The Question Returns Has Never Asked
Across a few events and conversations over the last couple of months, one idea has come up consistently and in different forms. Our CTO Carlos Blanco has been turning it over in the context of returns, and what it means for brands who are still using the same tools they had ten years ago.
Here is his thinking.
The question has changed
One of the more useful framings for what AI actually represents came up at an event in London recently. The argument was this: the first industrial revolution scaled human labour. What AI is scaling is human cognition. And that shift changes the fundamental question you are allowed to ask.
For most of human history, ambition has been a function of individual capability. You set your goals based on what you can do. The honest version of that question is: given what I know, what I can afford, and what I have time for, what is achievable?
CEOs have always operated under a different constraint. They do not ask what they can do. They ask whether it can be done, and then curate the capabilities needed to answer yes. The distinction matters because the two questions have entirely different ceilings.
What AI does is make the second question available to everyone. For the first time, ambition does not have to be bounded by individual skill. The question shifts from what can I do to can it be done. That is a larger change than it sounds. Most industries have not yet worked out what it means for them specifically.
Returns are still asking the old question
In returns, the old question sounds like this: given what the customer filled in on the return form, what do we do with this item?
That question has shaped the entire infrastructure. Dropdown reason codes. Manual inspection on arrival. Condition assessments made by whoever is working the returns bay that day. Decisions routed by habit rather than evidence. The whole system is built around what has always been extractable from a return, which is very little.
The reason codes most brands collect are not useless. Sizing issue, changed my mind, not as described. But they are a heavily filtered signal. Customers do not always know precisely why something did not work. The decision to return an item is often not a single moment of clarity. It is a gradual realisation that happens over days, sometimes weeks, and by the time the return form appears the original reasoning has already been simplified and partially lost.
As agent-mediated purchasing becomes more common this gets structurally worse. When an AI selects and orders an item on a customer’s behalf, the person filling in the return form may have had limited involvement in the original decision. What they write in the reason field reflects even less of the actual story.
Part of the problem is that a return form is the wrong type of interface for the question being asked. A form is a search interface. You present a query and receive a categorical answer. That works when the information is readily retrievable. It does not work when what matters is not a fact but a path. Why did this item come back is a navigational question. The customer’s path from purchase to return involves expectation, comparison, and realisation. Compressing that path into a dropdown does not summarise it. It erases it.
Brands have been asking one question of every return: what can we extract from it? That question has a ceiling. And the ceiling is low.
What any intelligent system actually needs
There is a principle that holds across every domain where AI is now being applied to real operational problems. The quality of the output is determined almost entirely by the quality of the context available before the system starts working. This is not a model selection problem. It is a data capture problem.
You see it in document retrieval systems, where encoding the right metadata at ingestion time matters more than the sophistication of the query. You see it in voice agents, where every decision the system has to make at runtime adds latency, and latency in a live conversation is experienced directly as uncertainty and incompetence. The teams shipping the most capable systems are not the ones with access to the best models. They are the ones who did the most work before the system was ever invoked.
The pattern is consistent: the moment of capture determines everything downstream. If the right context was not collected at the right time, no amount of downstream intelligence recovers it.
In returns, that moment is singular. It is when the customer has the item in hand, knows why it is coming back, and has not yet sent it. Everything after that is reconstruction. The item goes into transit. Days pass. The customer moves on. The warehouse receives something in a bag with a label. Whatever story existed when the return was initiated has been almost entirely lost by the time anyone with the capacity to act on it gets involved.
The question that changes everything
The new question is: can we capture the full story of why this item came back, at the only moment that story exists completely?
The answer is yes. But the window is narrow.
I wrote earlier this year about the photograph at the point of return, the starting point for condition data that most brands are not capturing. But condition is only one of the two questions. The other is why. Why did this item come back? The photograph does not answer that. Voice does.
At the moment a customer initiates a return, they have the item in front of them. They know exactly what they think of it. Their words in that moment, before the decision has been rationalised and before the reason code has compressed it into a category, carry genuine signal. So does the condition of the item before it is packed and handled. So does the product metadata that already exists from the original order.
Voice and image, captured together at initiation, before the item ships. That is what changes when you ask the new question.
A voice note describing the return in the customer’s own words is not the same thing as a dropdown selection. It captures intent. It captures the distinction between an item that was genuinely defective and one that simply did not match expectations. It captures the nuance that matters for routing. Whether an item is resellable as-is, whether it warrants repair, whether the issue is a product quality signal worth feeding back to the brand. None of that is in the reason codes. Most of it is in what the customer says in thirty seconds if you give them the opportunity.
An image at intake gives you condition before the item has been touched or processed. It creates an objective record that is independent of whoever handles the return that week and separates what the customer reported from what is actually true of the item.
Resell, repair, recycle, write off. That decision can be made in under sixty seconds with this evidence, rather than days later from incomplete information at the warehouse. Returns cost brands roughly forty percent of retail price to process. Most of that cost reflects decisions made without adequate context, resolved slowly, at the wrong stage.
The compounding case
Every return captured with voice and image is a labelled data point. The condition of the item. The customer’s own description of why it came back. The product category, brand, and material. The time since purchase. The routing decision made and the outcome that followed.
That dataset does not exist anywhere at scale. There is no corpus of returns data with this level of structure and specificity. The general models have not seen it because it has not been captured.
Brands that start capturing it now are building something that becomes genuinely difficult to replicate. A model trained on consistent, structured return data from a specific category is more useful than any general model. What a lightly worn cashmere piece looks like after six months versus eighteen. How a particular defect presents across different weights. What voice signals correlate with fraudulent returns versus genuine dissatisfaction. That specificity only comes from operational data captured consistently over time.
The ceiling on what can be extracted from a dropdown is fixed. The ceiling on what can be learned from structured voice and image data captured at scale is not. That is the difference between asking what can we do and asking whether it can be done.
Carlos Blanco is CTO and co-founder of CIRQUEL. If you are a brand, investor or operator thinking about returns, reverse logistics, or AI applications in physical supply chains, we would love to connect at cirquel.co.