Salesforce and Listen Labs: The Deal Is Not the Story

The deal is not the story 

On 29 September, Salesforce signed a definitive agreement to acquire Listen Labs, which it describes as an “AI-powered customer research and human simulation platform”. Salesforce completed the acquisition on 2 November without disclosing the purchase price. Before the agreement, Business Insider reported that Salesforce had held talks to buy Listen Labs for around US$2 billion. 

Most of the discussion since the announcement has focused on the price and whether AI interviewers will replace human moderators. Those are fair questions. But I’m more interested in what happens when companies can commission research, link it to customer records and share the findings with decision-makers on one platform.  
 
Esomar asked members and friends of its AI Alliance what they made of the deal. Vanessa Oshima, founder and CEO of Heart-Data, sees a change in how companies buy research. The customer relationship management system, or CRM, holds a record of what customers did. Research can help explain why. Oshima describes the deal as “shifting consumer voice and research from a professional service that is briefed and purchased into a software feature inside the CRM flow”. 

David Wright, co-founder of Hello Ara, sees a reason for the purchase beyond market research. In his view, Salesforce has struggled to make money from AI and bought “a team that rose to the top extremely quickly with an AI-native service”. 

Manuel Garcia-Garcia, who leads science activation at Ipsos, sees value in what Salesforce can do with Listen Labs across its existing products. Salesforce “is not buying a research business as traditionally defined; it is buying a capability that can be embedded across a much larger customer, marketing, sales, service, and agent ecosystem”. I share that view. The same research could inform decisions in marketing, sales and customer service. 

From an insight industry to a decision industry 

For years I have argued that we should think of ourselves as a decision industry. We say we help organisations make better decisions, yet most of what we sell is research. The client receives the findings and decides what to do with them, sometimes weeks later. In some cases, the findings never inform a decision. 

Salesforce’s other deals show how it plans to do this. It announced its agreement to buy Loops, which helps explain changes in business metrics, on 12 October and completed that acquisition on 2 November. On 20 October, it announced an agreement to buy Braze, whose software helps marketers respond to customer behaviour in real time. On 2 November, it announced an agreement to buy Cresta, which analyses customer service calls and provides AI voice agents. Taken together, the deals suggest Salesforce wants its software to collect customer feedback, explain what drives behaviour and help companies respond. 

Wright points to a problem that the deals leave open: “Data assets remain as important as ever. I don’t think anyone has fully solved how to make use of all a company’s data assets safely and productively.” Researchers can help companies decide which data to combine and what conclusions the combined evidence supports. 

The new economics of insights 

If research becomes part of software that companies already use, they may commission fewer projects from agencies. As Oshima puts it, “Fewer briefs go to agencies, and procurement shifts from buying projects to buying seats and tickets.” 

Agencies risk becoming interchangeable if clients can use the same platform to buy similar services from several suppliers. The platform owner sets the terms of access and how the results reach the client. As Garcia-Garcia says, “The question may not be how large the market becomes, but who can create the greatest decision impact from AI-enabled insights.” 

Dharmendra Jain, CEO of Actnable AI in Nairobi, expects small and midsize traditional agencies to feel the greatest impact. Barry Jennings, Esomar’s North America ambassador, expects corporate researchers to drive adoption, above all at large companies that have a team or partner with access to internal data and the systems that connect it. 

Access to the technology still leaves questions about who can take part and how well the method works for them. Jain warns of “growing scepticism about AI-moderated research” in emerging markets, where researchers must account for “linguistic diversity and varied digital literacy levels.” 

Research firms can help clients make those judgements. They can assess who took part, where the evidence came from and whether the method supports the conclusion. A firm independent of the platform provider can give clients a separate assessment of the evidence. 

What makes evidence trustworthy 

A system that brings research closer to decisions also brings its errors closer. Faster answers are not better answers. If what we feed into these systems cannot be checked, we will simply industrialise our mistakes. 

At the Esomar Congress in Valencia, Esomar President Anne-Sophie Damelincourt argued that the profession must preserve trust and the ethical practices that sustain it. “Trust is not sentiment. It is the load-bearing structure of everything we do.” 

Noriko Kishida, of Insight Core and the Japan Marketing Research Association, recalls the promise researchers make to respondents that they will report their answers in a form that does not identify them. Linking survey responses to a CRM can let a company trace an answer back to a named customer. Researchers need to explain any proposed link to participants, seek their consent and put safeguards in place to protect them. As Kishida warns, “Research would no longer be safe for respondents.” 

Jennings recalls customer surveys enriched with “a select set of variables” but “no identifiable information that would allow us to trace it to a specific customer.” Smaller projects, stronger impact. Wider market studies were still needed, but “for customer research it was a real goldmine.” 

So, less talking and more evidence. Today every provider describes accuracy with its own metric, a confidence score here, a match rate there, with no common baseline, and many buyers do not understand what they are shown. Every provider of AI research should be able to lift the bonnet. Garcia-Garcia calls for “much stronger standards of provenance”: users “should be able to distinguish clearly between what was directly observed from people, what was inferred by AI, and what was simulated.” Scale is no substitute for design: “A thousand automated interviews do not automatically provide stronger evidence than thirty carefully recruited conversations.” The advantage will go to those who document their methods, their limits and their progress. 

The systems around research are richer and better connected than our industry alone. That is why an independent, verifiable voice is worth more inside them, not less, provided we can show it is right. 

What simulations can tell us: exploring possible futures 

On 22 October, Listen Labs closed the loop. It launched Listen Twins, which it describes as “1:1 digital twins of your audience, each built from an in-depth interview with a real person”. Listen Labs says it updates the twins with the client’s own customer data. Each set is “custom-built for your use case”, which also means a twin is only as good as the use it was built for. Interview real people, feed them into the system, simulate them. The next move is obvious: explore scenarios, compare options, rehearse decisions about offers and markets that do not exist yet. 

Simulation is not prediction. It is a way to explore possible futures before putting money on one of them. 

Our contributors raise two different objections. Kishida’s is ethical: digital twins “infer reasons and opinions about things the person never said”, profiling people “based on information they never gave.” Oshima’s is methodological: synthetic data trained on internal CRM data risks “creating an echo-chamber decision environment that is merely validating what has already happened versus finding true insight.” 

Both point to the same requirement: say what a simulation is built on, and check it against reality. To its credit, Listen Labs traces each twin’s answer to the interview quote behind it, with a confidence score, and lets a twin say it does not know. That is the provenance we need. The open question is consent: do the people behind the twins know they will go on answering questions they were never asked? 

Not every simulation raises it. A twin models an identifiable person, so consent comes first; a population simulation models no one in particular, and must show that the population is right. 

In September, Aaru reported that its simulations ranked 340 pairs of brands in the same order as purchases recorded in a card panel 92.4% of the time. Rankings from a published survey of purchase consideration matched the card panel’s rankings 77.8% of the time across the same pairs. Aaru ran the test itself. Its model estimated purchases, while the survey measured which brands people would consider buying. 

Who vouches for the evidence? 

Garcia-Garcia sees researchers becoming “architects and stewards of evidence”, there to keep organisations from becoming “very efficient at confirming what they already believe.” That means defining what evidence a decision requires, checking how it was collected and explaining its limits to the people who use it. 

Oshima points to a consequence for junior researchers. They have learned to judge evidence by doing the work, from designing studies to moderating interviews and coding responses. As software takes on more of those tasks, agencies and client teams will need to teach those skills through training and coaching. 

Kishida wants Esomar to set out when researchers may link research responses to CRM records and what they should disclose to participants about who will use the data, “for what purpose, and how”. I agree. The Esomar AI Alliance should take up that question. 

Salesforce’s acquisitions give it the means to connect customer research with the tools companies use to act on it. Now the test is whether that integration helps companies make better decisions. 

As more research enters those systems, researchers have a responsibility to set the standards for the evidence companies use. We should make clear what came from participants and what a model inferred or simulated. We should test whether that evidence can support the decision at hand and make sure companies use participants’ data in ways those participants agreed to. Whoever makes the decision should be able to see how those checks were done.