sales not returned once the return window closes
After	Follow-up sales	Sales to opted-in shoppers within a fixed window after a message

Asked for, not stocked deserves a line in the business case even if conversion barely moves, because no till records it. Out-of-stocks are costly across retail; IHL Group's 2025 research, as summarised by Board, put the global cost of inventory distortion at $1.73 trillion a year. We explain how to collect it in capturing unmet demand in store.

Kept rate guards against a stylist that sells pieces people return. The online field experiments found returns did not rise; check that your pilot agrees.

Which metrics measure design and content AI?

Design and content AI pays mainly in cost and speed, and only sometimes in sales. Measure both.

Area	Metric	What good looks like
Content	Cost and time per approved asset	Against your current shoot or agency cost; Zalando reports campaigns in days rather than six weeks
Content	Approval rate and rework	Share of drafts approved first time; edits per asset
Content	Product accuracy	Assets rejected for altering the product: should be zero
Design	Brief to sample time	Days from an approved brief to a physical sample
Design	Samples per adopted style	Physical samples made for each style that goes into the range
Design	Pre-launch signal	Votes or try-ons on concepts shown to shoppers before sampling
Range	Sell-through and markdown	AI-assisted lines against comparable lines in the same season

Zalando's FY2025 results are the clearest public benchmark for content: 90% of product content AI-made, campaign creation from six weeks to days, and 70% more content. Walmart said AI cut its fashion production timelines by up to 18 weeks. Both are company-reported. Your own before-and-after on cost per asset is easy to measure and more convincing.

Be careful with sustainability claims. We found no peer-reviewed study measuring a net reduction in overproduction from AI, and image generation used over 60 times the energy of text generation in a 2024 study. If you claim fewer samples, count them, and count the compute.

How should you design the test?

Stores: matched controls and difference-in-differences

Match each pilot store with two or three control stores on sales and traffic trend over the previous 12 weeks, format, customer mix and assortment. Collect 8 to 12 weeks of baseline. Run for at least 8 to 12 weeks live. The result is the change in pilot stores minus the change in control stores over the same weeks; that removes the weather, the economy and campaigns that hit every store. A falling market makes this essential: Sensormatic reported US in-store traffic on Black Friday 2025 down 2.1% on 2024.

Online: randomise

Randomise shoppers, not pages or days, into AI and no-AI groups, and keep the split stable across visits. The Zalando and Fang studies are templates: one clear outcome, a pre-set test length, returns and ratings tracked alongside sales.

Design: matched lines

Randomising a range is rarely possible. Instead, pick comparable categories or capsules, use AI in some and not others in the same season, and compare sell-through and markdown. Treat the result as evidence, not proof, and repeat over two seasons.

How big does the test need to be?

Small effects need more stores or more weeks. A standard rule of thumb: to detect an effect with 80% power at 5% significance, the effect must be about 2.8 times its standard error. Work out how much your weekly pilot-minus-control difference varies in the baseline period; if a 3% lift in conversion is smaller than that noise allows you to see, add stores or weeks before you start, not after.

How do you avoid fooling yourself?

The platforms' published figures show the common traps.

Self-selection. Amazon reported that shoppers using its assistant were 60% more likely to complete a purchase. Shoppers who choose to use a tool were often going to buy anyway. Compare stores or randomised groups, never users with non-users.
Broad attribution. Amazon's $10 billion Rufus figure used a seven-day downstream attribution model, and Salesforce counts sales "driven" by AI and agents broadly. Fix your attribution window and channel rules before launch.
Different baselines. Adobe found AI-referred traffic converting 9% below other traffic, while Salesforce found it far better than social traffic. Name your comparison.
Engagement as outcome. Users, sessions and interactions show interest. Zalando's pilot reported 40% more high-value interactions; its later add-to-bag and returns figures are the ones that touch margin.
Novelty. Early weeks run hot. Read weeks nine to twelve separately.
Moving the goalposts. Write the decision rule on day one.

What goes into the cost side?

Licences or service fees, hardware and installation.
Product data work: attributes, size charts, images.
Staff time: training, hand-offs, and approvers' time, which is easy to forget.
Usage costs for model calls and image generation.
Governance: privacy assessment, legal review, image rights.

ROI is then incremental gross margin (not revenue) from the difference-in-differences result, plus measured cost savings, minus these costs, over a stated period. Report payback in months alongside it.

What does the calculation look like in practice?

Here is the arithmetic with round, illustrative numbers. They are not a forecast or a benchmark; replace every one with your own.

Baseline. Over 12 weeks before launch, pilot stores convert 20.0% of traffic and control stores 21.0%.
Pilot. Over 12 weeks live, pilot stores convert 21.5% and control stores 21.3%.
Difference-in-differences. Pilot stores moved +1.5 points; control stores moved +0.3 points; the estimated effect is +1.2 points of conversion.
Translate to margin. Multiply the extra transactions implied by +1.2 points of conversion on pilot-store traffic by the average transaction value, then by gross margin, not by revenue. Do the same for any change in ATV, and subtract margin lost to extra returns if the kept rate fell.
Add measured savings. For a content pilot, the cost per approved asset before and after, times the assets you actually used.
Subtract the full cost for the same period, including approvers' and staff time.
State the uncertainty. Give the range, not only the point estimate. If the range includes zero, say so, and decide whether to extend the pilot or stop.

Notice what the control stores did here: without them, the pilot would have claimed +1.5 points, a quarter more than the estimated effect. That gap is why the method matters more than the dashboard.

Are the metrics different for jewellery?

A little. At a jewellery counter, track the share of try-ons that lead to a piece being taken out of the case, the attach rate of a second piece (earrings with a pendant, a band with a ring), step-up to a higher purity or a larger stone, and appointment show-up for bridal and high jewellery. If prices are built from a metal rate, track how often a piece could not be priced because the rate was stale; it should be rare, and it tells you whether the rate process works. For fashion, add units per transaction from completed looks and the full-price sell-through of pieces tried on the mirror. Adobe's data suggests jewellery converts better than apparel from AI-referred traffic online, so if your pilot includes AI visibility work, report the two categories separately.

What should the weekly report look like?

One page per pilot store, every week: use, visit and sales layers against control; kept rate once the return window closes; the top ten "asked for, not stocked" themes; open approvals; and any errors. At day 90, one page for the decision: the effect with its confidence range, where it was strongest, the cost, and the recommendation.

Ateliq's analyst produces a weekly report for each store and one across all stores, covering what was tried, asked for, bought, kept and returned, with what to do next. With door counts imported, it shows the share of visitors who used the screens and went on to buy; with Shopify orders, an ERP or a payment provider connected, each hand-off is marked kept, partly returned or returned once the return window closes. Those are inputs to the method above, not a substitute for control stores.

What is the bottom line?

The five voices in this series disagree about the size of the prize, but the evidence agrees on how to find out. Choose a narrow job, measure it against a control, count the full cost, and expect a modest, uneven effect that you can then scale where it is strongest. Start again at part one for the full map, or see the whole series.