a16z's Anish Acharya: personal AI agents cost $1,000 per user per year today — but deflation is coming fast

Speaker 1: Have a good one. Sleep. We'll talk to you soon. Let me tell you about Figma. Agents meet the canvas. Your AI agents can now create and modify your Figma files with design system context. Our next guest is on the line, Anish from Anish, the general partner coming back on the show here with us to discuss All Things AI Personal Agents. How you doing, Anish?

Speaker 10: Nothing. That's weird. They got the squawking.

Speaker 1: What was that?

Speaker 2: Oh, that's the eagle. The eagle squawking. Your audio maybe get a little closer to your mic. The you sounded

Speaker 1: a

Speaker 2: little Can you

Speaker 1: do a little sound check for us?

Speaker 10: Test. Test.

Speaker 2: Test. Okay. You. Can

Speaker 1: Personal AI agents. The biggest debate point was, it going to destroy the restaurant industry? Is it going to gum up the restaurant industry? If you run a restaurant, is it the worst time or the best time? Are you going to be flooded with thousands of phone calls every day trying to book special occasions every five seconds? Or is there a white bill here? Is there a good solution where the the the humble restaurant tour sees a modest increase to the

Speaker 2: bottom line? Our debate. We I wanna know

Speaker 1: where you land.

Speaker 10: I think this would be great, man. I mean, look, I think the last gen of some aggregators were maybe not amazing for the restaurant industry. Mhmm. But if you think about the GM of Carbon, what does she want? You know, she wants to have the highest ARB people in the restaurant with the highest frequency and have a high fill rate and have some sort of entropy so new people can try the restaurant who might be future high spenders. Mhmm. I think your ability to fill the restaurant in a way that sort of globally alkamol is way higher. It's not just who clicked the website fast enough. That's one. And then two, I think for the long tail restaurateur, you know, they don't have to be amazing at marketing game or the kind of, you know, sort of marketing to the intermediary suppliers like Resy. So I actually think you're gonna get this like long tail effect where more people will visit more restaurants and for the restaurants that are in the part of the power law, they should actually get the best customer. I think it's very, very good.

Speaker 2: Yeah. Let's maybe start with Thanks. I'm curious how you think that the the tension right now between users who have personal agents that want them to be able to do anything are running into the wall on places like amazon.com. We saw somebody today that that got blocked from just buying something on adidas.com with with Muse, which again feels like a sort of casualty, like, that that that, you know, something They have a system in place to block more nefarious style bots, maybe like maybe it's sneaker bots or something like that. Who knows? Right? But it it feels like inevitable that users will be able to get their agents to operate on their behalf everywhere in the fullness of time. But the question is like, how long do we have that sort of tension between I think it's really just gonna come down. I expect the Adidas' of the world to adapt, but it's the it's the ad plat places that are monetizing with ads. How quickly will they fall or figure out alternative solutions?

Speaker 10: I think pretty quickly. I mean, we'll have to see how it plays out. I like what Noah said on the call this morning, which is there's gonna be a portion of your revenue that is driven by attention and a portion of your revenue that's driven by fulfilling transactions. Mhmm. And then every, you know, ecommerce retailer, they're gonna know, like, you know, Amazon is disproportionately advertising perhaps or high margin advertising. They're gonna wanna focus on protecting being the destination. Whereas, Adidas, if they're more of a sort of transaction fulfillment business versus a destination car sales business, I think they wanna quickly get the intermediaries out of the way. I think the distance is gonna be a lot of studios wants wanna think of themselves as destination businesses, but in reality, they're sort of transaction fulfillment businesses. Yeah.

Speaker 2: Yeah. Yeah.

Speaker 1: Are you getting pitches for we solve the problem created by Instinct and Muse yet? Because there in one way, there's already been, you know, business phone trees and transcription services and things that turn text message inbound into CRM stuff. There's AI customer support companies. But at the same time, like, this does feel like a slightly new twist, a slightly new tool relevant. But how are you dealing with that question in the era of personal agents, like the secondary businesses or maybe the secondary businesses are just instinct for business, muse for business, you know, the other side

Speaker 2: of it?

Speaker 10: We haven't seen as many yet. I mean, the question is what's the most extreme form of an adversarial relationship with your customer? And there's actually a terminal story about Brazil, which has the most lawsuits per capita of any country. And the reason for that is you can sue anyone digitally and sort of trivially. What happens is when, you know, your DoorDash is late, you just fire off a lawsuit like one. That's amazing.

Speaker 3: That is

Speaker 10: a very adversarial relationship and the merchants need to buy specialized software to fend out all the lawsuits. I think things like that will happen. You know, demand for refunds and protesting payments and all of these areas where merchants benefit from kind of customer apathy or being fully informed.

Speaker 2: Yeah.

Speaker 10: We haven't hit yet.

Speaker 1: I like this future. Like, if I you know, you asked me to order you a coffee the other day, you wanted regular milk instead of almond milk. I got you the almond milk one. You could've sued me for like 50¢ right then.

Speaker 2: Yeah. Sue a friend. Just Instantly.

Speaker 1: Yeah. Instantly, you know.

Speaker 2: That's really yeah. That's really crazy. If some if if if some of these systems get abused Yeah. Then they have to just remove the policy at which point as a consumer, you would have to actually press, like take like a legal Like some type of legal action to get the thing that used to be offered out of the box. Mhmm. Like that's kind of like the the darker

Speaker 10: I think it's an opportunity, Jordy. Actually, take it the opposite view. Like, think of every person, certainly of their lifetime, but maybe once a year has this just bureaucratic weight placed on them. Know, maybe you move and you have to change schools for your kid. You know, God forbid, maybe you get cancer or have a DUI. Like, there's just so much bureaucratic administrative work that the entrepreneur has to do, and I don't know that they actually have equal access to the legal system. It's hard to get an attorney and it's hard to go to small claims court. So I kind of like this vision of the consumer being extraordinarily empowered with these things you used to have to have a guy for.

Speaker 2: Yeah. No, that's a good point.

Speaker 1: How are you thinking about compute intensity in the era of personal AI agents? I've seen some, like, you know, wild speculation and rumors, but we've all been through the the realization that going from LLMs to reasoning systems to agentic coding, like every one was an exponential. We started producing 10 times the tokens, 10 times the energy, 10 times the cost. Of course, 10 times the capability. So the adoption happened. But Personal AI agents, do you have a feeling for how much it costs to serve someone? Or do you have any idea for how to even think about this? Or even does it require a new mode of thinking, is it just the same old?

Speaker 10: I think right now, our best estimates are a thousand dollars per user per year, which is obviously prohibitive if you're running this I think the the couple reasons to be optimistic if you're running Well, pause

Speaker 2: pause for one second too because when you think about making that back on transaction fees, like, the average person is not gonna be spent the average person just simply doesn't spend a $100,000 a year on, like, hotels and flights and things like that. And so there there there is, like, the cost has to drop meaningfully if you wanna be able to make up

Speaker 1: Mhmm.

Speaker 2: If you wanna be able to keep the services free and make it up with like Transaction volume. Percent or 5% Yeah. Fees.

Speaker 10: Yeah. The average American casino spends 5 to $6,000 on travel plus commerce, online travel plus commerce a year.

Speaker 2: Yeah.

Speaker 10: Yeah. That's got a very high take rate. Look, the number one thing that needs to happen is cost needed to deflate and they will. Yeah. Right? We've this over and over again, like, now, has kicked ass that computer use is very expensive. It will get cheap very quickly.

Speaker 1: Yep.

Speaker 10: The actually most exciting thing I saw in the last two weeks was Jeb. Yeah. With computers, like Jeb computer use is a 100 x cheaper and faster today.

Speaker 1: Yeah.

Speaker 10: So I think probably these kind of dark horse opportunities that it could deflate cost faster. Otherwise, you're kind of fighting a two front four on customer acquisition and compute capacity.

Speaker 1: Yeah. Yeah. That makes a lot of sense. Well, thank you so much stopping by.

Speaker 2: Oh, last last question I had. We were just purely speculating earlier that it feels very natural. Like agents now have a phone number and a computer. It feels very natural for them to have like their own dedicated sort of this is the first time that I've got I've I've heard a bunch of pitches for like why agents are gonna use stable coins and this and that. And I wasn't ever fully Yeah. Sold. But this feels like like you could imagine an instinct being like, hey, we're just gonna here's a credit card that you can use. It has some of these these benefits and maybe Mhmm. And it just feels like that's an easy way to like start monetizing all of this volume Yeah. Without having to go get the opt in from every single long tail retailer and service provider that you're driving a bunch of traffic to.

Speaker 10: Mhmm. Yeah. That's that's absolutely right. Like, unfortunately, just the take rate on those things is not high enough to still assume it's a good idea. I think the more interesting question, Jordy, is what happens when agents are independent economic actors and they can go look for work and get paid for doing work on your behalf and teach them your specialized knowledge and skills? Like, the whole agent economy thing is definitely coming and it's gonna be wild.

Speaker 1: That will be wild.

Speaker 2: Yeah. Because right now there's millions of people reaching out to businesses, for example Mhmm. Offering to effectively be a meat proxy for them to just do work that the models can already do and then just give it back to them. Yeah. And then at what point, you know, I'm sure there's people out there that have just scaled up outbound to the degree that they're effectively doing this already just as a company. But at what point can you have an Yeah. Something more like an independent agent. It's fascinating. Yeah. But, anyways, great to great to catch up on

Speaker 1: this.

Speaker 2: Very Yeah. Have chat more. Parts of you and come back on soon.

Speaker 1: Thanks so much. We'll talk to you soon. Goodbye. Cheers. Let me tell you about Cisco. Critical infrastructure for the AI era. Unlock seamless real time experiences and new value with Cisco. There's a Tesla Roadster update. It's been delayed to October 15. It was supposed to be October 7. They've been tracking the weather closely, but the severe conditions predicted. And because this event can only be held outdoors, they rescheduled it. So we will have to wait until October 15 to get the update on the new Tesla Roadster. Anyway, our next guest is already with us in the waiting room. Have Emma Conrad from San Francisco Compute.