# Can you build a synthetic B2B customer?

Date: 2026-11-21
Author: Richard Brooks
Canonical: https://richard-brooks.com/synthetic-b2b-customers/

<link rel="stylesheet" href="/essays/eth/figures.css">
<link rel="stylesheet" href="/essays/synth/figures.css">

Consumer goods companies [test a new product idea](https://en.wikipedia.org/wiki/Concept_testing) by showing it to a few hundred people and asking how likely they are to buy it. In October 2025 a team from [PyMC Labs](https://www.pymc-labs.com/) and [Colgate-Palmolive](https://www.colgatepalmolive.com/) published [a paper](https://arxiv.org/abs/2510.08338) arguing that an AI could do a good part of the answering.

They built "synthetic consumers" from a [language model](https://en.wikipedia.org/wiki/Large_language_model) and compared them with 9,300 real people across 57 surveys. The match was close enough for one trade site to report it as ["90% accuracy"](https://ppc.land/llms-achieve-90-accuracy-in-replicating-consumer-purchase-intent/).

B2B sellers have never had a panel. Nobody can survey 300 finance directors about one proposal to one account. So my question is a different one: can a synthetic customer be built for B2B?

[Mark Davies](https://www.cranfield.ac.uk/som/people/mark-davies-424615) and I have kind of been experimenting with this at [Value Matters](https://value-matters.net/) and in research for [our new book](https://value-matters.net/cvp/). This is a guide to what the research shows, where it breaks and what a proper synthetic B2B customer would need, as questions and answers, with five figures to try.

## What is a synthetic customer?

A language model that is told who to be, shown something and asked to react as that person.

Researchers call it silicon sampling, after [a 2023 study](https://doi.org/10.1017/pan.2023.2) that used the idea on [US election surveys](https://electionstudies.org/). The appeal is cost and speed. A human panel has to be recruited and paid. A synthetic one takes minutes and can be asked again tomorrow.

[Bain tested the idea](https://www.bain.com/insights/synthetic-customers-earn-their-stripes/) in May 2026 for a consumer technology company. It built digital twins from the firm's own customer data and reported that they reproduced about 90% of the key outcomes of an earlier study. Bain also sees a use in B2B, where a market may hold too few key customers, such as the finance chiefs of one industry, to sample in the usual way.

## What did the Colgate study find?

Two things, and the first was a failure.

Asked for a [score from 1 to 5](https://en.wikipedia.org/wiki/Likert_scale), the models usually said 3. Real people mostly said 4 or 5. A panel that answers 3 to everything makes every product look the same.

The fix was to stop asking for a number. Each synthetic consumer wrote a sentence or two about whether they would buy. A second step then measured how close those words sat to five reference statements, one for each point on the scale. The authors call it [semantic similarity](https://en.wikipedia.org/wiki/Semantic_similarity) rating, and [the code is public](https://github.com/pymc-labs/semantic-similarity-rating).

<figure class="eth-fig sy-fig" id="sy-f1">
<div class="eth-head"><span class="eth-no">Figure 1</span><button class="eth-replay" type="button">Replay</button></div>
<p class="eth-title">A number, or words</p>
<div class="sy-tog" role="group" aria-label="How the AI was asked"></div>
<div class="sy-hist" role="img" aria-label="Share of answers at each point from 1 to 5, real people beside synthetic consumers."></div>
<p class="sy-key"><span class="sy-sw real"></span>Real people <span class="sy-sw syn"></span>Synthetic consumers</p>
<p class="sy-out" aria-live="polite"></p>
<figcaption>Illustrative shapes, drawn to match the averages <a href="https://arxiv.org/html/2510.08338v3">the paper</a> reports for <a href="https://openai.com/index/hello-gpt-4o/">GPT-4o</a>: mean scores of 4.0 for people, 2.96 for the AI asked for a number and 3.77 for the AI asked for words, with similarity scores of 0.26 and 0.88. These are not the paper's raw data.</figcaption>
</figure>

With that change the spread of answers looked like the human one. The synthetic panels also put the 57 concepts in close to the same order as the people did.

## What does the 90% mean?

Less than the headline suggests.

Human panels are noisy. When one is split in half, the two halves disagree a little about which concept is best. The paper measures that ceiling, then reports the AI's result as a share of it.

On [the paper's own figures](https://arxiv.org/html/2510.08338v3), two halves of a human panel [agree](https://en.wikipedia.org/wiki/Pearson_correlation_coefficient) at about 0.8, on a scale where 1 is perfect. The best synthetic panel agreed with the humans at 0.72. That is 90% of the ceiling, which is a long way from being right nine times in ten.

<figure class="eth-fig sy-fig" id="sy-f2">
<div class="eth-head"><span class="eth-no">Figure 2</span><button class="eth-replay" type="button">Replay</button></div>
<p class="eth-title">What the 90% measures</p>
<div class="sy-bars" role="img" aria-label="Agreement with a human panel on which concepts score best. Two halves of the human panel: about 0.80. AI asked for words, then scored: 0.72. AI rating its own words: 0.69. AI asked for a number: 0.66."></div>
<figcaption>How closely each panel's ranking of the 57 concepts agreed with a human panel (<a href="https://en.wikipedia.org/wiki/Pearson_correlation_coefficient">correlation</a> of average scores). <a href="https://openai.com/index/hello-gpt-4o/">GPT-4o</a>, shown the concept as an image. The three AI figures are from <a href="https://arxiv.org/html/2510.08338v3">the paper's results table</a>. The human figure is calculated from the paper's own ratios.</figcaption>
</figure>

[Figure 2](#sy-f2) shows something the headlines missed. Asking for a number scored 0.66 on ranking, so the new method added six points. Its bigger contribution is in [Figure 1](#sy-f1): answers with a realistic spread.

## What did the persona do?

Most of the work.

Each synthetic consumer was given the age, income and other details of one real respondent. When the researchers took those details away, something odd happened. The spread of answers looked even more like the human one. The panel also lost the ability to tell a good concept from a weak one.

<figure class="eth-fig sy-fig" id="sy-f3">
<div class="eth-head"><span class="eth-no">Figure 3</span><button class="eth-replay" type="button">Replay</button></div>
<p class="eth-title">The same panel without the persona</p>
<div class="sy-tog" role="group" aria-label="With or without the persona"></div>
<div class="sy-meters"></div>
<p class="sy-out" aria-live="polite"></p>
<figcaption><a href="https://blog.google/technology/google-deepmind/google-gemini-ai-update-december-2024/">Gemini 2.0 Flash</a>, shown the concept as an image, using the best-performing set of reference statements in each case. Figures from <a href="https://arxiv.org/html/2510.08338v3">the paper's results table</a>.</figcaption>
</figure>

A synthetic panel can look right and carry very little information. The signal came from knowing who was being asked.

For anyone in B2B, I think that is the most useful finding in the paper. Knowing who is being asked is the job in [key account management](/insights-from-the-key-account-management-forum-conference/).

## Where does it break?

The authors are open about the limits. The reference statements were tuned by hand on the same 57 surveys. The test measured what people said they would do and never what they bought. Every author works for [PyMC Labs](https://www.pymc-labs.com/), a consultancy, or for [Colgate-Palmolive](https://www.colgatepalmolive.com/). It is a [preprint](https://en.wikipedia.org/wiki/Preprint), and at the time of writing no [peer-reviewed](https://en.wikipedia.org/wiki/Peer_review) version has appeared.

They also explain why it worked. Toothpaste is discussed all over the internet, so the model has read thousands of real opinions about it. Where that background is thin, they write, the method "will not conjure valid consumer preferences".

Two later studies tested the idea independently. Both are preprints as well.

[Heath and Alexander](https://arxiv.org/abs/2607.28550) applied the method to US political surveys in July 2026, using three newer models. The spread of answers improved again. The bias stayed: synthetic voters held more extreme views than real ones, and the method needed a small set of real answers to calibrate it.

[Pandey and Jajoo](https://arxiv.org/abs/2607.24649), at a company called [Kairosity](https://kairosity.ai), ran a sunscreen test with 94 people and compared the reasons as well as the scores. The synthetic reasons often came straight from the pitch. They call this "concept echoing". The model reads the selling points and plays them back as reasons to buy, including points that real people had rejected.

<figure class="eth-fig sy-fig" id="sy-f4">
<div class="eth-head"><span class="eth-no">Figure 4</span><button class="eth-replay" type="button">Replay</button></div>
<p class="eth-title">The pitch comes back as the reason</p>
<blockquote class="sy-prop"></blockquote>
<div class="sy-tog" role="group" aria-label="Which buyer"></div>
<div class="sy-claims" aria-live="polite"></div>
<p class="sy-out"></p>
<figcaption>An invented example of concept echoing, moved into B2B. Brooks Translations, the medical device maker and its head of regulatory affairs are all invented, and both sets of reactions were written for this figure.</figcaption>
</figure>

For a seller that is the failure that matters. An account manager rarely acts on a score. They act on the objection.

## Why is B2B harder?

Three reasons.

The first is the one the [Colgate authors](https://arxiv.org/abs/2510.08338) gave. A [procurement](/kraljic-matrix/) director's view of one supplier isn't on the internet, so the model has nothing to have read.

The second is numbers. A consumer study can check itself against 9,300 people. A key account has one buying group and one decision, so there is very little to check a synthetic buyer against.

The third is that the customer is a group. A [decision-making unit](https://en.wikipedia.org/wiki/Buying_center) is several people who want different things, and the outcome comes from the argument between them. All three studies simulate one person at a time.

Every piece of evidence above is consumer or political. None of it tests a B2B buyer.

## How have we kind of done it already?

In July 2026 I built the [DMU Simulator](https://dmu-simulator.netlify.app/) for [Value Matters](https://value-matters.net/), the business I run with [Mark Davies](https://www.cranfield.ac.uk/som/people/mark-davies-424615). It runs a [value proposition](https://en.wikipedia.org/wiki/Value_proposition) past a simulated buying group. The group gives a verdict. The tool maps where its members want incompatible things, the seller answers each objection, and a sharper proposition goes back through the same room.

We made one design decision early that the research now supports. Each seat is profiled on evidence rather than personality: the role, what the person is measured on, what they have said or signed off, and what they stand to lose if it goes wrong.

The [Colgate paper](https://arxiv.org/abs/2510.08338) found that the persona carried the signal. The [sunscreen study](https://arxiv.org/abs/2607.24649) found that a thin persona echoes the pitch. Both point the same way. A seat is only as good as what is known about the person in it.

<figure class="eth-fig sy-fig" id="sy-f5">
<div class="eth-head"><span class="eth-no">Figure 5</span><button class="eth-replay" type="button">Replay</button></div>
<p class="eth-title">A seat built from evidence</p>
<div class="sy-tog" role="group" aria-label="How much is known about this buyer"></div>
<dl class="sy-seat"></dl>
<p class="sy-obj" aria-live="polite"></p>
<p class="sy-out"></p>
<figcaption>The same invented buyer as Figure 4, with one more layer of evidence added at each step. The four layers are the ones the <a href="https://dmu-simulator.netlify.app/">DMU Simulator</a> profiles a seat on. The wording was written for this figure and is not output from the tool.</figcaption>
</figure>

I run a smaller version on [Cranfield's Key Account Management Best Practice programme](https://www.cranfield.ac.uk/som/kambp). An invented biscuit maker pitches to three invented supermarket buyers, each with a short CV that tells the AI who they are. I described that exercise in [What AI can now do for your most important customers](/ai-for-key-accounts/).

What we haven't done is the part the researchers did. Nobody has scored our simulated buying groups against what the real ones went on to decide. So I'd call it a rehearsal tool, and an unmeasured one so far.

## What would a proper one need?

Four things, on the evidence so far.

**Real material.** What the customer has said and written. The version I have in mind builds each seat from real customer information: our own notes, their feedback and what they have said in public. [Bain's test](https://www.bain.com/insights/synthetic-customers-earn-their-stripes/) points the same way, since its twins were built from the company's own customer data. A named person's data needs care, so I'd keep it to what the customer has told us or published, inside a tool the company has approved.

**Words before scores.** The seat reacts in sentences and the scoring happens separately. That is the [Colgate lesson](https://arxiv.org/abs/2510.08338).

**A source for every objection.** Each reaction points to the note, email or statement it came from. An objection with no source is probably the pitch coming back.

**A scorecard.** After the real meeting, what the customer said goes back in and is compared with what the seat predicted. The [political study](https://arxiv.org/abs/2607.28550) suggests a small set of real answers goes a long way.

## What can it be trusted for today?

| Job | Trust it? | Why |
| --- | --- | --- |
| Choosing between two versions of a proposition | Fairly well | Ranking is what the [Colgate study](https://arxiv.org/abs/2510.08338) measured, though only for consumers. |
| Finding the weak claim before the meeting | Yes | Synthetic consumers marked weak concepts down harder than people did. |
| Producing questions for the real customer | Yes | A wrong guess costs one question. |
| Explaining why the customer will say yes | No | Reasons are the weakest output, and they tend to [echo the pitch](https://arxiv.org/abs/2607.24649). |
| Forecasting a win | No | Nothing here measured a real purchase, and nothing was B2B. |

## Does it replace talking to the customer?

No. [Value is defined by the customer](/value-based-selling/) and created by the customer in use. A model that has never met them can only rehearse the conversation.

The use I trust most is the one from the classroom. The simulated buyer raises an objection, and the objection becomes a question for the real one. Every answer turns a guess into something known.

Which seat in your biggest account could you fill in from evidence today, and which would be a job title and a guess?

**Sources:** Maier, B. F. and others, "[LLMs reproduce human purchase intent via semantic similarity elicitation of Likert ratings](https://arxiv.org/abs/2510.08338)", arXiv preprint (October 2025), with [code](https://github.com/pymc-labs/semantic-similarity-rating); Heath, O. and Alexander, R., "[Correcting mode collapse in silicon sampling with semantic similarity rating](https://arxiv.org/abs/2607.28550)", arXiv preprint (July 2026); Pandey, A. and Jajoo, G., "[Reason-mediated behavioral models for auditing LLM social simulators](https://arxiv.org/abs/2607.24649)", arXiv preprint (July 2026); Argyle, L. P. and others, "[Out of one, many: using language models to simulate human samples](https://doi.org/10.1017/pan.2023.2)", *Political Analysis* (2023); Bain & Company, "[Synthetic customers earn their stripes](https://www.bain.com/insights/synthetic-customers-earn-their-stripes/)" (4 May 2026); PPC Land, "[LLMs achieve 90% accuracy in replicating consumer purchase intent](https://ppc.land/llms-achieve-90-accuracy-in-replicating-consumer-purchase-intent/)" (2025); [Value Matters](https://value-matters.net/), [DMU Simulator](https://dmu-simulator.netlify.app/); Davies, M., Brooks, R. and Holt, S., [*Creating Customer Value Propositions with AI*](https://value-matters.net/cvp/), [Kogan Page](https://www.koganpage.com/) (due July 2027); Cranfield School of Management, [Key Account Management Best Practice](https://www.cranfield.ac.uk/som/kambp).

<script src="/essays/synth/figures.js" defer></script>