The discussion covers the limitations of synthetic data compared to human-centric interviews, the technical challenges of programming agent judgment for follow-up questions, and the surprisingly high consumer acceptance and completion rates for AI-led studies. Mani shares his vision for Underheard to become an intelligence layer that enables companies to access customer insights on the same day to accelerate product development cycles.
Developing Voice Agents
Underheard is a Zurich-based company developing voice agents for customer interviews and market research. Can you tell us a little more about the company?
“We’re a new company,” says Mani. “We focus a lot on market research, specifically on trying to commoditize market research to companies so they can really understand their customers better. This is where the name Underheard also comes from, to really make those voices heard that usually companies might not have the option to. It stems from the notion that now everything is moving so fast, technology, product, everything has kind of become commoditized with AI. But at the same time, the thing that really matters, which is building products your customers like, building things they actually find useful, is becoming more challenging. Knowing what your customers want is not easy. You need to talk to them. You need to hear them. You need to understand them, and this is what we work on. We offer market research that can be done by any company of any size. That is our mission.”
So it’s an AI-based service developing voice agents?
“Yes,” says Madi. “It’s qualitative research in the sense of how you would have traditional researchers in big companies or even, of course, you might hire a third-party service to do market research for you. But you would need to spend a big budget. We are now doing market research for all different types of companies in size to hear what the customer says, and we use this with voice agent technology to make it affordable and most importantly quick. You need answers to your questions within days, and this is something we can provide that really compacts the time frame that it usually takes.”
Finding a Niche
What made you decide that this was the direction to go with Underheard, which is a pretty new venture developed as a startup? There’s a lot of developments happening with AI and synthetic panels and concepts like that. But why did you think this niche was a sweet spot?
“I think we actually reached that level quite early due to the experiences we’ve had,” says Mani. “One of our early customers was in the hospitality sector, and there we saw answers coming up from guests who stayed at this specific hotel. This was a continuous study we did with them. The results came in that usually you wouldn’t hear, whether that is about specific rooms having problems, someone wanting vegan options, details that we think synthetic data as it is currently is not the best at finding. From what we have found and seen, synthetic data is good at finding median data. But for very specific people, they are not the most common or average ones. Synthetic data doesn’t work as well from our experience. And I think the human touch is very important.”
He adds, “On the other hand, I do have to say, we are also very interested in synthetic data. I love market research. I think synthetic data is very interesting, and we’ve also dipped our toes into it. But it has a different use case, though, than replacing market research in my opinion. The use case is enhancing it. For example, if you conduct a study with 100 participants, and, of course, many of your questions might be answered. But, similarly, more questions usually do pop up. Usually, you get questions from these results that you can’t immediately answer, and synthetic data can expand on what you already have with more information.”
“This also plays into the concept of how does AI work? AI works a lot with the quality of the input. If we can feed synthetic data, AI could generate it with really good input that we have generated beforehand with actual customer data. I do believe there is definitely a place for synthetic data. It will improve over time. But I don’t think it’s a full replacement of human opinions and emotions. I think that’s very hard to simulate,” Mani says.
Judging the Agent
There’s a lot to unpack with AI about using human judgment and human emotion. What were some challenges getting Underheard to launch? As far as developing the AI interviewer, and using it and scaling it?
“To some extent, the space has become commoditized,” says Mani. “There’s a lot of different players. I think the hard part now is the judgment of the agent. For the agent to actually know when is it important to follow-up, when is it important to maybe stop at a certain point because these kind of human judgments are very tough to programmatically implant into an agent. There’s a lot of prompt engineering, a lot of systems and layers on top of it to achieve it. Still, I think this is technologically not the easiest thing to achieve.”
“Many researchers realize that there’s usually a time frame in an interview where a respondent becomes very open. And after a while, they become tired and maybe less responsive. To really get this feeling of when is the right time to ask questions, in what frame, and for how long, this is something a human is very good at, but agents still struggle. This is the technical challenge that we’re tackling a lot right now to make sure, aside from the voice and the conversation flowing well, that the interview itself is of a quality a researcher would respect as well.”
New Frontiers
With some of your current use cases, what’s the comfort level been like with the AI led conversations? Has it changed the way you think about the technology or where it’s headed next?
“Yes, I do think so. In the beginning, we were quite worried about the responses from people,” says Mani. “To some extent, we thought we’d get quite a few negative responses from people talking to agents. Especially here in Europe, people are still quite cautious. What we have found, though, is quite surprising, that people are very open. For example, in a study which we did recently, we had about a 92% completion rate of the study. In Europe, of course, you have to be transparent about it being an AI, and it wasn’t really an issue.”
He adds, “I think what we have found is that the problem people had with AI is the intelligence. We’re used to the typical AI that you might get from your health insurance provider where you’re on the phone call and it can’t really answer your questions. But if it is actually intelligent, if it feels like a proper conversation, people are surprisingly comfortable to the point where they actually almost disclose more information to an agent rather than to a human.”
Developing an Always-On Intelligence Layer
Consumer trust is so important. Now that your platform is available, what do you think is the next frontier? Where are you excited to build or expand?
“First of all, I think we’re never finished,” Mani says. “The platform per se stands, but I think like the conversation and the interview itself, we are continuously fine tuning and making sure it really gets good. I think none of these platforms yet have achieved a level where we can say this is indistinguishable from a human. On the other hand, I do think there are bigger goals. We have, in the end, the goal to be an intelligence layer for companies. Meaning, this doesn’t just involve market research. It means whenever you have something you want to work with, as a human, if you want to develop something, if you want to do a new marketing promotion, that you have access to information within the same day.”
“For example, a small company here in Switzerland might want to ask, what if we launch this new chocolate promotion? Maybe they want to launch a new brand. And what they used to do is send out promotional material to test. But on the brand side, it takes too long. Our goal is for them to have a knowledge base and be able to access customers the same day, get answers to the questions, and therefore, just have the product development cycle a lot leaner and faster. That is our vision.”
Contributor
-
View all postsMatthew Kramer is the Digital Editor for All Things Insights & All Things Innovation. He has over 20 years of experience working in publishing and media companies, on a variety of business-to-business publications, websites and trade shows.






















































































































































































































































































