Every item that comes through the door at Tatty's has already cost something before it reaches the sales floor. Someone has logged it, steamed it, photographed it, written a description, and put it on the rack. That all takes staff time, and none of it comes back if the item doesn't sell inside its cycle. For a consignment retailer, the sunk cost happens at intake. The only question left is whether the price gets the item out the door before the cycle ends.
At Tatty's, stock is processed in a central warehouse by a dedicated team, and each price came down to the judgement of whoever processed the item. A team member looks at a garment, weighs the brand, the condition and what similar things have sold for recently, and types in a number. That's a reasonable way to price one item. It's a much harder way to price a high volume of them consistently. Every person calibrates differently, memory of past sales fades, and nothing tells anyone afterwards whether last month's pricing was any good.
A suggested price, and how sure it is
We built a pricing assistant that sits inside the stock-entry screen the team already uses. As an item is entered, it returns a suggested price and a confidence band. A second model flags items that are unlikely to sell at all, at any price. That stock is often better donated, or discounted hard on day one, than kept in the cycle for weeks.
The team member still sets the price. Every suggestion is stored alongside the price they chose, so the system keeps a record of what it suggested, what a person decided, and what happened next. That record is how anyone can check, later, whether the model is earning its keep.
We trained it on the client's own sales history and tested it against their existing pricing before anyone relied on it. The case study covers how it is built and how we measured it.
The band widens when the data runs thin
The confidence band and the hard-to-sell flag are the actual product here. A single predicted number, with no indication of how sure the model is, invites people to either trust it blindly or write it off the first time it's wrong.
A confidence band tells the person pricing the item when the suggestion rests on plenty of similar past sales, and when the model is extrapolating from thin data: a designer label it has seen twice, a category that barely sells, a condition grade it's unsure how to weight. In those cases the band widens, and the decision stays with the person who can see the actual garment.
The hard-to-sell flag works the same way. The model says this item looks like the ones that don't move, which gives the team the chance to act earlier. They can discount it hard, bundle it, or move it on before it takes up weeks of shelf space.
What still takes judgement
A backtest on the client's own history gave us a sense of direction and scale. It can't tell us how buyers would actually have behaved at a different price, so we treat it as an estimate. The real answer comes from live stock, now that the system is pricing it, once it has run through full cycles.
The team now uses the assistant day to day, so that work has started: watching where people override it, checking whether those overrides were right, and feeding that back into the next round of training.
A system the team governs
A business of Tatty's size can now run a machine learning pricing system that would once have needed a dedicated data science team to justify. Accurate suggestions help, but the system is usable because it is explicit about where it's confident and where it isn't. The team can see where the model's reasoning runs thin, and step in there.
We think this is where this kind of tooling is heading for SMEs: a system a team governs, overriding it, watching where it's wrong, and feeding that back in. That loop is now running at Tatty's, built around the idea that the system should say when it's guessing.
The full case study, including how it's built and how we measured it, is at pattern.co.nz/work/tattys-ai-pricing-assistant.