Work / Case study 06

NuORDER by Lightspeed · 2025 · Concept complete · Research in progress

AI Assortment Assistant

Retail buyers spend hours gathering context before they add a single product to an assortment. This is one connected AI workflow that compresses that work, doing analysis first and creation second, so buyers spend their time refining, not starting from zero.

Domain
Wholesale · Retail planning
My role
Lead Product Designer
Timeline
2025
Status
Concept complete · In research
NuORDER Assortments grid after the AI wizard runs: a Balanced version added banner reading 75 styles, 86% category coverage, projected $250,250 retail at 55% margin, above a grid of product tiles, with the Assortment assistant open on the right offering suggested questions
Project gallery

The full arc at a glance - the analysis assistant that shipped first, the seven-question creation wizard, and the hand-off that connects them. Browse it here, or read the full story below.

View full

01The problem

Every assortment starts from a blank page.

Retail buyers build seasonal assortments by working from requirements set by their divisional merchandising manager, then manually reviewing previous buys, analysing past performance by category, and building a plan from scratch. Before a single product is added, a buyer has already spent hours pulling last season's numbers, cross-referencing category targets, and reconstructing a picture of what worked and what didn't.

This preparation work is real, necessary, and skilled. But much of the data needed to do it sits in systems the buyer has to navigate separately. The information exists - it just isn't connected at the moment decisions are being made.

ERP system Previous assortments Brand line sheets Sell-through reports

Four disconnected sources a buyer reconciles by hand - every season, before the building even begins.

02The opportunity

The platform already held the data.

NuORDER Assortments connects to a retailer's ERP, which means historical assortment data, past-season performance, and category-level sell-through signals are all available at the moment a buyer is planning their next buy.

The opportunity was to make that data visible and useful, not as a report the buyer had to go find, but as a layer of intelligence embedded directly into the assortment workflow.

03Phase 1 · AI analysis inside the assortment

Make AI useful where buyers already work.

Before designing a creation workflow, the first step was to make AI useful in the place buyers already spent their time: inside an existing assortment. The AI Assistant shipped as an analysis layer embedded in the assortment page, working on any assortment, whether a buyer built it manually, copied it from a previous season, or imported it from a template.

The assistant didn't make decisions. It helped buyers see things that are hard to see in a grid of 80 products, and gave them a starting point for improving what they had.

NuORDER Assortments grid with the Assortment assistant panel open on the right, answering a Break down margin prompt: blended margin 49% ranging 30-61%, five styles under the 40% floor highlighted, margin % by class bars, and a lowest-margin styles list, with follow-up questions below

The assistant sits alongside the assortment grid. Answers reference the buyer's live data. Here, a margin breakdown that flags the styles under the 40% floor and ranks the lowest-margin ones with follow-up questions to keep going.

Questions buyers could finally ask in place.

What is the category mix?
Which products are over- or under-represented?
How does this compare to last season?
Which styles are new versus proven?
Are there gaps in the selection?
Does it support the selected doors and deliveries?
What could be removed or replaced?
Shipping this first established trust before asking buyers to rely on AI for something more consequential.
Exploration · How it answers

A blank chat box is intimidating in a brand-new paradigm, so the default state opens with a greeting and suggested questions - showing buyers what the assistant can do and giving them a confident first move. From there I explored how an answer should feel: scannable findings grounded in the live buy, not a wall of text.

Two assistant states side by side: a default state with a greeting and suggested question chips (Summarize this buy, Break down margin, Find gaps and outliers, Compare to S24 plan, Which styles are under 40% margin), and an answered state showing a margin breakdown, margin-by-class bars, an under-40% flag, a lowest-margin styles list, follow-up chips, and an open chat-history menu
Default and answered - the empty state opens with a greeting and suggested questions to scaffold the first move; running a prompt returns grouped, scannable findings - a margin breakdown, flagged outliers, a ranked list of lowest-margin styles, and follow-up chips to keep going - with saved chats kept per assortment.
Exploration · Where it lives

Because the assistant has to answer about the grid the buyer is looking at, I tested two placements against the live assortment - weighing how much of the buy stays in view while a buyer asks, and how a flagged answer can point straight back into the table.

AI assistant as a floating window anchored to the bottom-right of the NuORDER Assortments grid, with a greeting and suggested question chips over the product table
Floating window - anchored bottom-right and layered over the grid, so the buyer keeps the full assortment in view while they ask.
AI assistant docked as a bottom sheet that expands upward from the bottom edge of the assortment grid, showing the conversation and a margin breakdown while the table rows stay readable above
Bottom sheet - docked along the bottom edge and expanding upward; the grid columns stay readable above it.
04Phase 2 · The creation wizard

From a blank grid to a first draft.

The next step was to extend the AI experience earlier in the journey - into the moment before an assortment exists at all. Instead of starting from a blank grid and analysing it afterward, the wizard gives the buyer a first draft. A short set of questions, each one giving the assistant the context it needs to generate a relevant recommendation.

Entry point - Create with AI, or set up manually
New assortment modal with an assortment name field, deliveries selector, door chips, and two paths: Create with AI (recommended) or Setup manually

The buyer names the assortment, selects deliveries, and chooses doors. The AI path is positioned as the faster route, not the only route - buyers who already know exactly what they want can skip the wizard entirely.

Five questions, each narrowing the draft.

Every answer hands the assistant another piece of context - so the recommendation it returns is focused, not generic.

1BrandWho is this buy for?
Wizard step 1 of 5, Brand: choose which brand the assortment is for, with single-brand guidance

Single-brand buys are recommended - they can be shared directly back to the brand partner via Order Intent.

2SeasonWhat's the window?
Wizard step 2 of 5, Season: Fall/Winter, Spring/Summer, Holiday, or Resort options

Seasonal context narrows the product universe to what's relevant, so the assistant isn't pulling from the full catalog.

3CategoriesHow deep?
Wizard step 3 of 5, Category: multi-select chips for Footwear, Jackets, Tops, Bottoms, Dresses, Accessories or Full collection

A multi-select. The more specific the buyer, the more focused the output - or pick Full collection to cover the season.

4DirectionHow to build it?
Wizard step 4 of 5, Strategy: radio options including Prioritize bestsellers, Prioritize new collection, Balanced mix, Good/better/best, Match a price range

One consolidated question - how should this assortment be built? One answer, clear intent.

5SizeHow big?
Wizard step 5 of 5, Size: Small 25-50, Medium 50-100, Large 100-150 products

Small, Medium, or Large. This gives the assistant a target and prevents over-generating - a common way AI tools erode buyer trust.

6GenerationShow the work
Generation step: a Building your assortment options screen for Trix Fall/Winter 2025 (Jackets, Tops, Footwear) with a progress checklist - Reading the Trix linesheet and Scoring sell-through from past seasons complete, Matching styles to your strategy in progress, and Drafting three ways to assort it pending

Before results appear, a loading screen shows what the assistant is doing. Trust starts with understanding what produced the output.

05The assistant proposes · The buyer decides

Three ways to assort the buy - not one answer.

Rather than returning a single take-it-or-leave-it result, the assistant presents three versions to compare - each sized, costed, and scored from past seasons, with a side-by-side table. It leads with a recommendation, but the buyer picks whichever fits.

BalancedRECOMMENDED

Newness, proven, coverage

Proven sellers anchor the volume, newness keeps the floor current, and core carryover closes coverage gaps. The dependable choice for a full seasonal assortment.

Newest first

Fresh, fashion-forward floor

Front-loads this season's new collection, leaner and higher-AUR. Best for trend-led doors and early-season drops.

Commercially proven

The safe, commercial buy

Weighted to styles with the strongest historical sell-through. Best for high-volume doors and reorders.

Choose an assortment version screen: three cards - Balanced (recommended) with 76 products, Newness-first with 62, Commercial/proven with 82 - each with coverage, units, retail, margin, style mix and category mix, above a compare table

Each option carries the numbers a buyer reasons with, coverage, units, retail, margin, AUR, % new, and a compare table puts them side by side. The assistant has proposed; the buyer decides.

06Review & confirm

Nothing is added until the buyer says so.

Before anything lands in the assortment, the buyer reviews the full suggested product list. They can select all, untick individual styles, check the category distribution, and confirm.

This step preserves the principle that runs through the entire experience: the assistant generates a starting point; the buyer makes the call.

Refine - Balanced review screen: a list of 76 styles with checkboxes, Proven/New/Core badges and a reason on each row, with 75 of 76 selected and an Add 75 to assortment button
07 · One connected workflow
"The wizard creates the first draft. The assistant continues the work."

Once confirmed, the buyer lands on the assortment page with the AI Assistant already open and a dismissible banner explaining what was created. From here they're inside the analysis experience that shipped in Phase 1. So, no context switch, no starting over.

Phase 1 · shipped first

The analysis assistant

Helps buyers understand assortments they already have, category mix, gaps, new vs proven, what to remove. Established as a trusted tool before anything more consequential.

hands off ↑↓ continue refining
Phase 2 · built second

The creation wizard

Helps buyers generate a strong first draft faster. When it completes, it hands off directly to the assistant, which is already open, already grounded in the new assortment.

The assortment grid after generation: a Balanced version added banner summarising 75 styles, 86% coverage and projected retail and margin, with the Assortment assistant panel open on the right showing a greeting and suggested questions like Summarize this buy, Break down margin, Find gaps and outliers, Compare to S24 plan, and Which styles are under 40% margin

The summary banner names the recommendation used, the styles added, category coverage, and the balance of new versus proven - and the assistant is open right beside it, so the buyer keeps the conversation going without moving.

The AI Assistant helps buyers understand assortments they already have.
The creation wizard helps buyers generate a strong first draft faster.
When the wizard completes, it hands off directly to the assistant.
Buyers analyse, refine, and adjust without switching context or starting over.
08Research in progress

The concept is built. Now we validate it.

The concept is complete and the screens are built. User research is underway with retail buyers to validate the wizard flow, test the three-option recommendation model, and understand how buyers want to interact with the assistant after the initial assortment is created.

CalibrationDoes the wizard feel too long, too short, or well-calibrated?
TrustDo buyers trust the recommendations enough to use them as a starting point?
BehaviourHow often do buyers modify the AI-generated assortment versus accepting it as-is?
ExplainabilityWhat does "explaining a recommendation" need to look like to be genuinely useful?
09Reflection
"Designing for AI means designing for trust - and trust is built incrementally."

Shipping the assistant as an analysis tool before building the creation wizard wasn't a compromise - it was the right sequence. Buyers who'd already used it to understand an existing assortment were in a far better position to trust it when asked to generate one from scratch.

01

Sequence builds trust

Analysis first, creation second. Each phase shipped with real value, and the second landed on a foundation buyers already trusted.

02

One question beats two

Consolidating type and strategy into a single Direction step was a small decision with an outsized impact on how the wizard feels. When two questions ask the same thing, ask one.

03

Reduce uncertainty, don't maximise automation

The loading screen, three options instead of one, the review step, the assistant that stays open - each gave buyers enough visibility and control to make the AI a collaborator, not a black box.