AI in Apparel Merchandising: Where It Belongs, Where It Doesn't

Every apparel software vendor has been forced to publish an AI announcement in the last 18 months. Most of them are answers to a different question than the one apparel operators are actually asking. The vendor question is "how do we keep up with the AI cycle." The operator question is "where does this actually save time, and where is it noise."

We have an opinion on both. Here is the honest taxonomy.

Where AI clearly belongs

There are three places in apparel merchandising where machine learning moves the needle today, in production, with real teams:

Sell-through pattern detection during the selling window. Identifying which styles are tracking ahead of, at, or behind plan — and surfacing the magnitude — is a pattern-recognition problem with sufficient data to do well. The models do not need to be exotic; the value is in compressing the signal-to-action lag from weeks to days. This is the highest-leverage AI in apparel right now and the easiest one to verify.

Size curve construction from historical data. Building size curves from prior-season sell-through, accounting for stockouts and substitution, is a cleaner problem than it looks. The judgment-vs-data ratio in size planning has been wrong for a long time — most teams default to a regional template and then complain that the curve is wrong. AI-derived curves, validated against actuals, consistently outperform manual curves once a brand has 4+ seasons of clean history.

Anomaly detection in operational data. Stockouts, lead-time slips, IMU drift, channel-mix shifts. The data exists; the bandwidth to monitor it does not. ML models that flag the 1% of patterns worth investigating, out of the 99% that are noise, are how mid-market teams cover the gap left by enterprise BI. This is unglamorous and high-leverage.

Where AI is a distraction

There is a second category where AI is technically possible, sometimes useful, but not where mid-market teams should be focused right now:

Long-horizon demand forecasting at the SKU level. Apparel is fashion-volatile. The fundamental thing that makes a forecast useful — a stable underlying demand signal — does not exist for new styles in a fashion category. Models that promise SKU-level demand forecasting 9 months out are mostly redirecting the team's attention from the part of the problem they can actually fix (the buy commitment, the carry-over decision, the in-season reaction speed). The honest framing: forecasting helps at the category level over a long arc; it does not help at the SKU level for a new style.

Generative AI for assortment ideation. Tools that propose assortment combinations from style libraries are interesting, occasionally useful, and not where mid-market teams will get the next 5 points of margin. The bottleneck is not coming up with combinations; it is committing the right depth and positioning the right inventory. AI ideation is downstream of the constraint.

Conversational interfaces over the planning database. "Ask the data a question in natural language" sounds like a productivity unlock. In practice, planners spend 90% of their time on the same 6 questions, all of which are better served by a purpose-built dashboard. The conversational interface is a demo feature, not a production feature.

Where AI is snake oil

A third category, which is the one the apparel software market has been most active in over the last 18 months:

"AI-powered" anything where the AI is a generic LLM wrapper around the same forecasting logic that has been there for a decade. The label is new. The capability is not. If the announcement does not name the specific model, the training data, and the metric that improved, the announcement is marketing.

"Autonomous merchandising" or "AI buyer" framing. Apparel merchandising is a high-judgment, high-context discipline. The decisions that matter — what to commit to, who to source from, when to mark down — depend on inputs that are not in the data. Tools that pitch autonomous decision-making in this space are either over-claiming or have been built for a category of brand (pure-play digital, single-channel, replenishment-driven) that does not look like most apparel businesses.

"Predictive returns" and similar narrow ML claims, layered on top of unconnected planning data. Returns prediction is real and useful. Most of the vendors selling it are doing it on top of a planning data set that is not connected to the supply or assortment plan. The model output cannot drive a different decision because the decision-making layer above it cannot consume the signal. The AI works; the system around it does not.

Our position

For mid-market apparel brands, the practical AI roadmap looks like this:

1. →Year 1: Sell-through pattern detection. Anomaly flags on inventory and lead times. Size curve validation against actuals. Three things, all operational, all measurable in margin terms.
2. →Year 2: Replenishment proposals informed by current production WIP and seasonal-arc rules. Markdown-timing recommendations against weeks-of-supply by store. Same theme — applied ML where the data is rich and the signal is clean.
3. →Year 3+: Whatever the market has actually proven by then. Not whatever the vendor announcements claim now.

The brands that will benefit most from AI in apparel planning over the next five years are not the ones that adopt the most AI features. They are the ones that adopt connected planning first, so that when AI capabilities mature, the decision-making layer can actually use them. The bottleneck has always been the disconnected stack. AI does not fix that — it amplifies whatever is underneath.

See how the connected planning layer makes AI capabilities operational.