Why not just prompt your own LLM?

It's the question every executive is being asked right now.
We're going to give you the honest answer.

The honest answer

If your question is “what will affluent customers in their forties do when interest rates rise?”, you can prompt your own LLM with a segment profile and get an answer.

The answer will be confident. It will be plausible. It will be cheap.

It will also be a hallucination dressed up as evidence, and the next person who reviews the decision it shaped will not be able to tell you why it said what it said.

Three approaches, one of them works

Digital twins, synthetic personas, and synthetic twins are not three names for the same thing. The category looks crowded. It isn’t.

Synthetic persona
Synthetic twin (LLM impersonation)
Digital twin
Built from
An archetype generated from real survey data
One prompt call, no underlying data
A behavioural replica of one specific real person, built from 5+ hours of AI-moderated video interviews, continuously enriched
Unit of analysis
An average
A fully fabricated character generated by the model, with no real-person grounding
One real individual, traceable back to their interview
Variance
Collapsed by design (the amalgamation step averages out the most interesting parts of the audience)
Inherits all model biases unfiltered
N twins each answer in their own voice, you get a real distribution
Validation against ground truth
No underlying source to check
None
Yes, via the real person sitting behind the twin
Auditability
Limited
None
Sense-checkable down to the interview moment

Built from

Synthetic persona An archetype generated from real survey data
Synthetic twin (LLM impersonation) One prompt call, no underlying data
Digital twin A behavioural replica of one specific real person, built from 5+ hours of AI-moderated video interviews, continuously enriched

Unit of analysis

Synthetic persona An average
Synthetic twin (LLM impersonation) A fully fabricated character generated by the model, with no real-person grounding
Digital twin One real individual, traceable back to their interview

Variance

Synthetic persona Collapsed by design (the amalgamation step averages out the most interesting parts of the audience)
Synthetic twin (LLM impersonation) Inherits all model biases unfiltered
Digital twin N twins each answer in their own voice, you get a real distribution

Validation against ground truth

Synthetic persona No underlying source to check
Synthetic twin (LLM impersonation) None
Digital twin Yes, via the real person sitting behind the twin

Auditability

Synthetic persona Limited
Synthetic twin (LLM impersonation) None
Digital twin Sense-checkable down to the interview moment
When the answer at the tails matters, early adopters, edge cases, expert outliers, only twins surface them. Personas and synthetic twins are consumed once and discarded. Digital twins are an asset that appreciates.

86% vs 57% on the same task.

We ran a head-to-head: predict bank choice for a defined consumer segment, validated against real respondent panellists.

Bank choice prediction · Brox vs LLM vs generic

Validated against real respondent panellists

Brox digital twins
Standalone LLM (segment-prompted)
Generic AI baseline (no segment context)

Source: Brox internal benchmark, validated against real respondent panellists.

Across other validated tasks (price sensitivity, concept testing, message resonance, switching intent, ad creative testing) the gap held. Real-human-twin panels beat LLM impersonation on every single one.

Why the gap exists

Three reasons.

LLMs are trained on what people wrote, not how they behave.

The world’s text is biased toward what people say in public. Most decisions are not made in public. Survey responses, interview transcripts, behavioural prompts and decision-trees, the raw material of how a real human actually decides, is not in the training set of any frontier model.

LLM-impersonation collapses the variance.

When you ask an LLM to “be” a 47-year-old nurse in Atlanta, you get the average of every 47-year-old nurse the model has ever read about. Real humans are weirder, sharper, more contradictory, more accurate. You need the variance, not the mean.

There is no audit trail.

An LLM cannot tell you why it said what it said, because no real person was ever consulted. With Brox, when a twin says something unexpected, you drop into the underlying interview transcript and see why.

An LLM cannot tell you why it said what it said.
A digital twin always can.

When an LLM is the right tool

We’re not anti-LLM. We use Claude every day. There are tasks where a prompted LLM is the best tool: drafting marketing copy, summarising a meeting, debugging code, brainstorming a campaign, generating creative variants for testing.

But for any decision where you need to predict what real humans will do, with traceability and an audit trail, drug launch, credit policy, ad spend, pricing, M&A diligence, crisis simulation, the LLM is the wrong instrument.

Find out more

We could write pages and pages of marketing and sales fluff but instead we'd prefer just to show you the product.

Contact us and we'll set up a call, during that call you should come prepared with some business and research requests and we'll go through them together live. We'll also explain how we build our digital twins, why you can trust them, how to deploy them, how much it costs (it's a lot), how we validate them and anything else you fancy talking about. You'll see ROI immediately.

Currently live in the US, UK, Japan and Turkey and launching in much of the Middle East and APAC pretty soon.