Sell Through · Field Notes · Part 8 of 7
Where AI Fits Inside the Five Steps
The original Sell Through Field Notes series — seven essays on a store-origin SPA discipline — was written without much reference to AI. The framework holds without it. The five steps describe a way of running a specialty apparel business that worked before generative AI existed and will keep working when the current AI cycle is forgotten.
But that’s not the right place to stop. The honest reading, six months into actually deploying AI co-pilots inside a specialty apparel company, is that the framework didn’t change. The cost of operating it did. Some steps that used to require a department now require a person and a co-pilot. Some still require the department, but the work shifts from preparation to judgment. And one step — the one that matters most — gets exactly as hard as it ever was, because what it asks for isn’t automatable.
This eighth essay is the addendum to the original seven. It reads the five steps through an AI-native operating layer and tries to be precise about which work changes, which doesn’t, and where the boundary currently sits.
Step 1 — Define the ideal business structure
Before AI: A founder, a finance lead, and an experienced operator sit together for several weeks. They pull every cost line apart, benchmark against peer companies, argue about what each line should be as a share of sales, and arrive at a target P&L architecture that the rest of the business will be governed against.
After AI: The same conversation. Same weeks. Same arguments. The AI co-pilot doesn’t decide what the business should look like, because nothing in its training data answers that question — it can give you the distribution of what other companies do, but the question Step 1 asks is what you intend to do, which is a judgment about ambition.
What the co-pilot does change is the inputs. Benchmark queries that used to take a research analyst two weeks now take an afternoon. “What’s the distribution ratio for store labor in Japanese SPA brands between ¥10B and ¥50B in revenue?” used to mean a phone call to a consulting firm; now it means a query.
The work that disappears: data gathering, benchmark synthesis, comparable-company research.
The work that stays: the decision about what kind of business you want to be.
Step 2 — Manage the business by the plan
Before AI: A weekly P&L meeting. Someone — usually accounting — spends two to three days pulling actual results, formatting them against the plan, and producing the variance report the meeting will use. The meeting itself takes ninety minutes. By the time decisions are made, the data is between five and ten days old, and the gap between observation and action is meaningful.
After AI: The actual-vs-plan reconciliation runs continuously. By the time anyone walks into the meeting room, the variance report has already been generated, anomalies have already been flagged, and the unusual line items have a first-pass explanation attached. The meeting opens with the interpretation, not the assembly. Decisions land Monday afternoon rather than Friday afternoon. Three to four days of operational cycle time gets returned to the business.
This is the change that compounds the fastest. The meeting cadence didn’t change. The skill set in the meeting changed.
The work that disappears: report assembly, variance calculation, basic anomaly detection.
The work that stays: interpreting why a variance happened and deciding what to do about it.
Step 3 — Control costs
Before AI: Each distribution ratio (head office 5%, real estate 20%, store labor 14%, etc.) has a defined target. The mechanism to hold to it is a combination of policy, monthly review, and individual judgment. Drift is normal. Catching drift early requires someone whose job is to watch the numbers.
After AI: The watching scales. Distribution ratios can be tracked continuously, by line, by department, by store. A drift of more than half a point triggers a notification. The system surfaces the combination of drift signals — when head-office cost goes up and store labor stays flat, that’s a particular kind of problem; when the reverse happens, that’s a different one — and routes those signals to the right human.
But the response to the signal still requires judgment. Cutting head-office cost when it’s drifting up is not a mechanical decision; it’s a question about which capability the company can afford to give up. The AI knows the numbers are off. It doesn’t know which of three corrective actions matches the long-term strategy.
The work that disappears: continuous monitoring, threshold management, multi-signal correlation.
The work that stays: deciding which corrective action matches the company’s long-term strategy.
Step 4 — Respond to demand (the 52-week MD cycle)
This is where the AI conversation has been loudest, and where the actual change is most subtle.
Before AI: The merchandising team builds a 52-week plan with seasonal pivots. The plan represents a best-effort estimate of demand by week, by SKU, by store, with capacity, lead-time, and supplier constraints baked in. The MD team revises the plan against actual sell-through every week.
After AI: The plan can be revised more often, with more variables, more granularly. Forecasting models give the MD team a probabilistic view rather than a single point estimate. Capacity, weather, social-media signals, and competitive activity can be folded into the model as data sources rather than as private knowledge in one person’s head.
What doesn’t change: the seasonal pivot still requires a taste decision. Whether next spring is a brown season or a cream season is not a forecasting question; it’s an editorial one. The forecast tells you what to do with that decision once it’s made. The decision itself is the work.
The clearest signal that AI has actually landed in MD is not that the forecasts got better. It’s that the merchandiser starts the week thinking about the editorial question — what is the brand’s next move — instead of about whether the spreadsheet reconciled. We documented this concretely in the recent essay on operating headroom: when the operating MD started using AI tools on her own initiative, the time that opened up went straight into what more can we do to grow revenue?, not into more forecasts.
The work that disappears: data preparation, variance tracking, multi-source signal aggregation.
The work that stays: the editorial decision about what the brand will offer next.
Step 5 — Create customers
This is where the framework is least disturbed.
Defining the target customer, articulating the points of difference, holding to a brand position over a decade — these are not tasks AI accelerates. They are tasks AI is structurally bad at, because the relevant data (what the brand wants to mean to its customers) doesn’t exist in the AI’s training set. It exists in the founder’s head, in the company’s history, and in the unspoken consensus of the team.
What AI changes about Step 5 is at the margins. Generative tools can help draft campaign copy, propose visual references, surface trend signals from social platforms. None of that is the work of deciding what the brand stands for. The decision still has to be made by people who hold the brand’s accumulated context, and it still has to be defended over time against drift, fashion pressure, and the temptation to optimize for the current quarter.
The work that disappears: tactical content production, signal monitoring at scale.
The work that stays: virtually all of it.
The line as it actually sits
If I had to draw the boundary between what AI helps and what AI doesn’t, it would look like this:
| Activity | Who does it now |
|---|---|
| Data gathering, benchmark research, report assembly | AI co-pilot |
| Variance calculation, anomaly flagging, signal correlation | AI co-pilot |
| Forecast generation, scenario modeling, sensitivity analysis | AI co-pilot |
| Interpretation of why a variance happened | Operator + co-pilot |
| Decision about what to do in response | Operator |
| Definition of what the business is for | Founder / executive team |
| Editorial decisions about brand direction | Founder / MD lead |
| Long-horizon trade-offs between capabilities | Founder / executive team |
The line is not where most AI commentary places it. It’s not “AI replaces analysts.” It’s that AI compresses preparation time so dramatically that the bottleneck moves to judgment. The skill the operating team needs more of — not less — is the ability to interpret signals and make consequential decisions inside meetings, in real time, without the buffer of “let me get back to you next week with the data.”
Specialty apparel companies that adopt AI well are not smaller. They are more capable at the same size — because the part of the work that compounds (judgment) is freed from the part that doesn’t (preparation). That’s the addendum to the original five steps.
The framework holds. The cost of running it has come down. The skill of running it has changed shape.
That is what AI has actually done.
This essay extends the seven-essay Sell Through Field Notes series originally published in May 2026. The framework is detailed at length in Sell Through: The Five-Step System for Building a High-Margin Apparel Brand (Amazon Kindle, 2026). KITAGATA’s other writing — including the Japanese-language monthly “Next OS for Specialty Apparel” column — is at note.com/kitagata.