Description
Sarah Kling recounts her first encounter with internal AI tools that transformed days of analysis into seconds of work. She says this efficiency allows her team to scale research across international markets like Fire TV and tablets at a fraction of traditional costs. However, Sarah cautions that speed should not come at the expense of quality. She emphasizes the concept of the thinking cave, where researchers give data space to breathe to avoid producing flattened insights.
The conversation covers the risks of AI hallucinations and explains why researchers must act as stewards of truth by verifying every output. Sarah also addresses the unique privacy challenges of generative AI compared to social media, describing it as a black box where consumers feel a loss of control. Looking toward the future, she anticipates a move toward interlinked workflows where bespoke tools integrate securely with internal data, a progression that requires a beginner's mind and a commitment to the craft of discernment that machines cannot replicate.
Episode Resources
- Sarah Kling on LinkedIn
- Amazon Website
- Stephanie Vance on LinkedIn
- Molly Strawn-Carreño on LinkedIn
- The Curiosity Current: A Market Research Podcast on Apple Podcasts
- The Curiosity Current: A Market Research Podcast on Spotify
- The Curiosity Current: A Market Research Podcast on YouTube
Transcript
Sarah - 00:00:01:
Setting the other way to keep the insights insightful is to make sure you have humans who are not only looking over what the AI gives you back, but really taking it to the next level of, I'm gonna sit with us for a day or two or three, whatever it is. I'm gonna read it all through. I'm gonna go off into my thinking cave or whatever. Just let it settle, and then I'm gonna come back and reread it. And just like it would with any report I'm writing or any insights I'm putting together, I'm gonna give it space to breathe. Without that breathing room, though, it definitely can get flattened.
Molly - 00:00:32:
Hello, fellow insight seekers. I'm your host, Molly, and welcome to The Curiosity Current. We're so glad to have you here.
Stephanie - 00:00:39:
And I'm your host, Stephanie. We're here to dive into the fast-moving waters of market research, where curiosity isn't just encouraged, it's essential.
Molly - 00:00:49:
Each episode, we'll explore what's shaping the world of consumer behavior from fresh trends and new tech to the stories behind the data.
Stephanie - 00:00:57:
From bold innovations to the human quirks that move markets, we'll explore how curiosity fuels smarter research and sharper insights.
Molly - 00:01:06:
So, whether you're deep into the data or just here for the fun of discovery, grab your life vest and join us as we ride the curiosity current.
Stephanie - 00:01:16:
Today on The Curiosity Current, we're joined by Sarah Kling, Senior UX Researcher and Global Insights Lead for Amazon devices.
Molly - 00:01:25:
Sarah has spent more than two decades building research programs that drive real product and business impact, from startups all the way to global technology companies. Her work spans UX research, privacy, global insights, and increasingly, the strategic implementation of AI inside of research workflows.
Stephanie - 00:01:43:
At Amazon, Sarah leads apps and games research across Fire TV, tablets, and related services, reaching hundreds of millions of users globally. She also created the global insights program, helping scale research across international markets while exploring how AI can accelerate research without replacing human judgment.
Molly - 00:02:03:
So, today, we're diving into one of the biggest conversations happening right now in enterprise research, the promise of AI. What's actually working, what still isn't, and where AI may take the industry over the next 12 to 18 months.
Stephanie - 00:02:16:
Sarah, welcome to the show.
Sarah - 00:02:18:
Thank you. Thanks for having me. Pleasure to be here.
Stephanie - 00:02:21:
Awesome. Well, let's get right into it, Sarah. You have worked across startups, agencies, and some of the world's largest tech companies, and currently, you're helping pioneer AI integration into research workflows at Amazon. I am curious, was there a particular moment or experience where it really clicked for you that AI was going to fundamentally change how research teams operate?
Sarah - 00:02:47:
Oh, that's an amazing question. I think the first real click moment that I had was about a year and a half ago. I was using one of our internal AI tools. Amazon makes its own, and we dock food, our products, as well as things like Alexa, out in the market and stuff from AWS. But I was using one of our rudimentary systems internally, and I had a huge file of open-ended responses from a survey. And I hadn't yet done the thing that people were talking about, which is, “Oh, why don't you upload them into tool X and have it do some analysis and pattern recognition for you?” And so I thought, “Well, why not?” You know? I'm still distrustful that it's gonna do it right, but I don't have time to hand-code it or, you know, go through it, you know, other ways. So, let's do this. So, I uploaded my file and asked it a couple of questions in the prompt about the specific things around the open ends, and it spit out a very, I would say, accurate analysis when I cross-checked it in, like, seconds. And I think I got a little teary because, yeah, I realized that a lot of that hype was true, but also that I was just doing this a massive amount of work in such a tiny amount of time and producing it at a quality level that I could have at least gotten to, but it would have taken me days that I didn't have. And so from then on, I started uploading even back studies and open ends that had been partially analyzed or for speed, you know, speed to insight. We really hadn't been able to go deep on it. And I just started experimenting in my spare time and quickly became a convert of all of the things that we're probably gonna talk about today, even though there are some caveats, but that was the moment, so.
Stephanie - 00:04:25:
I love that. I feel like the open ends, I like to think of the ways in a lot because we hear, you know, from folks a lot on that, and that is a common one. Another one is that I put a spreadsheet in there, I put a whole thing of cross tabs that would have taken me forever to go through, and it's always those moments that take a tedious, very long task down to seconds that humbles you, but, like, awakens your mind to the possibility.
Sarah - 00:04:49:
Yep. Exactly. The crosstalk came later, but yes. Exactly.
Stephanie - 00:04:53:
The quant data moment came much later, but yeah, exactly. That was exactly it.
Molly - 00:04:58:
I think teary-eyed is a great way to describe that in both ways.
Stephanie - 00:05:04:
No more Post-it notes.
Sarah - 00:05:06:
Right.
Molly - 00:05:07:
I wanted to drill into something that you said, Sarah, which was around hype. There seems to always be in this implementation cycle a gap between what's the hype around AI and what's the true reality of actually implementing it, especially inside of a large organization. So, from where you sit, what parts of that AI adoption are genuinely delivering value versus what is just noise and hype?
Sarah - 00:05:31:
Yeah. And I mean, do you want me to focus specifically on research? Because, obviously, there's a lot of different ways we could talk about the hype even within, like, productivity in researchers' lives. I think if I focus on research and where the hype hits, I think there are a couple of areas that pop up immediately, which is, one, the promise that or the belief in some circles that AI can just do all of this for us, and we, as researchers, will become obsolete, right? Panic has been present. I think it's finally subsiding now. But, you know, a couple of years ago, there was a lot of understandable panic and hype that AI was gonna just be able to replace a majority of the work that researchers, and data scientists, and even possibly academics can do. The promise of it, of course, is nothing close, or the promise of it wasn't even that, but the belief that it could get there is one area. The other is just that it can do every single thing that you could do in research analysis, and replace your human brain's ability to make discerning decisions about what you're seeing in data and insights. And, again, experience is showing us daily in some cases for me that that is just not replaceable by AI. AI is great at pattern matching, it's great at some of the analysis, it's gotten so much better with quant data, etc, but it hasn't been able to replace the human brain's ability to interpret data and insights contextually, and that is just something that we keep seeing. And so when we experiment, or our stakeholders who are not researchers try things, experimentation or otherwise, we can just very quickly see where that goes off the rails too. So, I think those are the two areas that have really stuck out for me as a lot of promise, a lot of hype, and haven't panned out yet. And I'm not sure with the human side of it if it's ever going to, but I definitely don't believe we're gonna be replaced by AI researcher avatars anytime in the near future either.
Molly - 00:07:23:
I feel like that fear is not just in research. That's in a huge swath of people who are worried about AI taking their jobs, and I always hear it. But you're not gonna be replaced by AI, but you could be replaced by someone who knows how to utilize AI. So, a researcher that knows how to utilize these tools to increase productivity is going to be… that’s the next skill set that everyone should be skilling up to.
Sarah - 00:07:54:
Yep. Exactly. So, well said. Yeah. And that's what we've had versions of that, too. And even internally at Amazon, where we've had a mandate from the top down, leadership on down, pretty early on in the GenAI revolution to find ways to to enhance productivity, whatever that means for your job role or team, find ways to utilize tools that we have, and/or find tools in the world, externally, you know, if you need something bespoke. But do it in a way that's mindful because we have quality bars and standards, and we have, you know, relationships with our customers that matter. So, we don't wanna cut corners for the sake of AI, but we do need to keep finding ways to do it. But there's been a less aggressive like, we're gonna replace you with AI or else. It's more of a we expect you to augment, however that looks for you is up to you to really figure out and bring back to, you know, your management and your teams to demonstrate. And I think it's one of the healthier approaches I've heard about in the world, but it's still, because things transfer fast, and there's so many tools, and people are trying new stuff. And we're not even in research, doing it with a level of rigor we probably should be. So, it's one thing if you're automating your calendar and your inbox summaries with AI tools, quite another when you're automating a giant segmentation study and trying to really leapfrog ahead in it, not the same as pattern recognition in Slack messages, right? So, totally sure.
Stephanie - 00:09:20:
You know, in that vein, you've talked about moving beyond, you know, basic generative AI usage, throwing things into tool X, if you will, into more strategic, maybe agentic implementation. I'm curious, for you, what separates meaningful AI integration from simply layering AI tools onto, like, existing workflows?
Sarah - 00:09:40:
That's a really great question, and I wish I had a solid answer to it. I think the short answer is we're very much in the midst of figuring that out. I think where we've seen the most gains come in research and AI, I'll speak for my team and group internally at Amazon, is that we have augmented workflows, we have augmented data processing, we have augmented pattern recognition, the ability to summarize, you know, qualitative data, we have some vendors we've worked with who have excellent research teams who do that on our behalf, or we can extend our team with them, and they're using AI tooling to do that, but as far as agentic AI goes, we haven't really found a place yet for that in our research workflows in the sense that we're doing some experimentation with mechanisms that might be considered more agentic to gather customer feedback, but it's super early days. And we're still just trying to figure out if that makes sense yet. But right now, it's really just, again, lifting up those low-level tasks and getting them off our lap in a way that really enables us to focus on more strategic insights and faster speed to impact. And then, you know, finding vendors and partners that do the same kind of work for us to help us extend that. And that's been effective. I mean, I have to say, like, when I launched the global insights program, I know you guys mentioned it in the intro, a couple of years ago, we put together the framework for it, and then it really launched in pilot form in 2025, and now I'm expanding it in 2026. A lot of the questions were just speed to insight. Like, if we can now do the same study replicated in five countries or seven countries because we have the same device that we sell, we have Fire TV customers everywhere, for example. And now, we can really understand across the different countries, like what the common needs are and what the differences are. AI has really helped accelerate our ability to do that and cut the cost down in over traditional methods, right? If we had to do different studies with translation, localization, and agencies, the cost might even be prohibitive. It certainly would be staggered, and we definitely wouldn't be able to look across our insights as well as, you know, vertically.
Stephanie - 00:11:45:
Yeah.
Sarah - 00:11:46:
And I think when I just look at the ROI and look at the number of studies completed, the depth of them, those numbers are, I would consider basic. We don't measure our research impact in volume of studies, usually, right? Like, you know.
Stephanie - 00:11:58:
Yeah. Yeah. Yeah. Yeah.
Sarah - 00:12:00:
20 great studies or 20 not great studies. That's not the point. But when I look at it, we've had the volume and the impact, and a lot of that's been AI-enabled. And so for us, it's an instant win. Are we getting better insights because of AI? I don't know if they are better. I just think they're faster, but we are still very much humans part of that process. And so, I think it's a very, again, we're still at early stages around the metrics, and we're still looking at ROIs. How do we run this with an agency, or run this across 8 countries with an agency there? This would've cost us hundreds of thousands of U.S dollars versus, like, tens of thousands of U.S dollars. And so those numbers have been pleasing to management and to our finance people. But, also, we're just happy to be able to get the work done at a level that we just couldn't before and to be able to connect those dots in ways that weren't possible. So, I'm still reveling in that as a researcher. And getting into our agentic stuff just seems like an amazing world to go into, but it's early days. So, I know it's a long-winded answer.
Stephanie - 00:12:57:
No. Totally. And I honestly find it really refreshing to hear that because I think there's a lot of pressure across many industries to move to these kinds of agentic experiences. But what you're really saying is that without even having done that and just deploying AI, you know, within the context of different research tasks, you're still able to have, you know, gotten to better times to insights as well as money savings. So, I mean, you're doing a lot without even having those agentic experiences.
Sarah - 00:13:26:
Right. And there's not a lot of demand from our leadership to do agentic, for example. I do research on agentic AI in consumers, by the way. So, you know, like, there's that word, and everyone's head, you know, glazes over, you know, outside of the industry. So, but I think, yeah, the demand will come, and there'll be good use cases, I think, that will make themselves apparent. But right now, the value that we're getting is so at a baseline, just so much better than we ever could have gotten 2, 3 years ago, that the reach, the value, and the ability to, again, connect those dots between things that just we knew we could have connected the dots, but no one had the time. And so now we're doing that, and I guess that's the crawl of the crawl, walk, run world, maybe, but I don't think of it as crawling. I think it's like we're supercharged in what we can deliver compared to two years ago, or even 18 months ago. So, I'm just thankful on a daily basis for those tools that enable that. And I know our vendors and partners are, and our stakeholders are as well. So, on the whole, it's been good. It's been a good run so far.
Molly - 00:14:28:
And when you talk about the experimentation and figuring out, actually, what works and doesn't. You talked about what's really working well, but I almost want to, I'm curious about the, like, how did you figure out what wasn't working very well? So, as teams are experimenting with AI, sometimes they struggle to move from this fun experience, it's novel, it's cool, to the actual business and operational impact down the road. So, where do you see in that sort of, like, this is fun versus this is actually making a difference? Where do organizations get stuck in that sort of transition period?
Sarah - 00:15:02:
Yeah. I think there's a couple of questions in there, Molly, that I'm gonna maybe try to tease apart. So, stop me if I'm taking this in a direction you didn't intend. But I think there's the question of how do we know we're gonna get results? Like, I gave the example when you asked me about the first time it clicked about uploading open ends and getting analysis back. But the rigorous side of that or semi-rigorous side is, how do I know it's true? Like, how do I know the AI didn't just hallucinate and make stuff up? So, there's a whole method around, well, how do I test and go to source? How do I ask the AI questions tool X in this case, about, “Hey, that quote, where did that come from? Tell me where it is in my spreadsheet. Give me the line number, you know, etc.” And really being able to go back and do those checks and remembering as I move forward with this and get excited to try more and more things, just open-ended analysis, to keep doing those rigorous checks and not just taking it for granted that it's right because there's no way to necessarily know. I mean, our models hallucinate less and less over time, but the errors or the misinterpretation gets more subtle, right? And so you still have to have that rigor. And sometimes you have to have even more rigor the more powerful the tool seems, because it is capable of hiding its mistakes better. And I think what we are always reminding ourselves of as researchers and trying to remind our stakeholders, who are non-researchers, is that AI isn't perfect. It has bias. The tools have bias. The models have bias. And, you know, sometimes the math doesn't ‘math’ as my boss likes to say. Like, you have to really look at, okay, the ad produced this beautiful report document that we have internal tools that do this, writing in the Amazon writing style. It's done all the right things. It looks great. It's the right number of pages. We have some very interesting guidelines around how Amazon does things for documentation. And it'll look perfect, but then when you dig into it, it's not that it's nonsense, but it doesn't hold up. And you can start to pick or even cherry-pick, like, let's figure out where that quote came from. Let's make sure it's contextualized, right? Did that matrix response in that survey, that data get analyzed? Right? Because we know that's a weakness in some of the models when they analyze matrix responses, you know, things like this. But this is all the learning that's come along the way. And, you know, we've learned the hard way where, like, we will run raw data from a study through several different tools now, and maybe our vendor will deliver something, and we'll look at it internally through two or three different lenses just to make sure everything matches up. And that's been added over time. So, the early days of trust were, oh, yeah. Sure. Those numbers are right. That must be correct. And now it's like, nah, actually, in a way, I almost trusted less over time. So, then I don't trust it. It's just that I've had to add in so many more guardrails because I've had things come back where the math did math or the quote was made up or somebody who's a smart stakeholder is like, that doesn't match our behavioral data. Where did that number come from? You know, things like that that just kind of sneak in on you. And as researchers, especially, we're supposed to be the stewards of truth from data or as much truth as we can get. So, if we're introducing bias or we're introducing, you know, lower quality methods accidentally, that's on us. But I think it's also just having moments of embarrassment or moments of experimentation where we've put data, and then it's come back, and it's all garbled, you know, and things like that. And that's why we run things through multiple tools and do these other steps. But that's evolved for me over the last 18 months. I don't know, Molly. Did that get at all of your questions?
Molly - 00:18:25:
It definitely did. And I think it also started getting me to think about the differentiation, sort of too, between the rigorous research that you vetted, that you've checked, that the math is mathing, as you said, versus seeing AI as an ultimate source of truth. I think that's also where it gets a bit dangerous, if you have stakeholders that are saying, “Well, this is what the AI said. It's 100% true.” And then there's an additional layer for you to say, actually, this is why we're redoing this, or we're changing this. So, this is the perspective shift that you should have on this.
Sarah - 00:18:58:
Yeah. And I think that's exactly it. And I mean, I would say in my experience internally, luckily, we have, I think, a pretty healthy set of training around why AI is not the source of truth. So, we all have access to a lot of materials that are, like, please don't believe that. Like, it's like believing everything you read online is true, right, or in the newspaper or whatever source of news is. There's that, but I think that common sense, conventional wisdom side of it can get lost in the magic. At the moment, you're like, but I just ran it through this tool, and it produces a beautiful document, and everything looks good. And it really did catch five of the six points it's supposed to, but that sixth point was wacky, right? And that's where your whole, you know, your whole strategy or argument or presentation is gonna fall apart, as if, you know, one of your pieces of data or one of your conclusions or recommendations is off by a crazy amount, the other ones get cast also into doubt. So, treating AI as a tool to get you to the answers or the ideas or the, you know, the insights, if you will, or recommendations is great, but you definitely can't rely on it to be the BLN doll. And it's easy to get pulled into, but when you see a finished report that looks like it's done all your work for you, it's hard not to do the human thing. Sometimes it'd be like, oh, okay, good enough, right? But you're like, what does that actually mean in this case? I still have to do my job. And other people, even that researcher's same position, I think they're just less familiar with some of the rigorous ways to test it. So, that's part of our education as well.
Stephanie - 00:20:24:
It's so sneaky because, like, the writing is good, right? And good writing hides bad data sometimes.
Molly - 00:20:35:
Yeah. Hides anything.
Stephanie - 00:20:38:
Exactly. Yeah.
Sarah - 00:20:40:
Not good thinking, but great writing.
Stephanie - 00:20:42:
Yeah. Yeah. Exactly. I think a lot about the tension between acceleration, whether it's, you know, from AI or just all the other kinds of automations we have at our fingertips and deep insight. You know, these real kernels of truth that we're often trying to get to that take a lot of distillation to get there. From your perspective, how do we make sure that AI speeds up research without flattening the nuance that is where those insights live?
Sarah - 00:21:12:
I think it's an ongoing, I was just gonna say battle, but I think it's just an ongoing process. And that's a great question. I think I was mentioning the guardrails that have come into place over the last 18 months. And then as people get more acquainted with the different ways that we can test the same data or run the same data through different models and ensure that our results are lining, or all the models are wrong, that can happen too. But in general, you know, having those steps built into the process, into the checklist, also really helps us get to a place where the insides are less likely to get flattened, but that does slow down acceleration. I mean, it's still a million hubs faster to run, you know, our data. We have Claude Code, for example, at Amazon, we have a big relationship with Anthropic, and then we have our own models built on Claude and other LLMs. And we have different tools we can use. So, we're lucky to have different models at our disposal. And for us to go through the stops means it might take a few more days to finish an analysis because we have to really stop through the process. And that human part of stopping through the process is writing the prompts, it's really thinking through how you're going to check the data, it's not something you can offload to the AI to do itself, and having the AI check itself, you know, it's a little scary to think it can. So, I think that, you know, having those guardrails and stops in place is really critical to ensuring that we get to a place where we can have insights that aren't flattened. But then the other side of it is taking the time, the thinking time, and that, you might have all the raw materials accelerated, you might have the analysis and processing accelerated from AI, but you still are gonna need real thinking time. And that, I haven't seen get accelerated by AI because that's the human side of it.
Stephanie - 00:22:56:
Right.
Sarah - 00:22:57:
Setting the other way to keep the insights insightful is to make sure you have humans who are not only looking over what the AI gives you back, but really taking it to the next level of, I'm gonna sit with this for a day or two or three, whatever it is. I'm gonna read it all through. I'm gonna go off into my thinking cave or whatever. Just let it settle, and then I'm gonna come back and re-read it. And just like it would with any report I'm writing or any insights I'm putting together, I'm gonna give it space to breathe. Without that breathing room, though, it definitely can get flattened. And I think you have to weigh that also with, like, is it a small set of questions that just needs quick answers? Turn it around. It doesn't need the cave. Take your time. But in a lot of the work that my team does, especially, we aren't doing those kinds of small questions as much anymore because it's so easy to get this period.
Stephanie - 00:23:43:
Well, that tactical stuff you're doing. Yeah. The big strategic messy work. Yeah.
Sarah - 00:23:47:
Yeah. And I think what's interesting too, if I can add one more, just piece of this, is that in talking to, like, newer folks who are entering the industry or entering research and insights or even UX design or students in graduate programs or undergraduate who are thinking about it and and asking these big questions like, “Well, why should I even work on this craft if AI is gonna do it?” Right? Then this is your question. It's a fair question. And I've just done my best to emphasize. I'm saying this as an experienced professional who's come up through having to do it the old way, the hard way, and still having accelerated tools in front of me the whole time, but nothing like there is today with GenAI. And I would still say you have to learn the craft. You have to learn how to think. You have to learn how to process. No tool, no system, no computer can offer that from you. And that comes with experience. That comes with making mistakes. That comes with letting yourself experiment and make mistakes and being wrong sometimes, or coming back with half-baked insights and having another researcher go, “Hmmm, I don't see that.” You know? And that process has to stay the same. I can't see us changing that part of the thinking process to get to good insights. But there are calls to be made as to, like, when do you do that, and when do you just kind of go with the faster output. And that also comes towards variance, too.
Stephanie - 00:25:02:
Truly. Yeah. Great answer.
Molly - 00:25:04:
I wanna switch gears a little bit and talk about something that can be an elephant in the room at times with AI, which is privacy. And you've worked in privacy research at an enormous scale, including work connected to Meta's privacy redesign. So, how has that experience and your background shaped the way that you think about responsible and ethical AI adoption in research as it relates to privacy?
Sarah - 00:25:29:
Great question, and it's one that I definitely feel uneasy about, to be really honest, especially with my background. Not that I feel uneasy about taking processed data and running it through tools, if it's anonymized and aggregated, and we, like many big companies, have extremely strong guardrails around PII and customer privacy and things like that. But that said, with research data, it's a little different because we are going out and collecting new information, even if it's not linked to their Amazon account or, you know, showing their order history or something like that or their watch history. We are still getting faces and voices and all of the things that should be protected. And I think there is a lot to be answered about how these models are actually protecting information, whether it's corporate information and IP, or private information on behalf of all your video uploads of an interview set that you did. I do think the answers are murky at best. I mean, I think part of it, I have confidence that we have a partnership with Anthropic, I know they have very strong kind of guardrails and ethical statements they've made, which is more comforting. But, again, as a privacy researcher, I don't see that we're even close to guaranteeing anything like that. And then on the other side of it, because we do research around AI with our customers, we do know that privacy is a very, very serious concern. It's come back through a number of different studies. I can't go into the details. I don't have clearance to share all the information, but just know that it's a hot topic again for a lot of consumers who don't, and they know there's this AI thing or they might use, like, an AI, you know, native app on their phone or even on the desktop for personal use where they might use it at work and have something like ChatGPT in a corporate situation. All of that, you know, is true, but their concerns about privacy are way escalated again. And a lot of those privacy concerns had really kind of died off in the social media edge, where they were just kind of accepting that in the world of social media, where you get free services for giving up data about yourself, that was just the value exchange. And a lot of people, especially in cultures where social media has been around for quite some time, have been very kind of understanding or very accepting of that value exchange. Now, we're seeing that that's changing. We're seeing a lot more consumer concern come through about AI and privacy. And it's very interesting to see change, and again, because, you know, 5, 6 years ago, I was ensconced in privacy from a different angle looking at it, going, “Oh, interesting. I wonder what's happening with people's mental models around this.” And as a researcher who cares about maintaining their privacy, but also who cares about how this is gonna go in the world, I have a lot of questions that haven't been answered yet.
Stephanie - 00:28:15:
I guess it's super interesting because the value exchange is still there, right? Like, with AI, it's very tangible, but you're saying that it's not alleviating it in the same way that it did with social media. Do you think that's a time thing, or do you think it's a regulation thing? Like, what do you think? A mix of all of it?
Sarah - 00:28:33:
I have hypotheses about it being both the time thing and a regulation thing, and then, you know, I think it's a usage pattern thing. I think there's a lot of components to it, but I do think that at least what we're seeing is that there's definitely recognition that this is different. And I think coming out of, again, having a social media background for privacy, a lot of, and this is out there in the world, and I've written about it. But there are people's mental models around privacy pre-GenAI, especially social media, where I get to choose if I'm gonna, I'm logging on to a platform where I'm joining, you know, Facebook or Snapchat or whatever for TikTok, and I'm gonna decide what I share. And if I don't share, then I don't share, right? And if I do share, I am deciding who can see it, or it's public. There's still a lot more to it than that, but how this begins with data inference and how all of the advertising models in the world work, etc, etc. But that said, people had a sense of control, locus of control, and they could opt out. GenAI is different because it's not, first of all, they're not opting in to share, they are in a private container with the tool, with their, you know, you're definitely not sharing, I mean, you might share something you do on GenAI with someone else, but it's very much a one-to-one relationship when you interact with that unless you're in a situation like an Alexa device. And I'm not, by the way, not shilling for Alexa though. We know she does a good job. Alexa is in the household, and there's a lot of trust because Alexa's been in households for a long time, and it's in, like, this public status, like, more than 50% of U.S households have Alexa devices, and they're sharing, typically, right? So, that kind of interaction, it sends as a group one. But when you get into this individual usage of, like, a Gen AI tool, it's expected that just you are interacting with that privately, and you're not choosing to share photos. You know? You're not choosing videos, and you're not in that model. So, I think they're inherently different, and people don't know what that black box of generative AI is doing. There's not that strong layperson understanding of what an LLM is yet. And maybe over time, that will change, but it's still a big black box. But they don't know if their data's getting munched in, to use the scientific word, with everyone else's. There's just a lot of questions that I, too, as a consumer, have, you know, taking off my researcher hat. So, as a researcher, you know, keeping a close eye on attitudes and sentiments and beliefs and understanding among the models is critical, but it's also really interesting to, like, wonder, is it gonna change, and how quickly would it change? Like, where would it go when people do have more exposure? And then how much more clamor will there be for regulation, you know, when that happens, right? All the things that could come out of it naturally. So, sorry, I think I just bumped my desk. So, it's an interim topic that could go on for the rest of the hour, but yeah.
Stephanie - 00:31:18:
Totally. Yeah. Well, and I mean, I guess I just wanna keep talking about this because the other thing that I think of too is that with social media, like, we felt like we understood the boundaries of what it was. And, like, with AI, it's like, and it ends in the singularity. Right? So, it’s like between from here to wherever it ends up feels very unknown to people, and that's quite scary, right?
Sarah - 00:31:38:
Yeah. It does appear infinite, to some degree. It is. Yeah. Exactly.
Stephanie - 00:31:44:
So, well, to pull us back from that kind of doom and gloom, you were talking a little bit earlier about, you know, research. Like the skills that researchers still need to have to exercise and do their craft, right? Like, to build that, what I think of is that skill of discernment. You still have to go through all the learning that a researcher has to know. I'm curious, what else do you think, like, as AI becomes more prominent as part of the work life of a researcher, how are research teams being evaluated differently inside enterprise organizations? What other skills do they need to have?
Sarah - 00:32:22:
I would say it's a great question. I think I can speak from experience at my company and from talking with some of our colleagues in the industry at larger companies. I think there is that acceleration to insight that we talked about a little bit earlier. So, yeah, can you do more with the same amount? Can you do more faster? Can you get the same insights in two weeks instead of four? These are the kinds of things that are being looked at. I wouldn't say we're necessarily being measured on that officially yet. Like, our performance review season, sorry, guys, when it's upon us, we're not gonna ask, did you accelerate by two more weeks? You know? Like, it's not that precise. But I think there's an understanding of, like, you can say that you have been able to increase impactful output and delivered more of that in this way or that way, and AI is embedded in the ability to deliver that, right? And so, as researchers, I would say we're always understaffed. I think we're always stretched as comp, you know, I don't know anybody who comes on a team that says, “Yes. We have enough people on it.”
Stephanie - 00:33:23:
Totally. Yeah.
Sarah - 00:33:24:
Right? But we are able to, like, extend our reach and extend our impact with AI tools. And so we're able to make that case and then come back and explain, “Okay. How did we do with that?” Right? So, what matters to my leadership when I come back and say, “Okay, here's what we spent this year on vendors and AI tools and stuff, here's what we got from that, here are the areas that impacted that were critical for business and customers, and this is how it all baked in together. And that's what they're measuring.” And so, as researchers, we're still doing the main job that we've done. It's just a matter of now there's another piece of the story as to how we got there, and we may be going deeper, faster, better with the insights, but we're still expected to come back with impact. Right? It's still measured on, well, great, how does this help the business? How does this help our customers? And that hasn't changed. So for us, it's really just showing where AI fits into the process and then, ideally, delivering bigger, broader insights and impact. But, again, it depends on what you're measuring. And for us, it depends on the business question that we're solving, too. So, that, you know, still varies.
Molly - 00:34:23:
Well, Sarah, you've said a couple of times throughout our conversation that AI is not necessarily something that's gonna replace researchers, but it's something instead that's going to enhance the human insight. In practice, what does that strong human-AI collaboration actually look like when you're doing the research?
Sarah - 00:34:42:
I think human-AI collaboration, there's a little more that I can say. I think one of the big topics that we've come into more recently on my team is quality control and understanding, like, where, so, and I'll say this with regard to using bespoke tools that have our, like, AI-first or AI-forward platforms where we have vendors. So, they're outside of Amazon. We're not using our direct tools to do this. But we're going to them and saying, “Okay. We're running this study. You guys are gonna come back with x, y, and z for us for this study. It's on your platform. We can play around with it if we want to, or you guys can deliver a report.” And there are AIs involved in pieces of that process, and it's doing, you know, quality control on respondents in the study. They're using it to perhaps put together a study plan to generate a survey instrument to put together an analysis of the raw data. And there's different phases along the way where AI gets deployed on their side, and we don't have insight into it. And it's not that they wouldn't tell us. It's not a secret. But in the olden days, you know, a few years ago, three years ago, if you went to an agency and said, “Help us with this big study, and here's the brief or here's the plan, and here's what we need. Here's the questions we need.” There's the deadlines, and everybody agrees to it. And they go off and do their thing, and you check in periodically and review milestone work and, you know, standard model, right? You don't really ask a lot of questions about what they're doing off on their end because you trust their experts, and they're gonna go do their thing, right? And then you get your results, and you might ask questions about that. But at the end of the day, you're not there with them on every step of the process because that's why you hired them anyway, right? But with AI, it was kind of weird because we have a lot of vendors that have a human in the loop. I think that's a phrase that's become pretty common. And we're, like, which part of the loop is the human in? Right? And I'll take this the way, but we see some loopy stuff come back, no pun intended, like, from AI results. We're like, was that you guys or was that the AI? Like, what happened there? And so we've had several interesting, I'll call them, feedback conversations with some vendors recently where they've been working to put more AI or offload some of the human side of human in the loop to more AI on their end, and we see the results instantly. And I don't mean in a good way. Like, we're seeing results come back. Like, did the AI write that? You guys don't usually write like that. You know? Like, it's a very not a fun conversation to have, but it's, I mean, it's part of the partnership. It's part of the iteration. And so the collaboration aspect of it, which I think is really a healthy one, is that we as researchers go to our vendors and researchers, and we're all like, “Hey, look. Okay, look, that's AI slop. Come on, guys. That writing is classic, you know, negative parallelism.” You know? We don't write like that. You know? Or it might be something where the insight's just not that crisp, or we're like, yeah, we missed a piece of the, you know, there was a pretty key objective in our brief, and we're only getting half an answer. And we know we know we collected the data, so let's go back and do it again. And again, it's because AI is doing some of the lifting for them, and so then they have to put more human in the loop. But our collaboration is saying, like, we're gonna really go over this work, and we're gonna ask the questions. We're gonna ask the hard questions, and we're gonna give the feedback because we all wanna get better at it. And if they're training their models to do certain things for bespoke research work, we're there to help train. I mean, I wouldn't say we're volunteering to help train, but we are. Like, the fact we are. And that's a lot of the collaboration that we're doing is really us going to vendors who are there to help us that we get a lot of value out of. But we're also saying, this is still bumpy. And we're gonna tell you when it's bumpy, and you know, we're glad you're willing to receive it and fix it. That's been a lot.
Stephanie - 00:38:21:
That’s super valuable for your vendors, I'm sure, for you to be able to do that and have that in the spirit of collaboration.
Sarah - 00:38:27:
Yeah. I think so. They say so. So, but it makes it, you know, you've gotta be willing to do that too. And I think the idea of having a turnkey vendor that you can just turn over the whole study to, that uses AI, doesn't exist yet. It's not, you know, it's not there yet. And if it ever will be, I'll be impressed. But I think you're always gonna have to be involved in collaborating, and you know, whether the vendor is willing to do the collaboration or not is a different issue. But I think if you're signing up to do this, you have to be willing to give the feedback and recognize the imperfections, or the, I shouldn't, it's not even perfect, but recognize when something can go off a cliff. Like, we had one vendor who I'd absolutely think is a great vendor, and we love working with them. They made some major changes to their system, and what it produced was crazy bad. And we're like, oh, that is not, is it turning a knob, you know? And I’ve worked in some rally, so I'm like, we see stuff go wrong all the time. Somebody pushed a new build. You know? Like, you know, it's kind of like you see the fire and the rain. You're like, but I had to be like, what the heck happened, guys? No. Really. And they're like, oh, that was, I’m like, don't give me the marketing answer. Please just give me anything. Go back and look at what happened. Somebody pushed the wrong, you know, like, that kind of thing, and we got it started. But I had to be willing to be like, come on, guys. You know? But also recognizing, like, this is not a fatal error. It's just we're all growing up together in this. And in a lot of ways, I feel like every day is a new day in a way that I haven't felt for a while, because when you've been around for a while and professionally, you're like, and I've seen it all. Yeah. There's stuff where I'm like, I have no idea what's happening here. I feel like I'm, you know, I imagine there are people getting right out of graduate programs or school that know more about what's happening and why that went wrong. Like, it's that level of, like, you know, I'm getting educated too. So, the beginner's mind is really essential more than ever.
Stephanie - 00:40:21:
Yeah. I like that.
Molly - 00:40:23:
Well, this is awesome, looking at the current state of things, but let's look ahead. Obviously, the rate of change is crazy when it comes to AI. But from your view, looking at the next 12 to 18 months, which part of the enterprise workflow do you think AI is going to completely change the most?
Sarah - 00:40:41:
I think, again, speaking from my team's perspective and also from a lot of collaborative conversations at my organization, you know, we don't, you know, we see a couple of things happening. One is that our internal tool suite and, again, we're at a company that makes AI products, so frame this as you will. But our internal tools are accelerating what they can do for us. And so as researchers, we have attention constantly between, do we just get raw data from our vendor, and we run it through our own tools and our own systems, our own templates just out of the gate? Like, we know we've got them. We know how to use them. They're getting better every day. You know, we have these. Or can we continue to look at bespoke, you know, third-party options for some of the deliverables that we otherwise would do ourselves? Because it is obviously a lot of work, whether you're getting a written report from a vendor and you're reviewing it, or you're writing it yourself with raw data that you got from a study you ran on their platform, the latter takes a lot of work, but they both do. And so the question is, which is more efficient for me in the long run as a researcher in the enterprise? Is it better to just optimize the end output, that last mile of deliverable with our tooling as it continues to grow, as we continue to, or we're continually encouraged to integrate in all aspects of our job lives, or should we look to our external vendors to do more of that, or should they become more specialized in what they do? Like, are our enterprise vendors who do AI-forward stuff really good at rounding out our research team, giving us that extra brain, right, for thinking, for finding insights, for designing methods, for designing instruments to collect the information and data that we need? They're good at optimizing that. They're good at doing aspects of that flow. And so we are now in a very interlinked workflow versus a throw it over the proverbial wall and have it come back, or have it completely automated because we just ran a full self-service. But either way, we have to bring the systems together. And that's one area that I think over the next 12 to 18 months, we're gonna be have we're gonna be figuring out a lot more of, or we're gonna see a lot more experimentation. And we're also seeing our vendors change, like what they're offering, too, as we collaborate and what they see in the market. It's no longer about just saying, “Hey, we can do self-moderated interviews with an avatar”, or you know, like, there's a lot of checklist items that a lot of vendors that are new to the market have, um, you know, all kinds of indexing on. Right? We're like, no, no, the value add isn't that. We got some of those tools where the value add is, can you give us the research brain help and help us get the data and do it in an expedient way at reach? And then we can figure out who mushes it together and makes the insights, right? So, that's one area. I think another area around, you know, enterprise, definitely, enterprise research is just controlling the data. So, as you can probably imagine, with any company, we have a lot of guidelines and requirements around the data being stored with us, right? We have dispensation for our research tools and have for a long time for data to be stored with them. There's a very rigorous Infosec process that has to happen for third-party systems, etc, like a lot of companies. But at the end of the day, you know, being in the older model with legacy tools, we were just collecting data, running a study. It lived on their platform. The platform was secure. IT signed off on it before we used it. Everybody's happy. Now, we're getting a lot of inferential data. We have vendors outside of our walls that do meta-analysis for us that can look across our studies, that can bring things in together. We can even do it cross-team, cross-workspace. But we now meld a lot of what we do for research with our internal brands, right? We have our own LLMs and our own tools that have a lot of context, and we can't make that context available. So, if I went to tool X, who's done five studies for me on a certain theme, I can do a meta-analysis now on their platform. Cost of five studies. Great. But I still have to web that to contacts, which I can never give them. So, one big question talking about with some of our partner vendors is how do we bring what you do into our world? Because we need to be able to bring your great meta-analysis into our context, the less manually, the better. We can't give you the context, but we can bring it into ours. And then if you're offering stuff like, “hey, we're gonna help, you know, we're gonna develop tools that let you create dynamic personas out of the data that you've collected and the studies you've run, we wanna bring those into our world because we know our stakeholders aren't gonna go to yet another system.” They wanna see everything into this increasingly robust set of tools that we have. So, there's a lot of conversations we're having, too, about how do we cross, how do we connect, right? Literally connect, and how do we bring in the right data? How do we do it securely? How do we bring it in so that it's value-added? And how do we bring in what they do that's value added? Then we're not gonna build those tools ourselves. I'm not gonna be programming a dynamic personas tool for anytime soon. I don't have that kind of time or talent, and it's not just a basic chat agent. Right? It's a bespoke tool, but I have vendors who can do a great job or have already, and I wanna bring what they have into our world. So, a lot of those questions are also on the table for the next 12 to 18 months.
Molly - 00:45:44:
What really stayed with me from this conversation is the reminder that AI is not valuable simply because it's fast or it does things easier; it's truly valuable when it helps researchers think better, connect ideas faster, and spend more time on the human side of insight instead of the more manual things that just eat up time.
Stephanie - 00:46:04:
Yeah. Yeah. I like that, Molly. It speaks to how much, like, intentionality it takes to implement AI well inside of a large organization. The tech itself is only part of the challenge. The workflows, the expectations, and the trust around it matter just as much.
Molly - 00:46:22:
Right. And too, this idea that the future of research is probably not a human versus AI thing, as Sarah mentioned. I think that's, you know, sort of going to the wayside, but it's rather the human with AI. What does that relationship look like, that working relationship? And it's the teams that figure out that balance thoughtfully that are likely going to be the ones in research, and any industry beyond that's gonna create the most impact.
Stephanie - 00:46:46:
For sure. Sarah, thank you so much for sharing your perspective on where AI is genuinely helping, where caution is still needed, and where you see the industry heading.
Sarah - 00:46:56:
Well, thank you, guys, for having me. It's been my pleasure. I've really enjoyed talking to you both, and I’m just happy to be able to share some of the perspectives and experiences. So, thanks again. My pleasure.
Molly - 00:47:05:
And to everyone listening, thank you for being part of The Curiosity Current. We'll see you next time.
Stephanie - 00:47:12:
The Curiosity Current is brought to you by aytm. To find out how aytm helps brands connect with consumers and bring insights to life, visit aytm.com. And to make sure you never miss an episode, subscribe to The Curiosity Current on Apple, Spotify, YouTube, or wherever you get your podcasts. Thanks for joining us, and we'll see you next time.
.jpeg)
