AI agents for consumer brands
We predict and simulate consumer behaviour. Then we build the agents that act on it.
Hugo holds a measured, maintained model of real consumers: who they are, what they think, what they need, and what they will buy next. We use it to predict where an audience is going and to simulate how it reacts to a product, a campaign or a price. Then our engineers work inside the brand, building and deploying custom agents for what that brand needs, so the predictions turn into work the business can use the same day. We invent the technology that makes better simulations possible. The agents are how it earns its keep.
Every number on this page is published with its method and its error rate.
The edge
Four layers per audience, built by observing real people.
This is the company. Everything else we ship runs on top of it. The map holds a named audience across four layers, built from live public conversation, images and video, and it is maintained continuously rather than commissioned once. Four decisions separate it from a scrape.
Reaching real people, not whoever the algorithm surfaced
Hashtag and keyword search produces a sample whose selection probability is set by an unobservable ranking algorithm, which cannot be corrected for afterwards. Hugo uses chain referral sampling, known in survey science as respondent-driven sampling, which reaches populations no register lists. The New York map reached 876 real accounts across 4,464 posts that way.
Seeing what people wear, not what the caption said
A fine-tuned garment detector runs on every image and video in the stream, at F1 0.597, with shoe recall lifted from 0.21 to 0.80. Only 2 percent of the frames in that run were outfit posts. The other 98 percent of observable clothing appeared incidentally, in posts about something else, which is precisely the behaviour keyword listening cannot see.
Embedding the garment rather than the photograph
Naive embeddings cluster by lighting and pose instead of by clothing. Training on the garment cut photo bias from 0.672 to 0.428 and reached pair accuracy of 0.959, so two people wearing the same jacket resolve to the same garment rather than to the same aesthetic.
Resolving to an actual product
Corroborated visual search takes a garment to a real, buyable SKU at recall@1 of 0.87 to 0.90 automated, and 0.76 when hand-audited. That is the step that turns a cultural observation into something a merchandiser can act on.
Why the claim holds
We publish our error bars. Ask anyone else in this category for theirs.
Saying we understand consumers better than anyone is worth nothing as an adjective, so we do not leave it as one. Every number above sits on this site with its method attached. Simulated audiences are run blind against published real-world surveys they have never seen: 71 percent winner accuracy and a mean Jensen-Shannon divergence of 0.039, across 70 held-out questions, 19 audiences, 11 countries and 5 languages. It generalises across language, at 0.028 divergence in Finland.
And the limits, in the same breath, because a measurement without them is a marketing claim. It only works where people talk in public. It over-represents the delighted and the annoyed, so frequency is directional rather than incidence. It cannot test a concept nobody has launched or discussed. The degree-weighted bias correction is designed and not yet implemented. A near-duplicate rule removes about 18 percent of crops, unevenly by category. Magnitudes flatten, so trust the ranking before the gap.
That is the whole basis of the claim, and it is checkable, which is the only kind worth making. Read the accuracy report, or the mapping programme it runs on.
Why it decides everything
Agents are easy now. That is exactly why the map is the product.
This is not a claim that other agent products are badly engineered. Most are fine, and the underlying models genuinely can do the reasoning these tasks need. Wrapping one in a brief and a schedule is a weekend of work, which is why there are suddenly so many.
Which is the point. When the reasoning is commodity, the only thing separating two agents is what they know. Ask an agent what a brand should make next season and it has to know what that brand's specific customers wanted and could not find. A general model has never met them. It knows the category's average opinion, which every competitor's agent also knows, and that is why those answers come back plausible, confident and interchangeable.
The constraint was never reasoning. It is observation. We built the observation first, and the agents are what became possible afterwards.
From prediction to action
A prediction changes nothing until something acts on it
Most consumer intelligence ends as a document. Somebody reads it, agrees with it, and the launch goes out the way it was always going to, because the finding and the decision lived in different rooms. We built the company so that a simulation ends as work done.
Every agent runs the same loop, and it runs whether or not anyone remembered to ask.
A real audience, measured continuously
Real people reached through their actual connections, what they wear read out of ordinary video, what they ask for read in their own words. Nothing here is a persona.
Where the audience is going, and how it would react
Which trends are spreading and who adopts next, fitted on the follow graph between real people and scored against a future the model never saw. A product, a campaign or a price can be put into that audience to see where it travels and where it stops.
One agent, one commercial decision
An agent is the prediction layer pointed at a single decision for a single brand. It weighs what the audience is about to want against what you sell, what you have in stock and what you have planned.
The work arrives where the decision is made
The agent does not send a finding. It sends the thing the team would have had to produce from the finding: the brief, the push list, the name. It lands in the Slack or Teams thread where that decision is already being argued, with the real posts behind it.
Then it keeps watching. How your audience reacts to what you launched is public too, so it shows up in the same model the next prediction is made from.
What acting looks like, agent by agent
| Agent | What it predicts | What it does about it | What the business gets |
|---|---|---|---|
| Catalogue | Which product you already stock is about to match a movement in your audience | Sends this week's push: the SKU, the audience, the reason and the timing | Revenue from stock you already hold |
| Product | Which unmet demand in your audience is growing | Writes the product brief: the item, who is asking for it, and what they said | A range decision made before the sales data could have shown it |
| Creator | Which account carries one product into the community that buys it | Names the creator, the product and the audience, scored by position in the graph | Seeding spend that reaches buyers |
| Competitor | Which reasons for choosing someone else are gaining ground | Sends the reasons, quoted, with your product set against each | A reply to the competitor while it still matters |
| Quality | Which fault will keep costing you customers | Sends a ranked fix list, with the complaints that never reached your support inbox | The same mistake not made twice |
| Simulation | How your real audience reacts to a concept, a price or a campaign | Runs it before launch and returns who adopts, where it stops, and the error rate | Money not spent on the version that would have failed |
| Custom | Whatever the decision needs | Built by our engineers inside your team, around your catalogue and calendar | An agent for the problem only you have |
These are where most brands start. They are examples, and the list is not a menu.
How we work with a brand
We build and deploy custom agents for what your brand needs
We run a forward deployed model. Our engineers work inside your team, learn the decisions that cost you the most when they go wrong, and build agents for those decisions on our simulation infrastructure. You do not buy a seat in a tool and work out what to do with it.
New technology for better simulations
Most of the company works on the prediction layer: reaching real audiences, reading what they wear and want, and modelling how influence moves between them. Every improvement there reaches every agent already deployed.
Engineers inside the brand
A forward deployed engineer sits with your product, brand and marketing teams and finds the decisions being made on instinct today: the range review, the weekly push, the seeding list, the reorder.
An agent for each decision
Each agent is built around your catalogue, your calendar and your audience, and wired into the places your team already works. How much it does on its own, and what needs a person to approve, is set with you for each decision.
It keeps working, and the next one is faster
Agents run continuously. When a new problem comes up, the infrastructure underneath is already in place, so the next agent is a matter of pointing it at a new decision.
A decision is specific to a catalogue, a calendar and a customer. An agent that fits every brand is an agent that knows none of them, which is why we build them one brand at a time.
The agents
The agents brands ask for first. Each one owns a decision.
Every agent we deploy is custom, and most deployments begin with one of these. Every agent below is the same shape: a decision a brand makes on a cycle, and the layer of the map it reads to make it well. They work continuously and report into the channel your team already uses.
The catalogue agent
What to push this week. It predicts which product already in your catalogue is about to match a movement in your audience. The fastest revenue in most catalogues is something the brand already stocks and nobody is putting in front of the people asking for it.
“What should we push this week, and why now?”
The product agent
What to make next. It reads the demand language in your audience: requests, workarounds, compromise purchases and abandonment, grouped by the job the product does. None of that is in your sales data, which can only contain products you already sell.
“What are our customers asking for that we do not make?”
The creator agent
The one creator who would move one product. It reads the graph, and follower count barely enters into it. Influence travels through trends and communities, so the question is which creator sits inside the community that buys this product.
“Who should seed the new runner to Stockholm women aged 22 to 30?”
The competitor agent
Why a competitor gets picked over you. People say out loud why they bought elsewhere. It collects those reasons verbatim, ranks them by how far each one moves a decision, and points at the product of yours that answers it.
“Why are we losing the under 30s to them?”
The quality agent
What to fix in what you already sell. Reviews you own are written by people who stayed. Public conversation includes the ones who returned it and the ones who never came back. It finds the fault that keeps recurring and what it is costing you.
“Why did the autumn knitwear underperform?”
The simulation agent
Test it on the audience before you spend. Put a concept, a price or a campaign into a model of your real audience while it is still cheap to change your mind, and see who picks it up and where it stops. We run our simulations blind against outcomes the model never saw, and publish how close they land.
“Will this travel beyond the people who already follow us?”
The custom agent
The decision your team keeps having to make. We own every layer beneath the answer, so the list is not fixed. Tell us the decision and our engineers build the agent and the workflow around it, against your category and inside the tools your team already uses. That takes weeks.
“Build us the one nobody sells.”
Next
What the same infrastructure supports, and we have not built for anyone yet.
Stated plainly rather than implied, because a roadmap presented as a product is the fastest way to lose a buyer.
The pricing agent. Where a price sits against what an audience believes the product is worth, and what they compare it to when deciding. The map already holds the comparison set and the language people use about value, which is the input a pricing decision needs and rarely has.
The market expansion agent. Which market or audience holds unmet demand for what you already make. The same four layers, run against a population you do not sell to yet.
The test
Bring a decision you already made and know the outcome of.
The honest way to evaluate this is not a demo of our choosing. Pick a launch, a range call or a campaign from the last year where you now know how it went, and let the agent answer it blind against the audience as it stood at the time. If it tells you what you already learned the expensive way, that is the whole argument. If it does not, you have lost an afternoon and learned something real about the category.
Frequently asked questions
Does Hugo build custom agents for each brand?
Yes. Hugo runs a forward deployed model: its engineers work inside a consumer brand's team, identify the decisions that matter most, and build and deploy custom agents for those decisions on Hugo's consumer simulation and prediction infrastructure. Each agent is built around that brand's catalogue, calendar and audience and wired into the places the team already works. How much an agent does on its own, and what a person approves, is set with the brand for each decision. Most deployments start with a catalogue, product, creator, competitor, quality or simulation agent, and extend from there.
How do Hugo's agents act on a consumer simulation?
Each agent runs a loop: observe a real audience continuously, predict where it is going and simulate how it reacts to a product, a campaign or a price, weigh that against one specific commercial decision for one brand, and then produce the work the decision needs. The product agent writes the product brief, the quality agent sends a ranked fix list, the catalogue agent sends the week's push with the SKU, audience and timing, the creator agent names the creator for a product, and the simulation agent runs a concept, price or campaign on the audience before launch. The work arrives in the brand's own Slack or Teams thread with the real posts behind it. Agents are custom built for each brand, and how much one does on its own is set with the brand for each decision.
What is the difference between consumer intelligence and an agent that acts on it?
Consumer intelligence ends as a finding somebody has to read, interpret and turn into a decision, which is where most of it stops. An agent is the prediction layer pointed at one decision, so its output is the thing the team would have had to produce from the finding: a brief, a push list, a name. It runs continuously instead of when someone commissions it, and the audience's public reaction to the decision shows up in the next prediction.
Why is the consumer map the thing that matters, not the agent?
Because every decision a brand wants automated is bounded by knowledge, not by reasoning. Deciding what to make next requires knowing what a specific audience asked for and could not find. No amount of model quality substitutes for that observation. The agents are the hands. The map is the part that is hard to build, hard to copy, and the reason the answers are specific to your customers rather than to your category.
What makes Hugo's understanding of consumers better than anyone else's?
That it is measured and published rather than asserted. Hugo reaches real people by chain referral sampling instead of hashtag search, identifies the garments they actually wear rather than the words in their captions, resolves those garments to buyable products, and reports the error rate of every one of those steps. Ask anyone else in this category for their recall, their divergence against held-out real-world surveys, or their sampling method, and see what comes back.
How is this different from wrapping ChatGPT or Claude in a prompt?
A general model reasons over what it absorbed in training, which is the internet's average opinion about a category. It has never observed your customers. Hugo observes a named audience, learns a behavioural model per person, and connects people, products and trends on a graph. The agent reasons over that. One interprets, the other retrieves and evidences.
What does Hugo actually observe, and how accurate is it?
Public consumer conversation, images and video across social platforms, forums, reviews and the open web. Garment detection runs at F1 0.597 with shoe recall lifted from 0.21 to 0.80. Product resolution reaches recall@1 of 0.87 to 0.90 automated and 0.76 hand-audited. Simulated audiences answer held-out real-world survey questions with 71 percent winner accuracy and a mean Jensen-Shannon divergence of 0.039, across 70 questions, 19 audiences, 11 countries and 5 languages.
What are the limits?
It only works where people talk in public. It skews toward the delighted and the annoyed, so frequency is directional rather than incidence. It cannot test a concept nobody has launched or discussed. The degree-weighted sampling-bias correction is designed but not yet implemented, and a near-duplicate rule removes about 18 percent of crops unevenly by category. Magnitudes flatten, so trust the ranking before the gap.
Which agents does Hugo build?
Hugo builds and deploys custom agents for each brand, so the list is not fixed. The ones consumer brands ask for first are the catalogue agent (what to push from what you already stock), the product agent (what to make next), the creator agent (who carries a product into the community that buys it), the competitor agent (why someone else gets chosen), the quality agent (what to fix), and the simulation agent (test a concept, price or campaign on a model of your real audience before spending). A pricing agent and a market expansion agent are supported by the same infrastructure and have not been built for a customer yet.
Do the agents replace a research team?
No. They remove the part of the work that is retrieval and waiting, which is most of the elapsed time in a research cycle. Judgement about what to do with the answer stays with the team. Heads of insights use Hugo to answer more questions per quarter, not to answer fewer questions with fewer people.
Where do the agents run?
In the brand's own Slack or Microsoft Teams, so people ask the way they already ask each other. The web app holds historical runs, longer reports and exports. Most answers arrive without anyone leaving the thread.
Related
Keep reading
Mapping what consumers actually wear · Audience simulations built on math, not LLMs · How fashion brands use AI · Hugo for fashion · AI for building a consumer brand · Hugo vs Aaru · Hugo vs Simile