Salesforce announced that it has signed a definitive agreement to acquire Listen Labs, a platform whose AI agents design research studies, recruit participants, conduct interviews and synthesise the findings. Financial terms have not been disclosed. The deal is expected to close in Q4 FY2027, subject to customary conditions and regulatory clearance.
Salesforce says Listen will complement Marketing Cloud, Service Cloud and its wider AI portfolio, and that it can reduce research projects that once took months to days.
More than a survey tool
Most AI in customer experience analyses feedback customers have already provided. Listen's agents go and collect it. According to Salesforce, they run audio or video interviews around the clock in more than 120 languages, drawing on a network of more than 50 million participants, and adapt their follow-up questions to what respondents say rather than following a fixed script.
This is the distinction worth holding onto: AI analysing customer feedback is one thing, and AI conducting the research is another. TechCrunch reported before the announcement that Listen's customers included Microsoft, Canva, Anthropic and Sweetgreen.
The digital twin question
The more contentious part of the deal is digital twins. Salesforce describes them as simulations built from research data, used to test how customers might respond to a product, message or idea before launch. Two separate claims are bundled together here. The first is whether AI can moderate interviews comparably to human researchers. The second is whether a model built from those interviews can reliably simulate how those customers might respond. The evidence for each differs.
On the first, the early evidence is promising. A paper posted to arXiv on 24 September, days before the announcement, tested AI-moderated interviews against human-moderated and static ones with 317 participants. In the study, AI moderation produced interviews comparable to human moderation in depth, covered more themes and recovered more customer needs for the same research budget. Participants showed greater emotional engagement with human interviewers. The human-moderated group was small, at 24 participants, against 139 for AI moderation and 154 for static interviews.
On the second, the results are more mixed. In the same paper, twins built from AI-moderated interviews predicted consumer responses better than demographics-only personas, but the richer interview data did not improve quantitative predictions over static interviews. The paper links prediction errors to differences in thinking styles between people and their twins, as well as to questions that fall outside the data used to build the model.
A study in the NIM Marketing Intelligence Review points the same way. Synthetic respondents captured broad patterns but overstated positive attitudes and gave more uniform answers than real consumers. The authors suggest synthetic data may currently be better suited to early-stage concept testing and lower-stakes applications than to decisions requiring precise insights.
Neither study validates Listen's commercial twins specifically. They provide evidence about the broader techniques Salesforce is betting on. A twin is not the customer. It is a model built from evidence about customers, and it is only as good as that evidence.
Why it fits Salesforce
Salesforce already holds vast amounts of structured behavioural data: purchases, service interactions, sales activity and marketing engagement. This shows aspects of what customers did. Listen adds what they say and why.
The deal extends an argument we made earlier, that Salesforce increasingly wants to be the context, data, permissions and action layer behind agents. Customer understanding becomes one more layer in that stack, sitting alongside the records and workflows agents already draw on.
What changes if this works
Traditionally, research runs from research to report to human decision. A report lands, someone reads it, and a team decides what to do.
If the technology matures, one possible sequence is research, then model, then simulation, then AI-assisted or agentic decision. Interviews feed an evidence base, the evidence base supports simulations, and those simulations inform actions taken inside the CRM, with real customers consulted again where the model is uncertain. Research would stop being a separate project and become an input to systems that act.
This is an interpretation of where the pieces could fit, not a workflow Salesforce has described. It does explain why a CRM vendor would want a research company.
Simulated customers still need real ones
Before any of this reaches production, buyers should ask three questions. How representative is the research behind the twin? How far can the model be pushed beyond that evidence? And when should the company still ask real people?
The research so far suggests these systems are better treated as a complement to real customer research than a replacement for it. The technology can plausibly scale customer understanding. Whether it can reliably simulate what customers will do is a much harder claim.