A chatbot built five B2B buyers for Germany's biggest sales study
The ECC CLUB Studie 2026, published by Cologne's ECC Köln research institute, did not interview its five buyer personas. It generated them. The datengetriebene Effizienz-Einkäufer (data-driven efficiency buyer), a conservative relationship buyer, a digital native, a risk-averse security buyer, and a transactional price hunter were built through AI and interviewed via a chatbot built for the study [1].
Their answers fed into the results alongside 25 real executive interviews and earlier ECC research [1]. This is not a one-off stunt. It is the latest instance of a broader shift already underway in market research: agencies and institutes supplementing, or in some cases substituting, human panels with large language models configured to answer as if they were a target buyer.
The practice has a name, synthetic respondents, and a small but growing body of validation research behind it. For B2B commerce leaders who commission or read this kind of study to shape go-to-market decisions, what that research shows matters more than the novelty of the method.
How did ECC build its personas, and what does it still need to disclose?
ECC Köln says the personas rest on prior empirical studies from IFH and ECC on B2B buying behavior, turned into agents governed by a role profile, company context, priorities, pain points and response rules [1]. Two questions remain open: which language model powers the chatbot, and whether its answers were checked against real buyers. The institute says it will address both in the coming days [1].
Crucially, ECC itself draws a hard line in the study's published key insights, flagging every persona statement as AI-generated and not an original quote from a real person, kept separate from genuine survey results such as the information channels buyers reported using [1]. No percentages are attached to the personas. ECC frames them as a way to make different buyer expectations explorable, not measurable [1].
Where synthetic respondents hold up, and where they break down
Synthetic respondents track real consumers closely on broad, low-specificity questions like price or material, but diverge sharply once a question turns to a specific brand or niche category. That is the core finding of a comparison the Nürnberg Institute for Market Decisions (NIM) ran, pitting 500 real US consumers against 500 GPT-4-based respondents built from identical demographic profiles [1].
The gap showed up unevenly across categories:
NIM found the AI systematically overestimated large brands, underestimated small ones, and answered with less variation than real people. The institute says this makes synthetic respondents useful for directional signals but not a substitute for real data, particularly in niche markets [1].
| Question category | Agreement between AI and real consumers |
|---|---|
| General purchase criteria (price, material) | Close alignment [1] |
| Soft drink brand preferences | Diverged on 75% of questions [1] |
| Sportswear brand preferences | Diverged on 80% of questions [1] |
Does academic research reach the same conclusion?
Academic research outside market research reaches a similar conclusion: synthetic respondents compress variation, even when they land close to population averages. A 2024 Political Analysis study by James Bisbee and colleagues found ChatGPT-generated answers matched the average values of a large US election survey reasonably well, but with much less spread than real responses [1].
“Synthetic respondents are a tool for directional signals, not a replacement.”
The same prompt also produced different results across April, June, and July 2023, a reproducibility problem that matters for anyone planning to run the same synthetic panel twice [1].
Why is B2B buying the hardest test case?
No validation study specifically for B2B buyers exists yet, but B2B purchasing is precisely where synthetic respondents are weakest: niche categories, multi-year contracts, and several stakeholders in one decision [1]. B2B buying committees are also growing. Forrester's State of Business Buying, 2026 puts the average purchase decision at 13 internal and nine external stakeholders, a figure that doubles when the purchase includes generative AI features [4].
That same Forrester survey found buyers are using generative AI for research while actively distrusting its output: 36% of respondents felt more confident in their decision because they used generative AI, but 20% said they felt less confident after encountering unreliable or inaccurate information [4].
Buyers increasingly validate what AI tools produce against peers, analysts and providers before acting on it [4]. Applying an unvalidated AI-generated persona to model those same buyers inherits the same risk from the other direction.
What should you check before trusting a synthetic panel?
A synthetic persona used in a published study should be treated as a hypothesis about a buyer type, not as a measurement of what buyers actually think. Before accepting results built on AI-generated respondents, buyers and researchers can check four things [1]:
Forrester's own framing of the category reinforces the same discipline from the vendor side. It defines synthetic personas as digital representations of specific job roles built to simulate buyer attitudes and behaviors inside AI workflows, explicitly positioned to extend human research rather than replace it [3].
It warns that personas built from a generic prompt without proprietary data and primary research produce profiles that look useful but miss the complexity of real buyers [3].
- 1Which language model generated the responses.
- 2What empirical data the personas were built on.
- 3Whether persona answers were validated against real respondents.
- 4Whether persona-based and human-based results are reported separately in the findings.
A vendor category is forming around synthetic audiences
The caution from researchers has not slowed commercial interest. Gartner's June 2026 research on synthetic data for marketing tracks four vendors, Arima, Verve, Panoplai and Semilattice, building dedicated synthetic-audience platforms. Gartner frames the push as a response to consumer insight becoming harder and more expensive as privacy rules tighten and AI agents enter the buying journey [2].
That is a marketing-side read on the same dynamic ECC Köln just applied to B2B sales research: real data collection is getting costlier and slower, and AI-generated stand-ins are the fast, cheap alternative being tested to fill the gap.
The tension is not only methodological. A survey of real panel participants presented by Hariet Köstner and Michael Schiller in the Jahrbuch der Marktforschung found many respondents worried that synthetic respondents threaten the authenticity and trust that underpin market research in the first place [1]. Vendors building synthetic panels are, in effect, competing for credibility with the very audience they are trying to simulate.
What this means for B2B commerce teams
Treat any study that includes AI-generated personas as two separate data sets, not one. The real survey results and the synthetic persona commentary carry different evidential weight, and a study that doesn't visibly separate them should be read with more skepticism, not less.
For niche B2B categories specifically, the comparison research available so far points the same direction: synthetic respondents can suggest where to look, but they should not be the basis for a go-to-market decision on their own.
Sources
- [1]etailment, Fünf Einkäufer aus dem Chatbot: Was KI-Personas in der B2B-Forschung leisten — etailment.de
- [2]Gartner, Coolest Vendor Innovations in Synthetic Data for Marketing — gartner.com
- [3]Forrester, The Real Promise Of Synthetic Personas Is Better Activation — forrester.com
- [4]Digital Commerce 360, Forrester: B2B buying groups expand as they question AI — digitalcommerce360.com
