What are AI agents for fashion brands?
An AI agent is only as good as the data infrastructure beneath it. For fashion brands, this is the core problem: the live, garment-level understanding of what a consumer audience wants to buy next does not exist in a form an agent can use. Ask any capable assistant what a brand should make next and it will reach for training data that is six to twelve months old, or search a handful of articles. The result is a generic answer, not because the model is weak, but because the specific, live infrastructure it needs is missing.
We build that infrastructure. Hugo creates a standing model of a consumer audience from public conversation, images, and video. It is not a study or a dashboard. It is a live map of who consumers are, what they wear, and what they want, resolved down to the specific garment. This is the hard part, and it is the part we own. On top of this infrastructure, we build and deploy agents that work for a brand around the clock, surfacing decisions in the tools they already use. This article explains how that infrastructure is built and what it makes possible.
What is a fashion audience map?
Mapping an audience means building a live, standing model of who your consumers are, what they think, what they need, and what they will buy next. It is not a report or a study. It is infrastructure. Unlike social listening, which tracks keywords you already know, an audience map is designed to find the demand that has no name yet. It is built from the ground up by observing real people and the products they use, not by asking them questions or prompting an AI to imagine their behavior.
In our work mapping cities like Paris, London, and New York, this means resolving consumer behavior down to the specific garment. Our New York map, for example, identified 832 distinct garments across 876 real accounts. This is the substrate. An agent is that infrastructure pointed at a single commercial decision: what to stock, who to partner with, which product to fix. It works for the brand around the clock, because the infrastructure it runs on is always on.
How do you find the right people to map?
Finding a representative audience is the first hard problem. Search bars and follower lists are poor instruments for this. They return the loudest voices or the most obvious candidates, not the people who genuinely represent a scene or a city. To solve this, we use a method from survey science called respondent-driven sampling, or chain referral. We start with a small set of seed accounts that fit the audience definition, then trace the real connections, engagements, and co-occurrences outward.
This process is how we reached the 876 accounts that form our New York map. It is designed to find the clusters of real people who influence each other, not just the individuals with the highest follower counts. It is a slower, more computationally intensive process than just scraping a hashtag, but it is the only way to build a map that reflects the real structure of an audience rather than the structure of a social media platform's algorithm.
How do you know what people are actually wearing?
You observe it from images and video, at scale. The critical finding from our research is that only 2% of public posts are explicit "outfit posts". The other 98% is where the real signal lives: in the background of a photo, in a video of something else, in ordinary life. This is the data that most methods miss entirely. Our process is built to capture it.
For the New York map, we processed 4,464 posts from the 876 accounts, which yielded 9,146 garment crops. We use computer vision models to detect and crop every piece of clothing, then use vector search to cluster them. This allows us to identify the same exact item being worn by different people. In New York, we found 169 garments that appeared across multiple, otherwise unconnected accounts. This is how a trend is measured before it has a name: as an object propagating through a network of real people.
What does an audience map show you?
A finished map is a graph. The nodes are people and the garments they wear, and the edges are the connections between them. It is a live model of how style moves through a specific community. For example, in our New York map, a single pair of white sneakers sits on seven accounts. By following the connections, we see the same people also wearing a specific crew tee, a style of jeans, and a pair of leggings. Their neighbors in the graph carry it further, down to a woven leather bag that appears on four other accounts.
This is not an abstract analysis. Every connection is auditable. Every garment cluster links back to the original images it was found in. A brand can see the evidence for every claim. The map is not a static document; it is standing infrastructure that updates as the audience changes. It is the foundation on which agents can be built to answer commercial questions.
What is an AI agent in this context?
An AI agent is the audience map pointed at a single commercial decision. Instead of a dashboard you have to remember to check, an agent works continuously and reports its findings into the place a brand already makes its decisions, like a dedicated Slack channel. It does not just provide analysis; it recommends a specific action. For example, an agent might report: "The woven leather bag seen on four accounts in the New York map is gaining traction. It is not in your current catalogue. We recommend developing a version with these specific attributes for the Spring collection." The research layer is the component, the agent is the product.
Because the underlying infrastructure is a general model of the audience, we can deploy as many agents as a brand has questions. The same map that powers an agent for finding new product opportunities can also power an agent for identifying the right creator to partner with for a specific launch. The work is done by the agent, not by a human analyst having to reinterpret a report.
What kinds of agents can a fashion brand deploy?
The agents are built to do specific work. They are not general-purpose chatbots. For fashion brands, the first agents they deploy typically fall into four categories, each answering a question that is notoriously difficult to answer with traditional methods.
- Catalogue Opportunities: This agent finds the hidden revenue in a brand's existing catalogue. It cross-references what the audience is trying to buy, or wearing from other brands, against what the brand already sells. Often, the fastest path to growth is marketing a product you already have to an audience that wants it but does not know it exists.
- New Product Recommendations: This agent answers what to make next, at the SKU level. It identifies the specific garments, styles, and attributes that an audience is demonstrating demand for but cannot find.
- Product Feedback: This agent synthesizes what consumers are saying about the products they already own. It goes beyond sentiment scoring to find the specific, recurring issues and points of praise that drive purchasing decisions.
- Creator Matching: This agent finds the single creator who will be most influential for a specific product and a specific audience. It moves beyond follower counts to measure influence on the real audience graph, identifying who actually moves product.
How do you simulate a product launch before it happens?
You simulate it on the real audience graph you have already built. A simulation agent tests a decision, like a new product or a marketing campaign, by propagating it across the measured network of real people. It uses the observed patterns of influence and adoption to predict how far a trend will spread, who will adopt it first, and which communities will be most receptive. This is fundamentally different from asking a large language model to role-play as your customers.
Our research into this area, published at hugo.love/research/audience-simulations-without-llms/, shows why this distinction is critical. A language model is a reasoning engine trained to be rational. Consumers are not. Real buying decisions are driven by habit, mood, status, and envy, with the reason invented after the fact. An LLM, when asked to simulate a consumer, constructs a rational account of a decision that was never rational in the first place.
Why are AI personas a flawed tool for fashion research?
AI personas, or synthetic respondents, are a flawed tool because they are not a real population. Ten thousand agents built on one foundation model is not a sample of ten thousand people; it is one brain wearing ten thousand name tags. They share the same underlying weights, the same biases, and the same training data. Their errors are correlated, which means averaging their responses does not reveal a true signal, it just launders a shared bias into a confident-sounding number.
The literature is beginning to confirm this. A 2026 audit of 37 major language models found they resembled each other more closely than any of them resembled the human populations they were meant to simulate. Worse, the alignment process that makes these models safe and helpful actively deletes the very mechanisms that drive fashion: status judgments, in-group signaling, and class inference. A model trained to refuse to be prejudiced is a poor instrument for measuring prejudice, which is a core driver of luxury and taste.
How can a simulation be tested against reality?
You test it on a world where you already know the answer. The only way to trust a simulation is to run it on historical data and see if it correctly predicts what actually happened. For our simulation methods, we took a real, measured audience of 40,212 people in Germany and mapped the spread of 578 real trends. We then held back the most recent data and tasked our models with predicting who would adopt which trend next.
Our fitted model, which learns the coefficients of influence directly from the graph, achieved an AUC score of 0.86 in predicting the 7,151 held-out adoptions. For comparison, we gave a frontier language model the exact same context and data. It scored 0.78. A plain statistical model, fitted to the real human behavior, beat the most advanced reasoning engine because influence is a mathematical property of a network, not a problem of literary interpretation.
How do you measure real influence, not just followers?
Follower count is a vanity metric, not a measure of influence. An account with millions of followers may have a wide reach, but that does not mean they cause anyone to buy a product or adopt a style. Real influence is measured by observing who adopts a trend after seeing someone else with it. It is a mathematical property of the audience graph, not a simple count of subscribers.
In our work building a real audience graph of 40,212 people in Germany, this distinction becomes clear. When we measured influence based on who actually led to new adoptions, the results were inverted from what follower counts would suggest. One account with only 99 followers was the leading source of influence for 596 separate trend adoptions within the network. Meanwhile, accounts with hundreds of thousands of followers had almost no measurable impact on behavior. This is the kind of finding that only emerges from observing the real network, and it is critical for decisions like creator matching.
How is this different from a traditional research agency?
A traditional research agency delivers a document. They run a project, conduct surveys or focus groups, and return a deck of slides with their findings. This process takes weeks or months, and the output is a static snapshot of a moment in time. It answers the question someone thought to ask, and it arrives after the decision window has often closed. A human team still has to interpret the deck and turn it into a decision.
Our approach delivers an agent that does work. The infrastructure is standing and always on, not commissioned per study. The agent tells you when it finds something, you do not have to ask it. And it reports its findings, with a recommended action, directly into a brand's workflow. It closes the loop between insight and action, which is the step where most of the value from traditional research leaks out.
What are the best questions to ask an AI agent about fashion?
The best questions are commercial decisions. Instead of asking "what are people saying about denim?", you deploy an agent to answer "which styles of jeans should we stock next season for our target audience in London?". Instead of "how is our brand sentiment?", you ask "which product in our current catalogue has the biggest gap between audience demand and our marketing effort?".
The goal is to move from analysis to action. The most valuable questions an agent can answer are those that lead directly to a decision that makes the brand money or saves it from a costly mistake. For example: "Which three micro-influencers have the most actual influence on outerwear adoption among 25-34 year olds in Paris?" This is a specific, actionable question that a general research tool cannot answer, but an agent built on a real audience map can.
What are the limitations of this approach?
This method has clear and important limitations. First, it only works where people are public about their lives and the products they use. It cannot map private communities or B2B audiences who do not post. Second, the data is directional, not a census. It tells you what is gaining momentum and with whom, but it cannot tell you "how many" in a way that would satisfy a regulator. For that, you need a representative panel from a provider like NIQ or Kantar.
Finally, it cannot test a true counterfactual for something that has never existed. We can simulate a product launch based on its attributes and how similar attributes have propagated before, but we cannot test a completely novel concept with no historical precedent in the data. For pure concept testing of unlaunched products, elicited research methods like AI-moderated interviews remain a better fit. We have not solved those problems, and that is the honest state of this work.
How do I start mapping my brand's audience?
The process begins with defining the audience you want to understand. This could be the customers of a key competitor, the population of a target city, or the followers of a particular aesthetic. From that starting point, we build the first layer of the map by identifying the core members and expanding outwards through chain referral.
Once the map is built and the garment-level data is resolved, we can deploy the first agents to tackle the most pressing commercial questions. The infrastructure is ours, but the map and the agents are specific to your brand and your audience. The goal is to replace guesswork with a standing capability that informs decisions every day, not just once a quarter.