Title: Warning: Investors Must Prepare For The Agentic Takeover
Show: Joseph Carlson After Hours
Guest: Joseph Carlson
Date: 2026-SEP-18
URL: https://youtu.be/HKskNSEb5_c
Length: 43:28 (2608 s)
Note: Auto-captions cleaned — verbal fillers (uh), stutters/false starts ("I I think", "most most", "they're they're", "a a", "is is", "delegate de delegated") and a "[snorts]" noise tag removed; wording, numbers, names and every (mm:ss) cue otherwise verbatim. Caption mis-hearings are left as captioned: "Citrory/Citrony/Citroen/Citriny" = Citrini Research; "init/into it/in it/Inuit/intu" = Intuit (INTU); "Turboax" = TurboTax; "SMB Global/SM Global" = S&P Global (SPGI); "Dualingo" = Duolingo; "Whimo" = Waymo; "Smith/Smiths" = Smith's (a Kroger banner); "Sachin Adella" = Satya Nadella; "Pasi" = passkey; "Brainree" = Braintree (PayPal); "ChachiBT/chatbt/Chat GBT" = ChatGPT; "Grock" = Grok; "Asians" = agents; "OpenT/Open" = OpenTable. "My biggest holding today is Google, a $24,000 position" is as captioned (his GOOGL stake is ~$240k per earlier episodes).

(00:00) In February of this year, an article was released that shook the markets. It caused a widespread panic and it caused many stocks to sell down over 7 to 10% and they continue to go down just by the release of this article and it was a narrative about a possible future, a situation where agents have took over the world.
(00:20)  This is the article from Citrory. It was published February 22nd. Ever since then, many companies including SAS companies went through what's called the SAS apocalypse. Citrony Research was already the biggest financial substack in the world with millions of readers and this also made it even bigger. This went hyper viral. The article title was the 2028 global intelligence crisis.
(00:42)  Intelligence is abundant. We have AI that knows about everything and so everybody knows about everything. But that wasn't even their major point. Their biggest point of this article is friction going to zero. Now, how does friction go to zero? It goes to zero through agents. Agents are an extension of humans.
(00:59)  They're robots that can go out and do things that humans normally do. Humans normally go and they fill out forms. They look at user interfaces. They purchase things. They shop for things. They compare things together. All of that stuff is friction. And in a world where agents do all of that for you, it erases that friction. Now, Citroen didn't just talk about agentic technology and the potential of it.
(01:20)  They also listed specific companies and how they think it might work out. For example, they talked about Service Now decelerating. They made fake headlines like Mastercard's revenue decelerating down to 6% year-over-year growth. They said it would destroy the job market, that software companies were getting obliterated, and they highlighted Door Dash as the poster child of the type of company that would be disrupted by Agentic technology.
(01:40)  Now, while Citrony's article was very thoughtprovoking, it was well written. It was convincing at the time, it was also full of flaws. Many of these companies aren't going to be affected in the way that they've illustrated, and I'll give them the benefit of the doubt. We didn't really have agentic technology then like we do now.
(01:56)  Times changed quickly. In just a number of months, we've had technology released that is truly agentic technology. The first example of this in a big consumer way is Meta's Muse. Meta just recently released an app called Muse. I've mentioned it before, but the more that I've used this app, the more that I realize that it has a lot of these type of things in it.
(02:18)  It is reducing friction dramatically. And for those of you that think that Meta's Muse is like another ChachiBT, another Grock, another anthropics claude, it's not. It is entirely different. It feels very different. It's the first time since the very beginning of using AI where I thought it was incredible AI could code that I was wowed again.
(02:39)  I opened up the Metamuse app and it was just immediately different than ChatBT. It wasn't an empty chat box that spits out information. It was this helpful assistant that will help you accomplish anything you want to do. It said it could do things like book hotel reservations. It can book restaurants for you.
(02:57)  It can join your email and clear out your email box and unsubscribe from unwanted emails. It can organize trips. It can plan things for you. Or it can shop on your behalf. It can just shop for you. I thought, can it really book restaurants? And it said yes, it can. It can link up to Open Table or it can verify directly on the restaurant's website.
(03:17)  And that was impressive in and of itself. But it got me thinking, when I can book a restaurant through OpenTable without visiting OpenTable's app, that changes things. I'm working through an agent that's talking to OpenTable and I don't have to see OpenTable's sponsored listings. I don't have to deal with their little promotions.
(03:35)  I don't even have to have the Open Table app on my phone. It can go in and just book things on my behalf right from this text box. Another thing that was interesting is it doesn't just do restaurant reservations, but it also does the commerce in a serious way. I asked it to look up Nike shoes, very specific ones that my brother-in-law was looking for, and it searched the web far and wide.
(03:56)  It brought up its own little browser. It search through all different websites, looking at reviews, trying to find the best prices in the exact shoe match. It was like having a secretary or an executive assistant spending the next 30 minutes of their time specifically trying to accomplish that task. And then just today, I decided that I wanted to try out more of the shopping and see if it could tackle a harder task.
(04:18)  I have four kids, one of them is a newborn. So, I have three kids that were feeding normally that we have to shop at the grocery store. And for two of them, they have specific allergies like a dairy and nut allergies. And I plugged all of this in to Muse. I said, "Here's my kids' ages.
(04:32)  Here's the allergies that they have. Could you make a week plan, a meal plan for them based on this and make sure you check the allergies of each food?" said, "No problem. It'll help out with that." It recommended Smith because Smith has a lot of allergy alternative foods. They have lots of substitutes. It went on to Smith's website.
(04:48)  It made an entire meal plan for them. Added every single item. And while it did it, you can actually watch it live. And you'll notice as it goes through these, not only does it verify that it's the right item for me, it's what I want, but it also picks the one with the lowest price. It attaches any coupons to it, any way to save.
(05:04)  It'll actually search online for any applicable way to lower the price for you as well. And this is just a preview of the new agentic world that we've entered into. Muse is just the beginning. We know that chatbt, we know OpenAI is going to be coming out with their version of this very soon. Same with Anthropic and same with all the rest.
(05:21)  We'll have multiple agents to be able to extend what we want to do to be able to go out in the world and do things for us. And it's not just shopping for food or groceries or goods and services. It's also booking things, planning trips, booking hotel reservations, going on to booking holdings or Expedia may not be quite as necessary when you have an agent that can do it for you.
(05:42)  All of these type of things are going to change. And the question is, how does this impact all these different companies? How does it impact companies like Door Dash or Uber or Booking Holdings or Open Table? How is this going to impact doing different tasks like your taxes, filing out your taxes with Inuit Turbo Tax? How will it impact bigger companies, ones like Salesforce that work with sales teams? This is going to impact how everything is sold and everything is purchased.
(06:07)  It raises a lot of questions. But now that we see the technology working, we can make some assumptions of which companies are more exposed than others where investors should be concerned and where they should feel a little bit more certain. What I've done is I've looked at this technology and plotted it through my entire portfolio.
(06:25)  I wanted to see how Agentic Technology truly living in a frictionless world impacts my portfolio plus a number of other popular stocks. So, let's go ahead and jump in. The summary is that the largest risk right now being priced into the market is automation. That's what everybody's focused on. Oh, no. Jobs are being automated, processes are being automated.
(06:47)  It's becoming more efficient and there'll be job loss because of automation. But we have to change gears. Automation is what's been happening, but we're moving into a different category. Automation is not the biggest risk today. Losing your customer is the risk. Losing contact with your customer. The popular framing is that artificial intelligence replaces labor.
(07:07)  That matters, but it's not the most consequential economic change. The deeper change arrives when software moves from answering a question to completing a transaction. It compares options, negotiates, books, pays, monitors the results, and learns the user's preference. At that point, the agent becomes the demand gatekeeper.
(07:26)  In Citrony's research, they said as a stress test to ask whether a company's real product is quote, I will navigate complexity that you find tedious. And if the answer is yes, a capable agent attacks the reason the intermediary exists. Stay with me for a minute because we're going to get very specific, but I want to cover a few more generalities before we jump into specific stocks.
(07:48)  The other thing that I think is important to understand here is that the companies that are the most at risk are the ones that sell friction. And there's a lot of companies that only exist to make something that has a lot of friction have less friction. Friction is the product until a machine absorbs the friction.
(08:07)  The more a business monetizes search costs, paperwork, confusion, inertia, and comparison fatigue, the more directly an agent attacks its profit pool. I would evaluate every single company on two axes. How much of its value comes from removing tedious steps and how much scarce execution remains after those steps are automated.
(08:28)  The first creates vulnerability. The second creates resilience, high friction and low scarcity. This is the maximum exposure. These are the companies that I believe are truly in the crosshairs of Agentic Commerce. And there's one that we're going to get into. It's a company that I used to own called Booking Holdings. I am becoming increasingly concerned about this company.
(08:49)  I am more concerned about it today than I was when I sold it. So, I'll be going through that. But the ones that have by far the most exposure are routine travel comparisons. When you're looking at where to travel, where to book, where to get the best plane ticket, those are maximum exposure. Templated legal documents, simple tax filings.
(09:12)  We have init Turboax and simple tax filings. The most friction, they're taking something that's hard to navigate and they're removing friction. We have lead reselling, commodity, freelance tasks, and basic customer support. A general agent can perform the work and shop the provider. Then we have the ones that are a little bit less exposed.
(09:31)  High friction and high scarcity. Interface at risk, but the system survives. So this is the category that I would put Door Dash and Uber in. Door Dash has an interface and Uber has an interface. The bots can talk with the interface, but they can't replace the networks. That's a whole different layer.
(09:47)  Now, even in these situations, losing the connection with a customer is something that should be taken into account. But having that back-end network is very important for these companies. Then we get into the most resilient ones. These are the low friction and high scarcity ones. So we have semiconductor lithography. There's really no friction with those companies.
(10:04) They just create a product that's really needed. They're not going to be disrupted by this. They're going to be very resilient. We have payment acceptance. So companies like Mastercard and Visa, they're not taking a friction system and making it less friction. They have a network that you have to use. So that's one that I don't believe will be disrupted.
(10:23)  We also have credit ratings. companies like SMB Global and Moody's exclusive entertainment. We have Netflix and Disney. We have procurement scale. We have restaurants, live events, and physical experiences. So, you have Cheesecake Factory and Texas Roadhouse. These are companies with already very low friction and very high scarcity.
(10:39) Therefore, they're more resilient. Then we get into the companies that will actually benefit from this. The ones that own the identity, they own the memory and the intent. the platform that knows a user, receives standing instructions, controls permissions, and can complete transactions that become a new demand driver.
(10:57)  Obviously, Meta is one of these companies. They have Muse. They are the ones that are owning the identity, the memory, and the intent. But Meta also has a lot of other businesses that may actually be exposed. So, think about this. If Meta's Muse starts to aggregate all the consumer habits onto its one chat where now you're shopping for Nike products, you're shopping for Amazon products, you're shopping for your Smiths and Walmart and Target all through this single app.
(11:24)  All of a sudden, Meta is the aggregator. Meta can start putting these sponsored agent results in your feed or they can take a sliver of every transaction with your permission. There's a million ways that they can monetize because they are now the new toll collector. You have to go through them to get the product you want because they're the ones that reduce all the friction, know all of your behavior, know your intent.
(11:45)  They know what you want to buy. We've seen examples of this done before. This is exactly what Amazon did with their online store to many sellers. Amazon is the gatekeeper. They have Amazon.com. Everybody goes through them to get to all the millions of sellers. Now, before we get into specific companies, I still think that there's some things we need to outline specifically with how powerful the agents are and how it's different than Chat GBT or Gemini.
(12:08)  A lot of people still might conflate these two things together because there's some similarities, but I want to go through some of the distinctions and why agents are really different. So, we have stage one, and this is what we're already in. We're already in the chatbt era of review pages being pressured, how-to content being commoditized, search results and clicks, and basic customer service.
(12:31)  But it often leaves the incumbents in control of execution. A chatbot can recommend a flight while the user still books through an airline or an online travel agency. That's what people are doing right now. They're going to catch and saying, "Hey, plan out my vacation." Then they learn from that, but then they still use all the services in the way that they're using them before.
(12:51)  This is where things diverge with Agentic Commerce. Stage two, once the software can take authorized action, the incumbent may lose the session entirely. Find me a good hotel near the conference for under $350 with a quiet room and free cancellation. That becomes one instruction. The agent query supply interprets terms, chooses, books, and stores a confirmation, and updates the calendar.
(13:17)  The traditional comparison interface is no longer mandatory. You don't actually have to visit any of these companies websites to accomplish what you want to accomplish. That's a huge divergent from what we have today. Then we have stage three, memory. Memory converts convenience into distribution. Persistent memory creates a compounding advantage.
(13:37)  The agent knows that the user values arrival reliability over the lowest delivery fee, dislikes redeye flights, prefers aisle seats, avoids restaurants with loud dining rooms, and wants tax documents filed automatically. This is not simply personalization. It is a durable demand profile. So, if you look at this again, it's very valuable to just know and learn about a person because it makes the future instruction set even have less friction than the previous one.
(14:03)  The first time I do a shopping list, I have to go through my kids allergies. Every single shopping list after that, I don't unless their allergy profile changes, unless my family profile changes, unless one of them moves out or something. But until then, now every single time I shop, it's easier than the first. That makes it so that you have distribution, the people that control the memory profile, this advanced learning about the user themselves.
(14:27)  When you enter in the credentials of Smiths or Target or Walmart, it'll save that securely forever. You don't have to enter it in again. It already knows that and that creates that lock in and distribution. Then we have stage four. Commerce becomes machine to machine. This is a conversation I was recently having with my brother.
(14:45)  He said airlines are starting to price things by the minute based on the user. The airlines are trying to determine exactly how much you're willing to pay by increasing prices to make more money. That's true, but that's only part of the story. We aren't living in a world where the airlines will have dynamic or personalized pricing and humans will just book airline tickets by themselves.
(15:04) We live in a world where airlines will have dynamic and personalized pricing and humans will have agents to fight back. We will put their agents on them and they will battle it out. This is where we get to commerce becomes machine-to-achine. Google's universal commerce protocol is designed to connect agents, merchants, and payment providers across discovery, checkout, post-purchase support.
(15:25) Microsoft describes hosted agents, multi-agent workflows, persistent memory, and Microsoft 365 integration. Visa, Mastercard are building tokenized credentials, authorization controls, and ways to identify legitimate agents. The plumbing for this is being built in public. And how do we know that Meta's Muse or similar type of agents will aggregate all of this demand? Well, we've already outlined some of the advantages, but I want to go through a few more.
(15:50)  Many marketplaces look like networks, but behave economically like habits. Agents are unusually good at breaking habits that exist only because comparison is annoying. So, again, we go back to that first thing. If the product only exists because of comparison, you think about a company. If the only reason the company really exists is to make comparison easier, if that's like 70% of the value proposition, it's in a lot of trouble.
(16:16)  A person orders from an app already installed. They may have a subscription, saved addresses, familiar interfaces, and a learn expectation that the order will probably work. They do not check the restaurant website, competing delivery apps, pickup prices, credit card offers, estimated arrival quality, and current promotions every night.
(16:35)  The cost of checking exceeds the expected savings. So, the reason people don't do all this extra work to save money is it just takes too long. You're not going to scavenge the internet to try to save some money on your booking. They're not going to try other booking apps, even if it saves a little bit of money. An agent's cost structure is different.
(16:51)  It can check every route each time. It does not get tired, forget a coupon, or decide that seven tabs are too many. Citrony calls this weakness habitual intermediation. demand appeals loyal to the intermediary because human default to the familiar path not because the intermediary owns an irreplaceable capability.
(17:13)  So a lot of companies that we think have a huge moat because they seemingly have a lot of demand. That moat only exists because it's annoying to do something else and agents take away the annoyingness of it. Humans by default open familiar apps, browse ranked inventory, accept displayed prices and use saved payments. Hope fulfillment is good.
(17:34)  Agent default is they query all authorized channels, normalize fees and terms, score price, quality, speed and reliability and policy and execute the best fit that weakens app habits, default status, opaque fees, renewal inertia, and sponsored rankings. These advantages do not disappear instantly. Authentication, loyalty rewards, subscriptions, and exclusive inventory can preserve preference, but the agent can calculate whether the bundle is actually worth it.
(18:04) It turns soft loyalty into a consciously tested economic relationship. So, a lot of things people do just because they've been doing it just because the other options are a little bit more annoying, they're no longer going to be doing that just because other options are slightly more annoying.
(18:21)  these type of products and services on the back end will constantly be competing to keep the same relationship. It removes a lot of that inertia. So, now that we've gone over a bit of the technology, what I expect to happen, let's go ahead and look at some examples. And again, these aren't 100% guaranteed, but I believe these are far more accurate than what Citroen gave us a few months ago.
(18:41)  We'll start off looking at the specific company Booking Holdings. We have Booking and Kayak. Travel is the cleanest early test because comparison is valuable, terms are structured, and the consumer already expects an intermediary to assemble options. Booking holdings reported $8.1 billion of marketing expense in 2025, equal to about 30% of revenue.
(19:04)  So every $10 that they gain in revenue, $3 of that goes to marketing. It also discussed a higher direct room night mix, lower performance marketing return on investment, and changes in paid traffic mix. Those facts show both the strength and fragility of the model. Booking has global supply, conversion data, payments, customer support, and brand.
(19:25)  Yet, it still spends heavily to acquire the demand. The threat to booking holdings, and the reason that I'm so worried about this company is a gatekeeper swap. Historically, a meaningful part of the online travel economics flowed through search engines. Booking paid to appear wherever travels expressed intent. In an agentic market, the traveler may express that intent once to a personal assistant.
(19:46)  The agent asks clarifying questions, searches broadly, applies loyalty and cancellation preferences, and then executes. The paid search auction does not necessarily vanish. It can reappear as a sponsored recommendation or preferred supplier slot inside the agent. Travel has ugly expectations. Flights are canceled all the time.
(20:05)  Rooms do not match descriptions. Payment methods fail. Local taxes differ. Loyalty terms conflict and refunds require persistence. A general agent may prefer to call Bookings's infrastructure rather than rebuild global contracting and customer service. Booking can also build its own trip agent, use its firstparty data and increase direct engagement, package flights on the ground transport, and make Genius membership an explicit user experience.
(20:30) So, they do have something they can do. Booking can try to build its own agentic layer as well and use all of its data. They're certainly going to do that. But again, they're competing in a very uphill battle here. They are the most exposed because they don't have any real internal structure like Door Dash or an Uber where they have thousands of drivers or riders going around picking up food.
(20:53)  The big thing that they offer is putting together this trip, making sure that they have all the terms correctly, and compiling all of that as a gatekeeper. And this is exactly the type of thing that agents are good at. The bare case does not require bookings volume to collapse. And that's an important thing here. It requires the marginal traveler to become owned by a different interface.
(21:14)  So even if booking holdings is being routed all the customers losing that direct interface control where now they're just an API a backend being routed is a big change in their economics. If the agent can credibly route around booking booking's marketing bill may move from Google to the agent layer or appear as a lower net take rate rather than an explicit expense.
(21:37)  The agent would have a huge part of the control, the consumer preference, the ability to advertise, to do sponsored listings or take a pay if they control the consumer demand. Both Open Table and Door Dash are aggregate demand companies. But the resilience differs because Door Dash physically executes the transaction while Open primarily coordinates it.
(21:57)  Open table is super vulnerable. Here we have Door Dash, which does have a level of vulnerability, but it also has some part of it that's more defensible. The agent can compare Door Dash to Uber Eats, restaurant direct ordering, pickup, or subscription benefits and promotions. This type of comparison, reaching outside of Door Dash's interface when you open up the app, attacks that habit forming aspect.
(22:17)  But the delivery part of this is not a database lookup. It requires local carrier density, dispatching, batching, routing, fraud controls, merchant integrations, accurate prep instructions, customer support, refund, and service recovery. Door Dash has reported that revenue margin improvements was supported by list logistics efficiencies, advertising and fewer credits and refunds.
(22:37)  A reminder that operating quality materially shapes the economics. So when we look at these two type of companies, this is what we should be thinking about. Companies where you're just the interface and you're just a comparison tool or a booking tool like OpenT, you're super exposed to agents. Door Dash has exposure because it can reach outside of the interface, but you still have to go with whatever network has the most density and the best execution, which is likely Door Dash.
(23:02) Now, even with something that has as defensible of an infrastructure and network and logistics as Door Dash, even taking away the interface is still some intermediary risk. It's something to take into account. When I look at my holding in Door Dash, I can see that it's dropped down a little bit. And I believe it's because investors are becoming a little concerned about these type of things, which they should be factoring in now.
(23:23)  Now, I'm keeping Door Dash as a position because I think it will be more defensible than the other companies like Booking Holdings or into it. Speaking of into it and taxes, that's the next one that I want to address into it with tax preparation becomes ambient. For a simple filer, the enduring friction is collecting documents, answering repetitive questions, mapping facts and rules, and submitting forms.
(23:47)  A persistent financial agent with permissions to access can make much of this continuous rather than seasonal. That pressures the value of a guided form completion. So if you look at what Turboax does, it's basically a gigantic form that you have to fill out. Like every step of the way, it has little animations.
(24:03)  It guides you through it. It can have like tax people that zoom in or whatever. But it really is just a big form to fill out. Now, why do you pay $100 or $200 for Turbo Tax? Because taxes are important. You don't want to screw it up. But that's where this gets a lot of pressure. If you can have an agent that already knows a lot about you, your situation, and you feed it some of the documents, and it can piece it together with reliability, that makes it much more difficult to justify paying separately for in it.
(24:29)  Now, in it is starting to fight this. Of course, they're not going to just stand still and do nothing. All of these companies have their own plan. They're trying to implement this technology as well. When I sold into it, I was concerned about Turboax. But after really understanding what this technology is capable of, I'm far more concerned about in it IT shareholders today than I was when I sold I believe a lot of the value that in it offers with Turboax is being compressed and a lot of the tools even like Credit Karma are being compressed
(24:56) as well. Now again these don't happen overnight but the technology is progressing dramatically. We have other companies ones like Legal Zoom. Templates are exposed. Execution and judgment remain. Entity formation, standard contracts, wills, and compliance reminders are structured enough for agents to absorb much of the explanation and document assembly.
(25:16)  I know people that work at jobs as accountants and lawyers. And a lot of what they do is summarize documents and they try to understand the nuances of contracts. Agents are already doing that. AI is already summarizing these documents, flagging red flag things. they already do much of that work. So any type of software that you sell to do that is going to be pressured by agents.
(25:36) Doing stuff like routine logo variations, basic copy, transcription, research summaries, spreadsheet cleanups, simple coding translation, and first pass support can be completely directed by agents. All of this is completely replaceable. Home selling companies, lead generation companies are going to be dramatically influenced by Agentic technology.
(25:58)  insurance, housing, travel portals, they're going to weaken because comparison is continuous. Select quote, Go Health, and Ever Quote. Insurance distribution is built around complexity, limited consumer attention, and periodic shopping. Once people sign up for their insurance, they just don't want to look anymore. It's annoying.
(26:15) It's not fun to look at competing offers. But there's one person on X just yesterday that he said, "Here's my insurance document. Can you find me anything better?" In 20 minutes, the agent found insurance for $600 per year less than what they're paying with the very same terms. So, people are going to do shopping that they previously weren't doing because now it doesn't have the friction. We have Zillow.
(26:37)  Think about how at risk Zillow is. A housing agent, one like Muse, could search listings, model trade-offs, schedule tours, compare neighborhoods, monitor price changes, and gather disclosures. They're already capable of doing all of this. You can have it keep track of prices of items or multiple things. You can have it remind you of things.
(26:56)  You can have it watch and scour real websites. How are these companies going to compete when an agent baked in already does all of this? All of these companies are trying to insulate and become more defensible and build these technologies in themselves. But it's going to be very difficult, a challenge for all of these companies.
(27:12) Now, we move on to much bigger companies, ones that are in my portfolio and likely in yours. My biggest holding today is Google, a $24,000 position. $110,000 in the green on this one. So, it's been a great position. Let's go ahead and take a look at how Alphabet fares in a world of agents. There's some good here and there's some bad.
(27:31)  Alphabet reported $224 billion of Google search and other revenues in 2025, up from 198 billion in 2024. This scale matters because even a gradual shift in commercial intent can have large consequences. Traditional search monetizes a sequence of user queries and choices. An agent can collapse that sequence into one delegated request.
(27:55)  The person searching quote best hotel in Chicago or cheaper car insurance or tax software for freelancers creates multiple ad opportunities. An agent may gather requirements once, query providers through structured connections, and return one or two recommendations. Fewer visible options can mean fewer paid clicks and less opportunities to monetize browser friction.
(28:17)  So this is the downside is that basically if you search things through like a Muse or even a Gemini and it just finds the thing that you want without giving you a 100 results, there's less opportunities to advertise. But there are a lot of ways that Alphabet's moving that are going to protect it.
(28:33)  One of them is just raw distribution. They have Android, Chrome, search, Gmails, map, YouTube, all these different things. They have intent data, queries, location, video, emails, map, commercial signals can help authorize agents understand context. That's very valuable. They have a technical stack. We have Gemini, cloud infrastructure, developer tools, and enterprise integrations.
(28:53)  And then we have commercial protocol. The universal commercial protocol aims to connect discovery, checkout merchants, and post purchase service. So, Google has a lot of ways to deal with this and adapt that many other companies that are exposed don't. Google can also reinvent sponsored placements. The relevant ad unit may stop being a blue link and become a clearly disclosed recommendation, merchant incentive, or preferred fulfillment option or bid for an agent's consideration set.
(29:21)  The core capability matching commercial intent to supply and running an auction remains valuable if the user trusts the result. So basically, even if an agent skips Google's search, Google has so much information that they could work with the agent to still present a very valuable backend. They could really offer a lot of intent and things that the agent will need to make a good decision.
(29:43)  And Google could monetize that intent. This is much more defensible than other companies that are exposed. Next, we go into an odd combination. We have Amazon, Uber, and Dualingo. Each of these companies lose one thing, but it's a different thing. Amazon's product search is exposed. A shopping agent can compare Amazon with brand sites, Walmart, niche retailers, resale markets, and local inventory.
(30:03)  That weakens Amazon's role as a default place to begin a product search and may pressure sponsored listings if the agent filters aggressively. Yet, Amazon owns enormous seller selection, prime loyalty, fast fulfillment, returns, payments, reviews, advertising demand, and Amazon Web Services. Agents may prefer Amazon because the full transaction is reliable.
(30:26)  Amazon can also supply the infrastructure that runs agents and creates its own shopping assistant. They're already clearly doing that. Amazon's trying to build in their own agentic shopping system with Alexa. It's the new Alexa in the Amazon. But still, this is a potential problem because if you look at Muse, it'll shop on Amazon plus a million other websites.
(30:45) The risk is to some of the search and advertising economics. The opportunity is more transactions through superior execution. Again, the agents will gather all that demand on the front end. So the companies with the best back-end infrastructure, the best quality, the most reliability, the best customer support, they'll aggregate the demand on the back end.
(31:06)  With Uber, this is far less exposed because while the agent can compare Uber against Lyft, Whimo, Tesla, it can compare all of those together. Ultimately, it's going to go with the best combination of reliability, all your preferences, safety, travel time, all of that, plus price. And Uber already has massive demand density which will make it in most cases the best choice. We have Dualingo.
(31:28)  Now this is the same type of case we've had for a while with this one. The utility of it is exposed but the motivational aspects are not. When you shop with an agent the agent knows every language. So it can navigate any language of any product of any website and shop on your behalf. In some ways that reduces the practical need to learn a language.
(31:47)  A lot of people, in fact, the majority of people learning in Dualingo aren't doing it because they want to be able to just transact easier. They really want to be able to communicate and be part of a different culture. So, I would not extend that into Dualingo is highly exposed overall. Language learning is also a hobby, an identity, a project, a social signal, and a long-term achievement.
(32:06)  Streaks and leagues and character and curriculum design, progress tracking, habit formation matter. Users often want to become a person who speaks a language, not merely solve a translation problem. Next we get into how these agents will impact Microsoft and Meta. Microsoft is the agentic operating system. Microsoft sits inside work identity, emails, documents, meetings, operating systems, developer tools, databases, cloud infrastructure, and security.
(32:32)  Its foundry strategy describes hosted agents, multi-agent workflows, persistent memory, Microsoft 365 integration. The architecture puts Microsoft in a position to coordinate tasks across applications rather than merely add a chat box to each one. This is something that Sachin Adella has done incredibly well.
(32:50)  He is super forward thinking. He's been thinking about this for a while. He's not going to let Microsoft be disrupted by agents even when they don't have their leading AI model lab. The economic advantage is permission context. An enterprise agent needs to know which employees have access to which customer records, which policy governs an action, who approves spending, and how activity is audited.
(33:11) Microsoft is fairly insulated here. And ironically, while it's a software company that hasn't performed too well, it's likely far more insulated from agentic threat than many other companies. With Meta, we have a whole different ballgame because Meta sits on both sides of things. Meta has Muse, its own agentic technology, but Meta also has a massive advertising business.
(33:31) Advertising relies on shopping. Meta's core product monetizes attention and discovery, not primarily the user's need to navigate a tedious commercial task. An agent that removes friction from buying may actually make the feed ad more valuable. The user discovers a product socially, then delegates comparison and checkout.
(33:51)  Meta has said AI interactions can become another signal used to personalize content and ads. This creates a direct feedback loop. Conversations reveal interest and intentions. The feed creates demand. Messaging connects people and business and agents can help complete the action. When we actually look at this, meta is probably more insulated from a gentic threat than Google is because of the way that their advertising works.
(34:14)  Search ads often monetize declared intent at the moment a user is solving a problem. Feed ads often create or shape intent inside an entertainment or social environment. Agents compress the first more directly than the second. So, if you think about it, with an agent, you're far more likely to not run into a Google ad because you're not needing to search different things than you are to run into a meta ad.
(34:36)  Because even with agents, you still want to have Instagram res and you want to check up on Facebook or whatever. You're still using those services because those services aren't based on friction. They're based on entertainment. Both of these companies have to become the gatekeepers. Microsoft's route begins with work and permissions.
(34:54)  Meta begins with social identity, messaging, creators, and attention. The reason that I have both of these companies still in the portfolio today, and I've been buying Meta like crazy, because not only I believe the valuation is good, but Microsoft and Meta are likely well insulated and even beneficiaries of Agent technology, just the opposite of in it and booking holdings.
(35:14)  Now, next up, we get to the big payment processors, Visa and Mastercard. This is where Citrony basically said that these companies were going to go way down. They're going to get hurt somehow, and I never agreed with that. I continue to invest in Mastercard aggressively, and I still have a huge position in Mastercard today.
(35:31)  It is a $189,000 position. There's actually a reason to believe these companies will do much better in this new agentic commercial world. Visa and Mastercard represent trust on the transaction level and more actors and more complexity creates more need more need for trust. An autonomous transaction introduces new questions.
(35:48) Was the agent authorized? Was the amount within the user's limit? Which merchant received the credential? Can the payment be traced, disputed or revoked? How does the merchant distinguish a trusted agent from an automated fraud? Visa's position is Agentic Commerce works around tokenization, authentication, spend controls, payment, and infrastructure.
(36:08) Mastercard's agent pay introduces Agentic tokens built on existing tokenization. Pasi capabilities with partnerships spanning from Microsoft, IBM, Brainree, and Checkout.com. These are extensions of network roles. These companies are not trying to win the interface. They don't care. Visa Mastercard never really care about being the interface.
(36:29)  Visa, Mastercard just care about being on the transaction layer and this arguably introduces more transactions and more need to trust those transactions as humans are further away from the transaction. There's one step happening in between. So, there's a bigger need for trust. So, I'm keeping my holding a Mastercard because I believe the actual upside potential is bigger than any type of downside potential with that company.
(36:51)  Next, we get to ASML. Agent growth increases demand for physical bottlenecks. This should be pretty basic, but agents don't route around leading edge lithography. There's just no going around ASML. If you want to do anything that requires a chip, you're going through ASML. Not only do they not route around ASML, but they rely on ASML.
(37:12)  They consume inference and training capacity, which increases demand for advanced logic and memory. ASML's moat comes from physical accumulated engineering, supplier coordination, service infrastructure, and customer process integration. A generalurpose model cannot synthesize that capability through a better interface.
(37:32)  ASML continues to be the apex predator. Nobody hunts it. It's just at the top of the food chain. Nobody can touch it. Agents can't either. And not only can agents not threaten ASML, but they create further demand for ASML's incredibly monopolistic technology. So if we look at these companies in combination, Visa Mastercard secure the transaction, but ASML enables the computation.
(37:55)  Both of these companies participate in the agentic activity without needing to own the assistance brand. Not only are these companies not being circumvented and they're actually benefiting from AI, but the Agentic layer also has no need to replace them. It's not even their goal. It's not where they're planning on monetizing to begin with.
(38:11)  Now, let's move on to the next batch of companies. We have to ask how companies like S&P Global, Netflix, Costco, and Texas Roadhouse. These are four companies. How do agents really affect them? Well, not to any big degree. SM Global and Moody's is one of the questions people have. But the value of these companies is not from less friction.
(38:30)  It's from authoritative output. Credit ratings are not merely summaries of public information. They are regulated, embedded, and mandated in contracts backed by methodology and surveillance and issued by institutions who opinions market participants recognize. Proprietary data, benchmarks, workflows, and long histories become more useful when an agent can query them.
(38:51)  S&P Global describes subscription revenue built from data, valuation services, analytics, research, and rating information alongside ratings, and index linked economics. So, a lot of investors have this wrong. They think that because AI can put together good research summaries that it will pressure SMB Global, but that's just the opposite.
(39:10)  They are getting more demand because AI is using their data to put together research summaries. Generic analysis may become cheaper. trusted, licensable, auditable inputs do not. So, when you look at it, it is cheap to find generic analysis websites. I have one. I sell it for $10 a month. It's not that expensive.
(39:28)  But when you're running a fund that manages 10 billion dollars or you're a bank and you're lending out hundreds of billions of dollars, you can't go with just whatever makes economic sense for your small portfolio. You need to go with what makes economic sense for your bank or institution. In that case, you have to have perfectly accurate decision-grade data.
(39:44)  And that's where these companies come in. The reason that the ratings are so important is not because they could find that information somewhere else. It's because of the authority and trust and brand recognition behind them. We also have Costco. Agents can compare prices. They cannot duplicate the bundle. Costco's value is in procurement scale, limited assortment, private label quality, membership economics, trusted pricing, and physical operating model.
(40:11)  An agent may help a household decide what to buy and when, but it can also reinforce Costco by recognizing durable value and routing recurring baskets there. Costco is more of a supplier than a search interface. With Netflix, a general assistant can tell a user what to watch across services that may weaken Netflix's home screen discovery advantage.
(40:34)  But Netflix owns and licenses its own entertainment. The agent can point to a show. It cannot legally stream the catalog without a commercial relationship. I find that Netflix is very insulated from agentic technology. A restaurant agent can reserve a table, compare weight times, and suggest dishes. It cannot cook the steak, create the service culture, or reproduce the atmosphere.
(40:54)  The same is true for theme parks, gyms, hotels, concerts, travel experiences, and live sports. Asians may reshape discovery and booking while increasing utilization of the underlying physical product. I actually believe that Texas Roadhouse is not only fine, but because all their operations, the actual restaurant so good, they may actually get more increased demand as a result of Agentic Commerce.
(41:15)  Now, again, I don't know if this will play out perfectly this way, but I believe this is a more accurate look at the type of impact Agent technology will have than what Citriny came out in February. And we'll be able to update this and see how it's going in real time because again, in months time, we're going to have millions of people using agents to shop.
(41:31) It is right around the corner. We look at the companies that have the highest exposure today. These are the ones where the agent can take over the user interface. Now amongst this, I also believe there's a high variability. For example, I think that Door Dash is much less exposed than Open Table or into it or booking holdings because Door Dash has that physical infrastructure behind.
(41:54)  We have booking having some physical infrastructure but not nearly as much. We have intu tax prep. If filing becomes ambient, if it's much easier to organize documents, what does that do to in it? And we have all of these type of companies where the agent will already know a lot of the intent. We get to the companies where they're a little bit more moderate or they may be beneficiaries.
(42:14)  We have companies that are moderate like Uber, Dualingo, Google with their search ads, Amazon with their ads. In many of these, it's mixed cases. Part of the business is exposed. Part of it will actually do really well. And then we have the companies that are resilient. FICO is an authoritative brand is more resilient.
(42:29)  We have S&P Global and Moody's very resilient. Costco and Netflix very resilient. Texas Roadhouse, Cheesecake Factory, all of those type of companies are very resilient. And then we have some that should actually benefit. Microsoft and Meta should benefit from Agentic Commerce. Shopify should benefit. ASML should benefit as well.
(42:46)  So overall, that's my thoughts on how this could play out. Now, if the agentic threat turns out to not really be real, if people don't really use the agents, it'll just be because the agents weren't good enough to make it so that it was better than using the actual interface. If these companies can really create a much better experience and people just want to go directly to that experience, then the agentic commerce won't play out. We won't have this thesis play out.
(43:09) But I believe there's a very solid chance that agentic commerce is going to be a very real thing because people like reducing friction. They like the easier option. And in many of these cases, the agent is easier. I know this was an unusually long episode and a little bit more dense than most, but hopefully this was informative.
(43:26)  That's going to be it for this episode. Have a good one.