Synthetic customer modelling
An AI model prompted to reason and respond as a particular buyer. It powers every offer ideation technique with breadth and speed; the real customer still supplies the truth and the surprise.
A synthetic customer is a large language model prompted to reason and respond as a particular kind of buyer. The team describes the customer, their role, their pressures, their industry and the decision in front of them, and then talks to it: interviews it, puts a draft offer in front of it, asks what it would object to. The book calls this the most important idea in chapter 6, because it turns slow, expensive customer projection into something a team can run in an afternoon and repeat whenever the market moves.
Why take it seriously?
Because there is now research behind it. In 2024 a Stanford team interviewed a thousand real people and built an AI agent from each interview. Given the same surveys and personality tests, the agents reproduced their human counterparts’ answers with striking accuracy, and far better than models built on demographics alone (Park et al., 2024). Market research firms have moved quickly; NielsenIQ writes openly about synthetic respondents as a way to test concepts in hours rather than weeks (NielsenIQ, 2024).
How does it work in practice?
It starts with hypothesis projections: describing the people who will decide and projecting their response before investing. The book’s case is a global lubricants company testing large packs of engine oil in out-of-town stores in a new country. Its consumers were six personas plotted by willingness to pay against investment in car care, from the Minimalist to the Passionate. In a business market the personas are the seven roles of the decision-making unit. Before AI this took weeks and a research budget. A synthetic customer lets a team walk all six shoppers down the aisle, then do it again for another country, or put an offer to all seven roles and hear where each resists.
And it powers the rest of offer ideation too. The book places synthetic customer modelling at the hub, with the seven field-proven techniques around it. The figure shows what it does for each.
The engine at the hub
Synthetic customer modelling
A large language model prompted to reason and respond as a particular kind of buyer. It brings every technique to life at speed and scale; the real customer supplies the truth and the surprise.
What are its limits?
A synthetic customer simulates from patterns already in its training data. It reflects what is broadly known about a customer type well, and surfaces the thing nobody has said out loud badly. The Kano model shows why that matters: delighters are by definition seldom expressed and new to the world, so a machine trained on what has been said tends to return the basics, not the surprise. The answer the book gives is to feed it properly. A synthetic customer grounded in real voice-of-customer evidence, kept current, is a very different thing from one spun out of generic training data.
What it is not
It is not the customer. Treating its answers as findings rather than hypotheses is automation bias with a persona attached. The book’s angle is the right one: use the synthetic customer for breadth, the real one for truth. And the closing point of the chapter holds. The model cannot shake a hand, read a room or win the trust of a senior leader deciding whether to bet part of their career on a supplier. As the modelling gets better, that human connection matters more, not less.
Set out in full in: Creating Customer Value Propositions with AI (Kogan Page, 2027), chapter 6.
Related terms
Offer ideation
Diverge before you converge. The machine's gift is width; the cull is yours.
Decision-making unit (DMU)
Everyone on the customer side who decides, influences, uses, pays for or can veto a purchase, mapped as seven roles and their relationships rather than job titles.
The Kano model
Basic, performance and excitement needs, and the uncomfortable fact that none of them holds still: excitement decays into performance, and performance into basic.
Hallucination
A confident, well-formed answer that is simply false. Worst on multi-step reasoning and specialised domains, which describes key account work exactly.
Automation bias
The human tendency to trust the machine even when it is wrong. The danger is not that AI replaces your expertise but that it lulls it into disuse.
Human in the loop
Adopt, but keep your hands on the wheel. The machine amplifies judgement. It does not supply it.
See the method run on one real account in three days: the One Page Proposition.