Taste vs data. Twelve years of buying says it isn't a fight.
A junior buyer once asked me, half-frustrated, half-curious: "If the algorithm says to drop this dress, why am I still in the meeting?"
It was a fair question. We had built the algorithm. It read sell-through. It read web traffic. It read the colour mix of last quarter's drops, the regional bias of yesterday's footfall, the variance between weekday and weekend lift, the way Tier-1 metros leaned one way on a silhouette and Tier-2 cities leaned the other. It read more numbers in a Tuesday morning than any single human had read all season.
It surfaced what looked like a clean call: cut the SKU, free the open-to-buy, reinvest in the silhouette that was lifting.
I said: the algorithm gives you the floor. Your job is to give it the ceiling.
That is the whole essay. The rest is just an explanation.
The argument retail likes to have
There is a tired debate that surfaces every time a head buyer is hired, or a new merchandising VP gets profiled in a trade rag, or some consulting firm reissues last year's report on the future of retail with a new cover.
Is buying a craft or a science? Is the future of retail merch a buyer with taste or an algorithm with a feedback loop? Should the buyer be in the room when the model speaks?
It is a debate that flatters the people having it. It pretends there are two camps. There aren't. The good operators in this industry have always done both. The bad ones have always picked one.
The bad-taste-only buyer is the person who fell in love with a print at a Paris market, bought it heavy on instinct, ignored every weekly sell-through pull from the floor, and lost a season. The bad-data-only buyer is the person who optimised every SKU to the regression line of last year, shipped a range that was forensically correct and aesthetically dead, and lost the brand. They both walk into the end-of-quarter review with confidence. They both walk out with a markdown problem.
The debate flatters because it makes the failing buyer feel like a misunderstood artist and the failing analyst feel like a forward-thinking technologist. Both are wrong in the same way. They have refused to use one of their two hands.
What twelve years inside a buying floor actually teaches you
I have spent twelve years on the same shop floor, mostly in women's Indo-Western. I have run trials. I have written the systems that read those trials. I have hired buyers and watched some of them turn into the operators I trust more than my own judgement. I have sat through more line reviews than I can count and I still take notes in every single one.
Here is what the floor teaches you, eventually, if you are listening.
One. Data tells you what happened. Taste tells you what is about to happen. The first is rigorous, the second is honest about being a bet. You need both because the past doesn't repeat, but it does set the lower bound. The buyer who ignores the past is wrong about the floor. The analyst who ignores the future is wrong about the ceiling.
Two. Taste is not a feeling. This is the claim that gets the most pushback from buyers themselves, because they like the romance of I just know. But what they actually have is a compressed memory of a thousand decisions and their outcomes, sitting in a person's head, made fast because the deliberate-thinking path would take a week per decision. A buyer who has placed three hundred orders in the same category isn't guessing when they say that print won't work in Bengaluru. They are pattern-matching against a private dataset that they will never write down, and the right model has never been built. Their brain ran the regression a million times faster than the laptop did, on data the laptop never saw, and produced a one-word answer. The mistake is to call that magic. It is just compressed evidence.
Three. Data without a buyer in the room is a regression to last year's mean. A model trained on last year's data, by definition, doesn't know what is about to be different about this year. It doesn't know that a film is going to drop in October and reset what young women want to wear to a wedding. It doesn't know that the price of a fibre is going to spike and the market is going to push toward a substitute. It doesn't know that a competitor just put a designer on staff. The buyer's job is to be the person who tells the model where the slope just changed.
Four. Taste without data is theatre. A buyer who can't be challenged by the numbers gets to be wrong with confidence, which is the most expensive way to be wrong. If your gut produces a strong signal and the sell-through pull says the opposite, you owe the question of which one is wrong. Sometimes the gut. Sometimes the data. Either way, the answer is in front of you.
Five. Both of them have to share a vocabulary. If the buyer can't read a sell-through curve and the analyst can't read a colour swatch, the meeting is just two people performing their own discipline at each other. The fastest path to a good range is one person, sitting alone with both the sell-through report and the trend board, choosing. That is more rare than it sounds. Most organisations split it into two seats and then hold meetings to try to glue the seats back together.
Six. A great buyer's biggest job is not picking the winners. It is killing the losers fast. Anyone can fall in love with a print. The job is to fall out of love with it on the right day. Data is the thing that lets you fall out of love on time. Taste is the thing that put you in love to begin with.
Seven. If your data system isn't producing a decision faster than your gut, the system is the problem. Most retail dashboards are written by people who have never been on the floor. They produce reports that buyers ignore. The right test for any dashboard is: does the buyer who ran the last category check this dashboard before lunch? If not, the dashboard is decoration. The buyer's instinct is right to ignore it.
Eight. The relationship between the buyer and the model is a long marriage, not a date. The first six months feel competitive. The buyer thinks the model is going to replace them. The model gets the colour mix wrong on a hero range and the buyer feels vindicated. The second six months feel collaborative. The buyer notices the model caught a slow drift in a region they had missed. The years after that, you stop noticing the seam. The buyer asks the model questions they wouldn't ask a junior. The model surfaces patterns the buyer would have caught eventually, but two weeks later. The category gets faster. The team gets quieter. The customer notices nothing, which is the point.
What "taste" actually compresses
I want to spend a paragraph on something the industry rarely says out loud. The buyers who keep beating the algorithm are not artists. They are walking neural nets with twenty years of training data and no logging.
Think about what a senior buyer's brain has actually seen. Twenty years of wholesale fairs in Paris, Milan, Mumbai, Delhi. Three thousand line reviews. Tens of thousands of fit sessions. Every meeting where a vendor missed a delivery and the buyer learned why. Every meeting where a buyer said no, this colour will not work in October and got overruled and got vindicated. Every season where a competitor launched a silhouette six weeks before they did, and the customer noticed, and the next year's calendar shifted.
That is not magic. That is a dataset of a few million labelled events sitting in a single person's head, with feature engineering done by lived experience, and a loss function tuned by the markdown ledger.
The reason the senior buyer's first reaction is often right is not a sixth sense. It is that their brain just ran the regression in 800 milliseconds, on a dataset that no engineer has ever exported. The reason it is often wrong is that their brain ran the regression on yesterday's data without noticing today's slope change.
The model and the buyer are running the same kind of computation. The buyer's edge is the richness of their training set. The model's edge is its appetite for new data. The pairing is so obvious in hindsight that it is embarrassing how long the industry has tried to make them choose.
A practical version of what this looks like
The system we run on the floor is built around a short trial-drop cadence. It writes itself in a sentence: ship a small set of styles to a handful of profile-clustered stores, give them a four-week shelf read, decide on the next buy from the actual data, scale or kill. We do this on the hero categories. We do not do it on the long tail. Long tail still gets old-school open-to-buy on intuition; the trial-drop machinery is for the styles the season turns on.
That sentence is half data, half taste.
The data half is everything the system does on its own. It picks the profile-clustered store panel by reading store DNA. It assigns the right depth across stores so each one has enough units to produce a clean sell-through signal. It triggers the auto-replenishment on the hero SKUs when sell-through clears a threshold. It surfaces the markdown candidate when it doesn't. It compares week-on-week velocity to a baseline that knows the seasonality of the category, the region, and the price band. It does this every Monday morning, before the buyer has sat down at their desk.
The taste half is what no system picks. Which silhouettes go into the panel in the first place. Which fabric story the season is going to be built around. Which print is the one the customer is going to want, six weeks before the customer knows it. That call is made by humans, looking at fabric swatches, watching what twenty-two-year-olds were wearing at the cafe outside the office, listening to what their daughters say about a competitor's window, reading the trade press, walking the floor, and then choosing. The system makes their choice faster, more measurable, less wasteful. The system does not replace the choice.
You can build the system without the choice. People do. The result is a category that runs like clockwork and drifts into irrelevance over four years, because the system is reading the past and the past is no longer where the customer is going. You can make the choice without the system, too. People do. The result is a category that ships beautiful things at the wrong price and prints a markdown every season.
The interesting place is the one in the middle. A buyer who reads the sell-through curve like a sentence, and a system that respects the buyer's veto.
What this looks like in the room
Picture a Monday morning. The merchandising room. Eight people around a table. A wall screen showing last week's sell-through across the hero panel, colour-coded green and red. Three swatches in the buyer's hand. A trade-press cutting on their tablet. A cup of cardamom tea getting cold.
The buyer is looking at a green tile on the screen and frowning. The tile says a particular drape is selling above forecast in Tier-1 metros, below forecast in Tier-2 cities. The model is recommending depth-up on the drape across the whole panel.
The buyer says no.
The room turns to me. I am the person who built the model. I look at the tile. I look at the buyer's face. I ask one question: what is it that you are seeing that the model isn't?
The buyer says: the drape is selling in Tier-1 because two cities had unusually mild weather last week. The drape needs lift to wear well, and lift doesn't read in twenty-eight degrees with seventy-percent humidity. The model is reading the green tile and thinking demand. It is actually reading weather.
I check the panel. The two cities named are the two with the unusual readings. The other Tier-1 cities aren't lifting the drape. The model is sampling a small window and confounding the variables.
We don't depth-up. The next week, the weather normalises. The drape's sell-through normalises. The buyer was right. The model was right too, technically. It just didn't know what week it was looking at.
I made one note in the model spec that night. We now adjust for a weather variance signal on hero-panel reads. The buyer's veto became a feature. The next time we run the trial-drop loop, the model is one beat smarter and the buyer is one beat freer.
This is how a category gets faster. Not by replacing the buyer with the model, or the model with the buyer. By giving the model the buyer's last veto as a new feature.
The seat I am writing from
I do both. I write the systems and I take the call. Twelve years gave me enough trial-drop cycles to know when the model is right and the print is wrong, and when the print is right and the model is just looking at the wrong window. The buyers who report to me know the system will not be defended when their gut is right. They also know their gut will be questioned when the sell-through says otherwise.
The model has a name on the buying floor. The buyers have names too. None of the names sound like the future of retail. The future of retail just sounds like a person who knows where the floor is and where the ceiling is, and has the patience to keep walking the difference.
A short note for anyone hiring for the seat
If you are hiring a head buyer or a merchandising VP and you find yourself drafting the role around either the buyer who can read data or the operator who can run AI systems, you are still in the old debate. The seat is one person who does both. They are not common. They are not exotic either. Look for the buyer who can also write the spec, and the operator who can also pick the colour. They have always existed. The job description just hasn't caught up.
Ask candidates a specific question in the interview: tell me about a time the model was right and you overrode it, and the override was wrong. The candidate who can't answer that has either never built the system or has never trusted the system enough to lose to it. Either way, that is your answer.
Closing
The longest careers in this industry belong to the people who don't end up on either side of the debate. They use the model when it helps. They override it when it doesn't. They have the meeting and they take the note and they ship the range and they read the sell-through and they ship the next range and so on, and so on, and the brand keeps getting where it needs to get.
That is the seat. Always up for a conversation about it. Or anything else on the floor.