The lie at the centre of every decision
Every brand decision, every drug launch, every credit product, every fan strategy, every government policy, is built on assumptions about what humans will do. And almost every one of those assumptions is a guess.
For the bigger ones, people commission surveys, focus groups, and consultancy studies, and that helps. But it’s expensive, it’s slow, and it’s still not reality. We’ve gone from blind to half an eye open with bad cataracts, and we’ve called it progress.
The guess might be wrapped in a beautiful PowerPoint, modelled in a sophisticated spreadsheet, or backed by a five-figure focus group. It’s still a guess. The world has accepted this for decades. I started Brox because I don’t think we have to anymore.
And no, AI on its own can’t fix this. It can just pretend to.
Why now
Three things are happening at once.
First, AI can now hold a deeper conversation with a person than a human interviewer can, at massive scale. Hours of interviews with the same individual over time, asked the right way, tells you more about how someone actually thinks than thirty minutes with a moderator and a clipboard ever could. This was always theoretically possible. It was just exorbitantly expensive, slow, and unsystematic. Now it isn’t.
Second, every Fortune 500 is in a quiet panic about how to use that AI. Their boards keep asking the same question: “why don’t we just prompt our own LLM with a segment profile?” Because LLMs are trained on what humans have written, not how they behave. They’re remarkable, but without grounded data, calibration, and the right agentic scaffolding, they cannot predict what real people will actually do.
Third, regulated industries cannot afford unauditable inputs. The CRO at a pharma company will not let an unexplained model decide how to launch a drug (or at least they shouldn’t, though I’m sure some do). The Chief Compliance Officer at a bank will not let a hallucinated persona decide who gets a credit card. AI-assisted decision making demands provenance, and the dirty little secret no one wants to say out loud is why: every model makes mistakes eventually, ours included (very rarely, of course), and when they do you need to be able to trace why.
What we believe
1. Real humans are the data.
Not synthetic profiles. Not personas. Not made-up segments. Real people we’ve spent hours with, building behavioural twins deep enough to predict what they’ll do in scenarios that haven’t happened yet.
The distinction matters. An LLM can interpolate within its training data, giving you an answer that looks like the millions of answers it has already seen. But humans are asked, every day, to make decisions outside that distribution: a new drug, a new product, a price change, a policy shift. To predict behaviour in a situation that hasn’t happened yet, you need a model of why a person behaves the way they do, not just a record of what they’ve done before. That’s what we build. A mechanistic model of a real human, grounded in hours of conversation with them.
2. Provenance is the moat.
If you can’t trace an answer back to a real human moment, to a mechanistic behaviour, you don’t have an answer. You have an opinion expensive enough to look like one.
3. Vertical depth wins.
We chose pharma and financial services first because that’s where decisions about humans carry the highest stakes, the strictest compliance bar, and the deepest professional respect for evidence. (We also do sports and TV because those clients keep coming to us and we’re good at it. First world problems.)
4. Transparency, with limits.
We publish our benchmarks. We publish our predictions. We share our methodology with clients. The only way to win the trust argument is to show the work, continuously, including the failures and the limitations of what we’ve built. We won’t publish trade secrets, but everything that affects whether you can trust an answer, you get to see.
5. AI plus real people beats traditional surveys.
Both on accuracy of prediction, and on explaining why people behave the way they do, not just what they’ll do next. We know that’s a big claim. We’re happy to defend it with anyone who fancies a philosophical chat.
What we refuse to do
We will never publish an accuracy number that hasn’t been measured against a real-world outcome.
We will never compete on price with vendors selling LLM-prompted personas dressed up as data. Ours is the human-truth layer for the most consequential decisions in the world, and we tell prospective customers to go elsewhere when we’re not the right fit, which our sales team has mixed feelings about.
What this means for our customers
If you work at a pharma company, a bank, an insurer, a consulting firm, a sports league or team, or a government, our job is to give you a tool you can use the way you use a spreadsheet. Every day. On every decision. With the receipts in hand.
Not a quarterly study. Not a six-figure project. A live feed of human truth that any person or system in your organisation can call on demand before they act.
What this means for the AI industry
AI without human truth is a hallucination factory. We exist so that your team, and the AI agents your company is racing to deploy, have a reliable source of human signal to ground themselves against before every decision they make.
We are 15 humans across San Francisco, London, and Amsterdam, 30,000 real people who have trusted us with hours of their lives, and a growing list of the most serious companies in pharma, finance, sports, and consulting who have decided this is the way they want to make decisions.
The next decade of AI will be defined by what its decisions are grounded in. We’ve made our bet. If yours is the same, we’d love to hear from you.