Unsmoothed: The Beauty of Friction in AI | Dreamforce 2026

Hi, everybody. Good afternoon. Thank you for joining me as I give you a talk about the research that I've been doing at MIT and from my upcoming book, Unsmoothed. So when I'm not here in San Francisco, I spend most of my time researching the intersection of behavioral science, that is how people make decisions, artificial intelligence, which is a catch-all phrase for everything from machine learning to agents to GenAI, and how that intersection affects human decision-making and performance and what that means for strategy, for how we work, how we live, and the humans who are at the center of this. Now, I've been researching this for about 10 years. And there's so much going on, I literally never give the same talk twice. So I want to talk to you today about something I've actually never talked about before, which is some of the new ways that AI is smoothing out some of the frictions in human life. For example, now there are thousands, potentially at least a million AI immortality bots. of people who have passed on. The digital AI-generated afterlife industry is expected to really explode into $5 billion by the end of the year. And the objective for many of these firms and these industries is to make grief obsolete, to smooth it out. Now, you all know that I'm a researcher at MIT. and I told you what I focus on, but what you may not know is that I am living in loss. Last year, my husband passed away from colorectal cancer, and I would do anything to bring him back. And an AI bot, well, I have enough data and content to train it such that I could hear my husband's voice, I could see his image, and I could make him say things That he probably wouldn't say in real life, but that I'd like to have him say. But I'm not going to do that. And I want to tell you why the friction of grief and loss has added a new dimension to my research and why friction is not something we should always seek to smooth out. So when we think about friction, I often refer to it as the F-word, In managerial circles, the word friction is usually a negative thing, something to be removed. But my research shows that friction is something that we actually need to change our perspective on. And sometimes, particularly when AI is involved in systems, we need to seek beneficial friction. So I'm going to give you 3 things for you to take away today for your own playbook, not just for your work, but maybe even for your life. The first thing is to think systematically about friction. We often talk about friction as being tantamount to a pain point, but in actuality, we need to think about things dynamically within a system. So back in the 1950s, the field of system dynamics was invented at MIT. And what this field said was that in any system, we have these interconnected parts. And what we change at one touchpoint in the system can have a positive feedback loop or a negative feedback loop. But it won't show up at the touchpoint that we make the change. It'll show up somewhere down the line in the system. What I want to say to you then from a system dynamics point of view is when you use AI to remove friction at any touchpoint in a system, You should be thinking about where else in the system the results will show up and where could there be negative feedback loops that are unintended and where could there be positive feedback loops and where could we use friction to our benefit. One way to think about this is in this image here, you see a runner and you see the formula for friction. And I want us to think about friction the way in which physics talks about friction. Friction is neither good nor bad. It is a force. It is a force to be harnessed. Now, what if in this runner, this system of the runner and the ground and the air, we were to remove the friction? What would happen to that runner? Well, they'd probably try and run and slide all over the place, right? And lose control. Turns out friction plays a really important role here in this system. Without it, we lose 2 important things: the ability to accelerate— the runner can't accelerate without friction— and the ability to change direction. So I want you to look at the systems in which AI is working with humans and identify places where when we remove friction, Where else could we take care to add friction to stabilize the system? Now, what do I mean by friction? Well, there are 3 forms of friction that I want you to pay attention to. The first is the kind of friction that people talk about all the time and that I agree should be eradicated. That is wasteful friction. Effort that is not producing value Questions that have already been sorted and are rote, we don't need to waste effort and energy on that. Wasteful friction reduces value and it should be removed. But there are 2 other forms of friction that are beneficial that I want you to pay attention to when you're using AI. The first form of beneficial friction is load-bearing friction. That's right. If in this building I were to a load-bearing wall, well, the building probably wouldn't fall apart immediately. But if there were any tension, any strong wind, any rumble, that's exactly where the weakness would show up. So when you are removing friction with AI, think to yourself, where should I be adding load-bearing friction to reduce risk at the same time? The second form of beneficial friction is generative friction. That's creating something that didn't exist before. That is the connections that you're making here at Dreamforce. That is something that you experiment with and try. That is listening to your customers and having them come up with ideas. All of that is not frictionless, but it does generate value through the effort. When we think about systems this way and we think about 3 forms of friction, we're thinking about using agents to reduce wasteful friction, but simultaneously taking care to use the other 2 forms of the friction force to reduce risk and to increase value. So the second thing I want to leave you with is embracing the friction of Testing and learning to challenge your assumptions. I'm going to show you an example. Imagine you're trying to get to MIT. So this is where I work. And imagine you're driving and you're going to choose a route with a GPS app. So this is actually my route. You can see 3 options right here, options 1, 2, and 3. This is my real commute to work. So yes, Boston traffic is pretty bad, although I understand San Francisco can probably go toe-to-toe. So let me see, which one would you choose? Vote once. If you would choose option 1, please raise your hand. All right. Look around, see who would choose 1. All right. If you would choose option 2, please raise your hand. Pretty equal. And if you would choose option 3, please raise your hand. I saw someone Sort of. Sort of move, but no. Okay. Well, there's no wrong answer to this question. But what I want to use is this example to show you some of the research we've been doing, which is, after seeing how someone made a choice, their logic flow, their chain of thought, which of the people in the room would be more persuasive to you for a subsequent choice? In our experiments, what we looked at was, could there be, with agentic AI, a similar attraction as we see from human to human, such that agentic AI that reflects choices like you are ones that you're most likely to trust? And so we ran an experiment. In this experiment, people were randomly assigned to use agents in 2 conditions. In one condition, the logic flow of the agent was similar to theirs, a digital twin from a logic perspective. In the other condition, the logic flow of the AI agent was dissimilar to them. Not wrong, just a different chain of thought. And then the question we had is, which one would people be more likely to trust? Which one would they delegate more choices to? down the line. Now, interestingly, when we think about hiring people to our team, we tend to say it is optimal for us to find people who complement us and the way in which we think. Is that the way— the same way that people behave with agents? Turns out, no. When it comes to trust in AI, we see that to the degree that the AI agent Logic flow. Now remember, we're controlling for accuracy here. This is not a difference in right or wrong. This is simply in the path, the logic flow, the chain of thought that the AI agent follows and whether it is the same as yours. In the orange bar, we see the control condition where you do not see the logic flow of the AI agent. In the pink bar, we see an AI agent where you see the logic flow, but it is different. And in the green bar, we see an AI agent where you are shown that it has a similar logic flow to you. And you can see in the green bar, we have a significantly higher level of delegation for the task by treatment. Moreover, in the green bar, people think that it is more reasonable to follow that AI agent's logic flow. That AI agent thinks like me, so it makes more sense. Moreover, they describe the task as cognitively less effortful when they're using an AI agent that has a similar logic flow. So it turns out that the process is one where we trust AI agents that mimic our logic. And so we may want to think about adding a little bit of friction to when people use AI agents to understand the similarity or the dissimilarity and to intentionally in the system interrupt over-anchoring on an AI agent with a similar logic flow. Because the similar logic flow makes people think that the suggestions are more reasonable, which makes them expend less cognitive effort and makes them trust. Or potentially overtrust. Another example of beneficial friction is experimentation. The data are pretty clear that firms with more experimentation are more successful. In fact, 30% higher productivity gains than those with not. And then of course there's a negative feedback loop such that when experimentation is not invested in, The firm has more problems in terms of performance gains. Again, thinking about friction not as something to be removed uncritically, but understanding that in the case of testing, learning, failing, and experimentation, it's not frictionless, but it is effort that results in value. All right. I want to leave you with one more thing. Let's have a look at this. We see that 81% of professionals trust agentic AI with their customers' data. And we can see that there are some examples where there are indeed improvements. But what do customers think? Well, we see that only 36% of customers trust agentic AI with their personal information, and that trust actually fell in the past year. So how do we resolve this tension, this difference? My colleagues and I have done research on this, and I want to tell you that you should be thinking about 5 forms of value that you deliver to your customers, and some of them involve some friction. The first form of value is economic capital, and that's where everybody thinks. Give them a coupon code, give them a discount. Give them something free, and I'll tell you nobody's going to say no. But the problem is this is actually a frictionless form of value. It's easy for your competitor to replicate. So you want to go to a higher level of value. The next form of value we looked at is cultural capital. This is ritual. This is your principles. This is your lifestyle. This is your mission. And it turns out when your customers are aligned with regard to receiving cultural capital and involved in investing in cultural capital, then they perceive more value from their experience with the firm. Next, we have social capital. That's connections with other people, connections with other firms, connections with other folks in the field. Turns out that when you provide your customers with that, they receive value in return. Next form of value is information capital. You've made me a smarter customer because you've given me important information, but also you treat my information with respect and care. And last but not least, temporal capital. That is, you help me get to my goal in a timely and efficient manner, but also you anticipate my needs. What we find is that these 5 forms of value create a value exchange, and they're mutually beneficial such that the more forms of value your customers have, the more they believe they have agency in their customer experience, and the more likely they are to trust and have loyalty with the firm. When you are looking at using agents, think about places where you are creating value, And be careful about removing friction in places that actually would remove value. Because it turns out that humans are really good at helping with cultural capital and social capital and even information capital. And we don't want to remove the friction in a way that removes the value that your customers would receive in these different 5 ways. So the third is about value. And when we think about all of these 3 taken together, I think about how people talk about AI. I have an opportunity to talk with thousands of leaders every year in different kinds of industries. And a phrase I keep hearing over and over again, I thought maybe you have heard this phrase Frequently too. So if you look at this slide, what do you see that all of these have in common? What phrase, what buzzword keeps coming out? AI first. AI first. And I guess my question is, if AI is first, then who's second? The humans. The humans? Your customers? Your employees? So when we think about incorporating AI into our workflows, into our systems, we have to take care to not smooth out the human element. AI without providing value to the humans is a waste. You will not receive the ROI. But AI that is human-first thinks first about how do humans think, how can we create value, what do they value, and how can we not just build AI and agents that people trust but that are trustworthy. So my 3 takeaways for you today: polish comes from friction. We do have a tendency to avoid friction. When I think about myself in my grief process, would I like to fast-forward to the part where things don't hurt so much? Absolutely. But I think that would make me a little bit like in The Odyssey, the Lotus Eaters, who are detached from reality and don't grow and don't develop. and miss out on the important parts of the journey that we have as human beings. Our purpose on this planet is not to be seamless, to do what's easiest. It is to grow, and sometimes that comes with challenge. So remember, polish comes from friction, not from the removal of it. And these 3 takeaways I have for you is to to ask which forms of friction, those 3 forms of friction that I mentioned to you, wasteful friction, load-bearing friction, and generative friction, could you be removing either purposely or inadvertently when you incorporate AI into your system? Beware smoothing out the beneficial friction that is load-bearing and generative. Secondly, testing and learning and being less comfortable is part of this process. If you are not creating a culture of experimentation, and that's not easy to do, I challenge you to think about how you can lean into that and go through the process and the friction of creating the culture of experimentation because from what I've seen, both from my research and in other leaders, that's the secret sauce of innovation. And third, AI has some incredible potential, but delivering value is ultimately what its purpose is. And so sometimes that means the friction of being human first. Thank you all for your attention and for being here. It's been a pleasure. Enjoy the rest of Dreamforce. Thank you.