A Practical AI Assortment-Planning Workflow for Small Teams

Assortment planning is not the task of predicting which six garments will be fashionable. It is the decision of how much breadth and depth to buy under limits: cash, rack space, pack sizes, delivery windows, category roles, size coverage, margin, and the risk of unsold inventory.

AI can help estimate demand, identify similar products, or generate scenarios. It cannot remove the uncertainty from a new style with no history. A useful workflow therefore starts with a decision the team can explain and a baseline it can reproduce.

1. State the decision before selecting a model

Define the planning horizon, channels, location, customer, and unit of decision. Are you deciding which styles enter a six-week pop-up, how many packs a boutique buys for a season, or which online-only variants receive photography?

Write the objective in business language. “Maximize predicted demand” is incomplete. A small retailer may instead want to maximize expected contribution while preserving minimum category coverage, staying within cash and space limits, and restricting the loss from leftover inventory.

Every objective embeds a point of view. A model that rewards only margin can eliminate discovery items. One that rewards only unit sales can overfill the assortment with low-margin basics. Surface those tradeoffs before code turns them into a score.

2. Build a style-level fact table

For every candidate, record what the team knows and what it is estimating:

• style and variant identifiers;
• category and intended assortment role;
• pack size, minimum order, maximum available quantity, cost, price, and expected margin;
• launch and delivery windows;
• color, material, silhouette, size range, and fit relationship;
• display or storage burden;
• comparable styles and the reason they are comparable;
• demand estimate, uncertainty range, and source date;
• cancellation, return, markdown, and replenishment assumptions.

Do not let the system treat missing values as zero. “No return data” and “zero returns” are different facts. New styles also need a cold-start rule. The product-age fashion forecasting paper is one example of research designed around the unusual lifecycle of fashion products, but its reported results do not automatically transfer to another retailer, category, or season.

3. Establish a transparent baseline

Before adding AI, calculate a simple feasible buy. A spreadsheet, small script, or mixed-integer model can enforce constraints that a conversational prompt may overlook.

Google’s OR-Tools introduction describes optimization in terms of an objective and constraints and distinguishes feasible from optimal solutions. That language is useful even when the team never uses the library. A feasible assortment respects the rules. An optimal assortment is only optimal for the specific objective, data, and assumptions supplied.

FashionMember created a synthetic six-style baseline in content/data/FM-007-assortment-scenario.csv and a reproducible enumerator in scripts/fm007-assortment-scenario.php. It requires at least one pack of every fictional style, respects a $2,500 buy budget and 140 space units, and maximizes a simple expected-contribution expression with a penalty for units above the point forecast.

The script selected 15 units of JKT-01, 25 of TOP-01, 20 of TOP-02, 20 of DRE-01, 10 of BOT-01, and 5 of ACC-01. The fictional buy cost was $1,870, used all 140 space units, and produced an objective value of $4,410.

Those numbers are not a recommendation or performance result. They prove only that the code can produce one feasible answer from the supplied fictional assumptions. The model omits substitution, size-level demand, supplier risk, time, and forecast uncertainty.

4. Give AI bounded jobs

AI is more useful when assigned a component of the workflow rather than the final buy.

Comparable retrieval: group a new style with genuinely comparable historical products. Require the model to show which attributes drove the match and let a merchandiser reject superficial similarity.

Demand range preparation: produce low, base, and high estimates from documented inputs. Preserve the original data, method, and version; do not accept a number that cannot be traced.

Substitution hypotheses: identify which products might absorb demand if one item is absent. Treat the output as a testable hypothesis, not observed customer behavior.

Constraint checking: extract pack sizes, lead times, and delivery windows from approved vendor records, then verify every extracted value against the source.

Scenario explanation: translate changes between two runs into plain language: which constraint bound, which styles entered or left, and what assumption caused the change.

The final quantities should come from an auditable calculation and named human approval.

5. Model fashion-specific tradeoffs

Classic assortment research already shows why the problem is not simple. Kumar Rajaram’s fashion assortment methodology addresses breadth, depth, basics versus fashion goods, and the challenge of short lifecycles and limited history. More recent omnichannel assortment research examines physical and online assortments together rather than treating channels as isolated shelves.

A small team should at least test:

• what happens if demand is 25 percent lower or higher;
• what happens if one delivery is late;
• whether an online style can substitute for a store item;
• whether minimum category representation is worth its cost;
• how size curves change inventory risk;
• how return rates and markdown recovery alter contribution;
• whether the model repeatedly excludes new designers or unfamiliar customer needs.

There may be several defensible assortments. Show the frontier instead of hiding every choice inside one score.

6. Evaluate the process, not just sell-through

Use a forward holdout period or a pilot store when possible. Compare the AI-assisted workflow with the existing method and a simple numerical baseline.

Track forecast error by style age and category, full-price sell-through, stockouts, leftover units, gross-margin return on inventory, returns, transfers, editing time, and how often merchants override the suggestion. Record why they override it. An override may reveal model failure, new information, or an undocumented business rule.

NIST’s AI RMF Core calls for documenting expected benefits and costs, comparing with appropriate benchmarks, testing under deployment-like conditions, and monitoring over time. The framework is voluntary and broader than retail assortment planning, but those measurement principles fit the task.

A weekly operating rhythm

For a small independent team, the process can fit into one focused session:

1. Freeze the candidate list and source data.
2. Review missing facts, late vendors, and new information.
3. Run the existing baseline unchanged.
4. Run low, base, and high scenarios.
5. Use AI only for documented comparable, extraction, or explanation tasks.
6. Review constraint binding, exclusions, and unusual quantities.
7. Record human overrides with reasons.
8. Approve the buy and archive inputs, outputs, model version, and date.
9. Compare the decision with actual outcomes when enough time has passed.

The archive is more valuable than a polished dashboard. It lets the team learn which assumptions failed.

Where the assortment scenario stops

The worked example uses fictional data and a simplified single-period objective. It does not model customer choice, substitution, sizes, delivery timing, replenishment, returns, cash timing, or uncertainty distributions. Its enumeration approach is suitable only for a tiny scenario. A real assortment requires merchant, finance, operations, and supplier review and should not be delegated to an AI output.

Sources and verification

• Rajaram: Assortment planning in fashion retailing — primary research on breadth, depth, fashion lifecycles, and a retailer application.
• Google OR-Tools: About optimization — objectives, constraints, feasibility, and solver classes.
• Product age based demand forecast model for fashion retail — primary research on new-product lifecycle forecasting and assortment use.
• Omnichannel assortment optimization under the MNL model — primary research linking physical and online assortment decisions.
• NIST AI RMF Core — benefit/cost mapping, benchmarks, deployment-context testing, and monitoring.
• content/data/FM-007-assortment-scenario.csv and scripts/fm007-assortment-scenario.php — FashionMember’s fictional, reproducible baseline executed August 31, 2026.