Every AI-in-fashion article is about the runway end: generated campaign imagery, virtual try-on, trend forecasting models that claim to know next summer’s silhouette. Meanwhile the actual money in most brands leaks out the back office, in returns processing, supplier email archaeology, and spreadsheets that nobody has reconciled since the last drop.
I run operations for a DTC apparel brand doing seven figures. Over the past year we quietly put AI to work on three back-office problems, and the combined effect on our margins has been bigger than anything we’ve done on the marketing side. None of it will ever be a press release. All of it is repeatable by a two-person ops team.
Problem one: returns reasons are a gold mine stored as sludge
If you sell clothes online, 20 to 35 percent of it comes back. That’s the industry’s open secret, and most brands treat the returns pipeline purely as a cost center: receive, inspect, restock, refund, move on.
But every return arrives with a reason, and the reason field is where the gold is. The problem is the format. Our returns portal gives customers a dropdown plus a free-text box, and customers, bless them, use the free-text box like a diary. “The medium fit fine in the shoulders but the sleeves hit weird and honestly the green is more olive than the photos.” That is four separate signals in one sentence: partial fit issue, sleeve grading issue, color accuracy issue, photography issue.
Multiply by 400 returns a month and you have a dataset nobody reads. We used to have an intern tag a sample of 50 per month by hand. The tags were inconsistent month to month, and 50 is not a sample size you re-cut a size curve over.
What we do now
We pull the raw returns export into our ops sheet, and a formula classifies every free-text reason into our taxonomy: runs-small, runs-large, fabric-quality, color-not-as-pictured, style-changed-mind, damaged. Literally a formula in the reason column’s neighbor, the same way you’d write a VLOOKUP. We moved our ops sheets to a platform that has AI inside spreadsheet formulas, so the classification step is just =CLASSIFY pointed at the free-text column with our six categories, and it fills down the whole export like any other formula.
The unglamorous magic is that it’s a column, not a chatbot session. Nobody copies text into a separate tool and pastes labels back. New returns land, the column classifies, the pivot on the other tab updates. The AI usage is metered in dollars with a hard ceiling we set, so a 4,000-row month can’t surprise us on cost; in practice a month of classification costs us less than the coffee budget for the returns intern it replaced, and the intern now does work that needs a human.
What we found once every return was tagged: two SKUs accounted for most of our runs-small tags, both from the same factory, both graded off the same block. We fixed the block. Returns on those styles dropped by a third the next season. That is a pattern 50 hand-tagged rows per month never surfaced.
Problem two: supplier emails are unstructured POs
Our fabric mills and factories communicate exclusively in prose. “Hi, the 240gsm loopback is delayed to the 14th but we can ship the rib and the jersey Friday, quantities as per your last mail except the ecru which is short 40m.” Somewhere in that paragraph are three ship dates, three quantities, and a shortage that will quietly wreck a colorway if it doesn’t make it into the plan.
For years the workflow was: read email, retype into the production sheet, make one typo per week, discover the typo at the worst possible moment. Now the workflow is: paste the email text into a staging column, and an =EXTRACT formula pulls the structured fields (material, quantity, promised date, exceptions) into their own columns. A human still eyeballs the row before it’s confirmed, because a factory’s “the 14th” can mean two different months in the wrong week of December. But eyeballing a structured row takes ten seconds; retyping a paragraph took two minutes and carried the typo risk.
We do the same thing with weekly production status updates. A =SUMMARIZE formula condenses each factory’s long email thread into a two-line status per PO, which is what actually gets read at the Monday meeting. The threads are still there when we need the detail. The summary is what keeps the meeting to twenty minutes.
Problem three: the photo backlog (a smaller win, but real)
Between drops we sell through a long tail of core styles, and the PDPs for core styles rot: old photography, missing detail shots, flat lays from three brand iterations ago. We now run a quarterly audit where every core SKU gets a row and a checklist of required imagery. For the gaps we can’t shoot immediately, decent stock and licensed imagery fills the editorial slots (not the product shots, obviously, those must be the real garment). When we need editorial filler for lookbook-adjacent blog content, we usually go digging through Green Stock Media before paying a big library’s subscription price.
The audit itself is boring. That’s the point. Boring, on a schedule, beats heroic and never.
The pattern behind all three wins
Notice what these have in common. In each case the AI is not making a creative decision or predicting a trend. It is converting unstructured text that already exists in your business into structured columns you can pivot, chart, and act on. Returns prose becomes a tagged dataset. Supplier prose becomes dated, quantified rows. Long threads become two-line statuses.
Fashion ops has always been rich in this kind of dark data. Customer service transcripts, stockist feedback, fit notes from the sample review, wholesale order emails. The bottleneck was never insight, it was transcription. Nobody was ever going to pay a human to structure all of it, so it stayed sludge.
The second thing they share: the AI lives where the work already lives. Our team was never going to adopt a separate “AI returns intelligence platform” with its own login and its own dashboard. They already live in a grid. A function in a cell asks nothing of them they don’t already know. The whole rollout was one afternoon, most of which was arguing about the returns taxonomy, which is the part that actually deserved the argument.
Where I’d tell a smaller brand to start
If you’re doing even 50 returns a month, start with returns classification. It’s the highest signal per dollar, the taxonomy forces a useful internal conversation about what you believe your fit problems are, and the output feeds directly into next season’s size curves and size guide copy.
Then do supplier email extraction the first time a retyped date costs you money. It will happen this quarter. It always does.
The runway end of AI in fashion gets the headlines because it makes pictures. The back office end makes margin. I know which one my P&L prefers.