The Modelled Voice: A Boardroom Audit for Synthetic Respondent Evidence

Synthetic respondents have moved from conference speculation to commercial offer. In 2026, UK agencies and platforms are presenting modelled personas, simulated samples and AI-generated interview transcripts as a quicker, cheaper alternative to fieldwork. For decision-makers, the attraction is obvious: fewer weeks in the field, lower costs and answers on demand. The risk is equally clear. A synthetic respondent is not a person. It is a model output, shaped by training data, prompts, weighting assumptions and the commercial priorities of the supplier. Before such material enters a board paper, it needs an audit.

What synthetic evidence is, and what it is not

A synthetic respondent is a generated approximation of a respondent type. It may be built from survey microdata, census profiles, transaction records, social media content or older qualitative transcripts. The output can look like an interview, a completed questionnaire or a persona summary. It cannot offer testimony in the ordinary sense. There is no respondent to recontact, no refusal to interpret, no spontaneous remark and no accountability for a misleading answer.

That does not make synthetic evidence worthless. It can be useful for hypothesis generation, questionnaire testing, internal scenario planning and early exploration of language. It is much weaker as a substitute for measurement. Claims about incidence, market size, willingness to pay, regulatory behaviour or the views of vulnerable groups require observed evidence and a defensible sampling approach. A modelled voice can suggest where to look. It should not be the only reason a board decides what it is looking at.

Begin with the supplier's evidence chain

Procurement should start before the deliverable arrives. Ask the supplier to explain the evidence chain in plain English: what data trained the model, when that data was collected, which markets and populations it covers, how the sample was weighted and what human review took place. A supplier that cannot describe the chain is not offering research. It is offering an output.

The questions should be specific. Which UK datasets were used, and do they reflect the category, channel and customer type in question? How were under-represented groups handled? What model version produced the findings, and can the same inputs be reproduced? Were prompts or persona instructions supplied by the client, and could they have steered the answer? What validation was run against known benchmarks, observed survey results or independent data? What are the known limitations, and how are they disclosed in the deliverable?

These are not hostile questions. They are the same questions a careful buyer would ask of any new supplier. Synthetic methods simply make them more urgent because the output can be fluent enough to disguise thin foundations.

Plausibility is not validity

The boardroom danger is that synthetic evidence often reads well. A modelled quote can sound like a customer. A simulated segment can look tidy. Fluency, however, is not a proxy for validity. A synthetic panel may reproduce the assumptions in its training data, including old category conventions, biased language and gaps in coverage. If the underlying data under-represents small firms, regional buyers or complex buying committees, the model may smooth over exactly the uncertainty the board needs to see.

A practical audit therefore separates observed evidence from modelled evidence. The board paper should state which findings come from real respondents, which are modelled, and which are inferred by the author. A short evidence register can do this without turning the paper into a methods appendix. The register should list the source, date, population, method, sample basis, limitations and confidence rating for each material claim.

A traffic-light framework for board papers

The following framework is deliberately simple. It is intended to help decision-makers classify synthetic evidence before it is presented as decision-grade.

| Rating | Suitable uses | Conditions | | Green | Internal ideation, question wording, hypothesis generation, early framing | Labelled as synthetic; not used for external claims; no material decision rests on it alone | Amber | Directional input for discussion, early concept exploration, supplementing observed evidence | Triangulated with at least one independent source; limitations disclosed; human review documented | Red | Board approval, market sizing, pricing, regulated claims, vulnerable audiences, reputationally sensitive decisions | Not acceptable as sole evidence; requires observed fieldwork or equivalent independent verification |

Green does not mean good. It means low consequence. Amber is where most synthetic evidence belongs: useful for shaping a decision, not for settling one. Red is the category that matters most. If a supplier proposes synthetic respondents as the primary evidence for a material board decision, the buyer should treat the proposal as a warning, not a shortcut.

Procurement clauses for 2026

Contracts should catch up with the pitch. Buyers should require clear disclosure of any synthetic component in a deliverable, a written statement of data sources and model version, and a commitment not to use client data for model training without explicit permission. Where findings may inform regulated or high-stakes activity, the supplier should provide a validation record and a named human reviewer.

It is also reasonable to ask for a right to audit the evidence chain and to require that synthetic outputs are never presented as observed respondent quotes. These clauses protect the buyer and the supplier. They make the difference between a research partnership and an opaque software transaction.

The board should see the seams

Synthetic evidence is not automatically bad. It is an additional instrument, and like any instrument it has a range and a margin of error. The problem arises when the seams are hidden. A board paper should make the modelled nature of the evidence visible, state what it can and cannot support, and show how it has been triangulated. If the supplier cannot explain the training data, the weighting and the validation, the correct rating is red. A confident voice is not the same as a reliable one, and a board should never be asked to confuse the two.