EvidenceField

The problem

Remote panels, Zoom automation, web-form recruiting, and transcription pipelines all assume digitally fluent, English-adjacent participants. In Vietnam, Indonesia, Brazil, Mexico, and most of the markets where the next billion users live, participants often won’t — or can’t — do digital onboarding or remote studies. Global product teams either fly researchers in at enormous cost, or skip the research and guess.

The false fix

The AI industry’s answer is synthetic respondents: LLM personas that stand in for real participants. It demos beautifully — and the peer-reviewed evidence (Section 03) shows it is least valid exactly where research is hardest. GPT’s similarity to real humans falls with a country’s cultural distance from the United States at r = −.70. Pure-synthetic vendors accelerate research where it was already easy, and fail where the need is greatest.

Why now

• AI is finally good enough at the right jobs — transcription, translation, and cross-market synthesis of messy multilingual qualitative data — to collapse the cost of processing human research, without pretending to replace the humans.
• The failure modes of pure-synthetic research are now documented in top journals — buyers are about to get burned, and the honest alternative doesn’t exist yet.
• Product organizations are more global than their research stacks. The gap grows every quarter.

Founder edge

Eleven-plus years operating a $2B+ earner marketplace across 130+ countries: firsthand operating knowledge of why Western research and onboarding playbooks fail in Southeast Asia and Latin America, an operator’s network in those markets, and the product + operations discipline that the expert routing layer productizes.

Pure-synthetic vendors

Fantasy’s Synthetic Humans (the category’s flagship — Fast Company World Changing Ideas 2024; BP, LG, Ford, Spotify, Google engagements), Synthetic Users, Evidenza, Yabble. Common shape: a persona prompt layer on third-party foundation models. Thin moat — every model upgrade erodes last quarter’s tuning — and structurally weakest in non-WEIRD markets. Their documented failure modes are our “why us.”

Research incumbents

Kantar, NIQ, and Qualtrics have all publicly landed on “supplement, not replace” — NIQ: “synthetic respondents are not a replacement for human consumers… a supplement to your ideation process”; Qualtrics:“synthetic data augments human research, it does not replace it… models degrade without ongoing human data collection.” They validate the hybrid thesis but carry panel-era cost structures and lack emerging-market, low-tech field operations DNA.

The gap nobody covers

Emerging-market fieldwork operations + AI synthesis + expert routing, sold as one system. Pure-synthetic vendors structurally can’t serve non-WEIRD markets; incumbents have panels but not in-person operational muscle in SEA/LatAm; nobody productizes the routing decision itself.

The moat

The human fieldwork network and the routing expertise. Model upgrades erode the synthetic layer — which we deliberately treat as a commodity input — and strengthen our synthesis layer. The proprietary asset compounds: every study adds verified human data from markets where nobody else can collect it.

Where synthetic genuinely works

Cheap, fast exploratory work: ideation, questionnaire debugging, directional concept ranking (~5,400 responses for under $1 in an hour; effect-direction correlation with humans r ≈ 0.85). Always a precursor to human data, never a substitute. Better elicitation methods (e.g. PyMC Labs × Colgate-Palmolive’s semantic similarity rating, ~90% ranking attainment) keep improving this layer — which is why we treat it as a commodity input, not a moat.

Why humans stay in the loop — precisely

Human data is the binding constraint: even the best statistical hybrid — calibrating 100,000 LLM responses against 10,000 human ones — adds only ~13% effective sample size (Broska et al. 2025, Sociological Methods & Research). And the MIT meta-analysis of 106 human-AI experiments (Nature Human Behaviour 2024)shows combinations win when the human is the stronger performer and lose when the AI is. So the design rule is not “always add a human” — it is put the human where the human is superior: research design, field access, cultural interpretation, judgment. That rule is the product.

Risks & honest answers

• Frontier models close the validity gaps.Possible on variance and elicitation; the WEIRD data gap requires data that doesn’t exist online — which our field network generates. We commoditize the layer that improves and own the layer that doesn’t.
• Ops-heavy business. Yes — deliberately. Start narrow (2 markets, 1 vertical), let AI absorb the processing cost, and price as research value, not hours.
• Buyer skepticism of anything “synthetic.” Our positioning is the skepticism, with citations. We sell the honest version of the category.

The study lifecycle

1. Intake — decision to be made, markets, segments, timeline.
2. Market assessment — data-richness scoring per market: language coverage, digital penetration of the segment, prior research corpus.
3. Routing — expert assigns the mode mix per market and designs the instruments; client sees the routing rationale.
4. Synthetic pretest — where valid: debug instruments, rank concepts, sharpen hypotheses before spending on fieldwork.
5. Human fieldwork— the field network collects real responses, met in participants’ channel and language of comfort.
6. AI synthesis — transcription, translation, theme extraction, cross-market comparison; humans spot-check every load-bearing quote.
7. Decision-ready readout — recommendations with a full citation trail; every claim clickable back to its evidence.

What the client buys

Which product to build, which market to enter, which pricing survives contact with reality — at the speed of AI where AI is valid, with the confidence of real humans where it isn’t.