AI Agents Threaten Booking, Intuit by Bypassing Consumer Portals

Joseph Carlson After Hours

Companies that monetize transaction friction face margin collapse as autonomous agents become demand gatekeepers, shifting pricing power toward payment rails and physical logistics networks.

AI-assisted summary of Warning: Investors Must Prepare For The Agentic Takeover — original source. Check the original for full context.

Source published: 2026-09-18

Key ideas

The deployment of practical consumer AI agents like Meta Muse enables autonomous multi-step execution—including grocery ordering, trip planning, and reservations—directly on merchant websites, completely bypassing intermediary apps and search interfaces. (2:03)
Booking Holdings faces severe structural risk from a consumer gatekeeper swap, threatening an asset-light business model that spent $8.1 billion in 2025 (approximately 30% of gross revenue) solely to capture travel search intent. (18:10)
A critical divide exists between pure software coordinators and physical execution networks: platforms like OpenTable are easily circumvented by agents, whereas DoorDash and Uber retain defensible moats through physical driver density and logistics fulfillment. (21:45)
Visa, Mastercard, and ASML stand as core structural beneficiaries, providing the indispensable identity and authorization rails for machine-to-machine payments as well as the EUV lithography hardware required to power agent compute demand. (36:12)

Detailed notes

## Background Published on YouTube on September 18, 2026, by Joseph Carlson After Hours, this episode examines the investment implications of autonomous AI agents becoming the primary intermediaries of consumer transactions. Carlson outlines a friction-versus-scarcity evaluation framework based on his video description to determine which publicly traded businesses face structural disruption and which possess defensible moats as software assumes the role of demand gatekeeper.

## TL;DR Carlson analyzes the transition to agentic commerce, demonstrating through tools like Meta Muse that autonomous AI agents are moving beyond passive chat to execute end-to-end purchasing, booking, and price comparison. He asserts that the true threat to business models is not labor automation, but the loss of direct interface ownership as agents become consumer gatekeepers that bypass traditional search ads, sponsored listings, and friction-reliant aggregators. Evaluating companies on friction removal versus scarce physical or regulatory execution, he identifies severe exposure for asset-light middlemen like Booking Holdings, OpenTable, and Intuit TurboTax. Conversely, businesses anchored by physical logistics, regulatory mandates, computational hardware, or mission-critical payment rails—such as DoorDash, Uber, S&P Global, ASML, Visa, and Mastercard—remain insulated or poised to capture expanding transaction volume.

## Atmosphere & style The presentation carries a calm, methodical, and pragmatic tone, balancing serious economic warnings against the hyper-viral panic seen in financial commentary. Carlson communicates with the grounded perspective of a programmer-turned-investor, anchoring high-level architectural theories in vivid personal anecdotes, such as testing Meta Muse live to navigate his family's specific grocery allergies and budget limits. The experience feels like a transparent, behind-the-scenes portfolio review that dissects marketing hype to soberly determine where corporate cash flows are genuinely defensible.

## Bulletin - [02:03](https://www.youtube.com/watch?v=HKskNSEb5_c&t=123s) The deployment of practical consumer AI agents like Meta Muse enables autonomous multi-step execution—including grocery ordering, trip planning, and reservations—directly on merchant websites, completely bypassing intermediary apps and search interfaces. - [18:10](https://www.youtube.com/watch?v=HKskNSEb5_c&t=1090s) Booking Holdings faces severe structural risk from a consumer gatekeeper swap, threatening an asset-light business model that spent $8.1 billion in 2025 (approximately 30% of gross revenue) solely to capture travel search intent. - [21:45](https://www.youtube.com/watch?v=HKskNSEb5_c&t=1305s) A critical divide exists between pure software coordinators and physical execution networks: platforms like OpenTable are easily circumvented by agents, whereas DoorDash and Uber retain defensible moats through physical driver density and logistics fulfillment. - [36:12](https://www.youtube.com/watch?v=HKskNSEb5_c&t=2172s) Visa, Mastercard, and ASML stand as core structural beneficiaries, providing the indispensable identity and authorization rails for machine-to-machine payments as well as the EUV lithography hardware required to power agent compute demand.

## Key takeaways - [05:43](https://www.youtube.com/watch?v=HKskNSEb5_c&t=343s) While markets originally feared widespread labor automation following Citrini Research's viral report, the primary economic threat of the agentic era is companies losing their direct customer interface and relationship. - [02:03](https://www.youtube.com/watch?v=HKskNSEb5_c&t=123s) Meta Muse represents the transition from passive chatbot conversationalists to proactive agents capable of handling complex errands across the web, such as building allergy-compliant grocery carts and applying digital coupons. - [07:26](https://www.youtube.com/watch?v=HKskNSEb5_c&t=446s) The Friction Framework evaluates enterprise durability along two vectors: the degree to which value stems from removing tedious navigation versus the amount of scarce physical, legal, or infrastructural execution that remains. - [09:00](https://www.youtube.com/watch?v=HKskNSEb5_c&t=540s) Businesses built primarily on high friction removal and low asset scarcity—such as online travel agencies, lead-generation aggregators, and simple tax prep software—sit in the immediate blast zone of agentic disruption. - [11:57](https://www.youtube.com/watch?v=HKskNSEb5_c&t=717s) The agent economy progresses through four architectural stages: informational assistance, autonomous execution, persistent user memory lock-in, and dynamic machine-to-machine commerce negotiation. - [15:48](https://www.youtube.com/watch?v=HKskNSEb5_c&t=948s) Autonomous agents eliminate 'habitual intermediation' by replacing human cognitive fatigue and default app opening habits with instant, objective cross-platform price and fee normalization. - [18:10](https://www.youtube.com/watch?v=HKskNSEb5_c&t=1090s) Booking Holdings risks being reduced from a high-margin front-end destination to a commoditized backend fulfillment API as travel intent is captured directly by personal AI assistants. - [21:45](https://www.youtube.com/watch?v=HKskNSEb5_c&t=1305s) DoorDash maintains an operational moat that asset-light reservation platforms lack, because software agents cannot replicate physical driver networks, route optimization, batching economics, and merchant integrations. - [23:30](https://www.youtube.com/watch?v=HKskNSEb5_c&t=1410s) Intuit faces profound pricing pressure on TurboTax and Credit Karma as persistent financial agents transform episodic annual tax filing into continuous, ambient background reconciliation. - [25:01](https://www.youtube.com/watch?v=HKskNSEb5_c&t=1501s) Continuous real-time market monitoring by autonomous agents dismantles inertia-driven margins in insurance lead generation (such as SelectQuote) and boilerplate document preparation (such as LegalZoom). - [27:05](https://www.youtube.com/watch?v=HKskNSEb5_c&t=1625s) Alphabet experiences top-of-funnel search ad volume pressure as multi-click query journeys collapse into direct agent answers, but it retains defensive depth via Android, Chrome, YouTube, and its full-stack AI infrastructure. - [29:47](https://www.youtube.com/watch?v=HKskNSEb5_c&t=1787s) Amazon faces margin compression in high-margin sponsored search ads, yet Prime delivery speed, return logistics, and marketplace selection ensure agents frequently choose it for final order fulfillment. - [29:47](https://www.youtube.com/watch?v=HKskNSEb5_c&t=1787s) Duolingo is shielded from functional translation disruption because its primary monetization engine relies on gamification, user identity, and habit-forming entertainment rather than pure utility. - [32:15](https://www.youtube.com/watch?v=HKskNSEb5_c&t=1935s) Microsoft and Meta act as structural toll collectors: Microsoft anchors enterprise identity, document permissions, and multi-agent workflows, while Meta captures frictionless discovery-driven shopping through attention feeds. - [35:14](https://www.youtube.com/watch?v=HKskNSEb5_c&t=2114s) Visa and Mastercard are reinforced rather than replaced, as machine-delegated purchases dramatically heighten the necessity for tokenized credentials, spending limits, fraud prevention, and centralized settlement rails. - [36:12](https://www.youtube.com/watch?v=HKskNSEb5_c&t=2172s) ASML remains an impenetrable bottleneck at the top of the technology supply chain, wholly immune to agent automation while directly capturing value from surging model compute requirements. - [38:11](https://www.youtube.com/watch?v=HKskNSEb5_c&t=2291s) Companies anchored to regulatory mandates, physical scale, exclusive IP, or offline experiences—such as S&P Global, Costco, Netflix, and Texas Roadhouse—face negligible risk from autonomous software agents.

## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=HKskNSEb5_c&t=0s) The Agentic Commerce Thesis and Citrini's Viral Warning

In February, a hyper-viral article from Citrini Research—already the largest financial Substack in the world with millions of readers—shook the markets and triggered widespread panic. Published on February 22nd under the title *"The 2028 Global Intelligence Crisis,"* the piece painted a narrative of a future where AI agents have taken over the world. The release of this single article sent many stocks down 7% to 10% almost immediately, sparking a continued sell-off across software companies that came to be known as the "SaaS apocalypse."

While the article noted that intelligence is becoming abundant because AI knows about everything, its central thesis was about **"friction going to zero."** Humans routinely spend time filling out forms, navigating user interfaces, shopping, purchasing, and comparing options—all of which represent friction. As robotic extensions of humans, AI agents can go out and perform all of those tasks on a user's behalf, effectively erasing that friction. Citrini didn't speak only in abstractions; they named specific companies and projected how the disruption might play out:

- **ServiceNow:** Projected to experience significant growth deceleration. - **Mastercard:** Featured in a fictionalized future headline showing revenue growth decelerating down to 6% year-over-year. - **Software and Labor Markets:** Predicted that software companies would be obliterated and the broader job market destroyed. - **DoorDash:** Highlighted as the ultimate poster child of a business model that would be disrupted by agentic technology.

Although Citrini's article was well-written, thought-provoking, and convincing at the time, Joseph notes that it was also full of flaws regarding how many of these specific companies will actually be affected. Still, he gives the authors the benefit of the doubt: back in February, true consumer agentic technology did not really exist the way it does just a few months later.

[02:03](https://www.youtube.com/watch?v=HKskNSEb5_c&t=123s) Meta Muse: The First Real Consumer Agent in Action

The first major consumer release of truly agentic technology is Meta's new app, **Muse**. Having used the app extensively, Joseph emphasizes that anyone who thinks Meta Muse is just another ChatGPT, Grok, or Anthropic's Claude is mistaken. Opening Muse was the first time since initially seeing AI write code that he felt genuinely "wowed" again. Instead of an empty chat box that passively spits out information, Muse operates as a proactive assistant ready to accomplish tasks across the web:

- Booking hotel and restaurant reservations—either by linking to OpenTable or verifying directly on a restaurant's website. - Connecting to an email inbox to clear out clutter and automatically unsubscribe from unwanted emails. - Organizing trips, planning itineraries, and shopping autonomously on the user's behalf.

Testing the restaurant feature immediately revealed the strategic threat to existing platforms. When a user can book a table through OpenTable directly from a text box in Muse without ever opening OpenTable's app, the entire customer relationship changes. The user interacts solely with the agent, completely bypassing OpenTable's sponsored listings, promotional pop-ups, and even the need to keep the OpenTable app installed on their phone.

To test its e-commerce capabilities, Joseph first asked Muse to track down a very specific pair of Nike shoes that his brother-in-law was looking for. Muse launched its own built-in browser, searched far and wide across numerous websites, read through reviews, and hunted for the exact shoe match at the best price—functioning like an executive assistant or secretary dedicating 30 minutes of focused time to a single errand.

Pushing the app even further with a complex household chore, Joseph gave Muse a real-world grocery test for his family. He has four children—one is a newborn, and the other three eat regular groceries, with two of them having specific dairy and nut allergies. He entered the kids' ages and allergy restrictions into Muse and asked for a complete weekly meal plan. Muse replied "No problem," recommended shopping at Smith's because the grocer carries a wide range of allergy-friendly alternative foods and substitutes, and went straight to Smith's website. Watching the agent work live on screen, Joseph observed Muse build the entire week's meal plan and add every single item to the cart while simultaneously:

- Verifying that each product met the strict dairy and nut allergy requirements. - Selecting the item with the lowest price. - Attaching available digital coupons and searching online for any applicable discounts to lower the final bill.

Muse is only a preview of the agentic world now arriving. OpenAI, Anthropic, and every other major AI lab will soon roll out competing agents, giving consumers multiple digital extensions to shop for groceries, plan trips, and book hotels—making direct visits to platforms like Booking Holdings or Expedia far less necessary.

[05:43](https://www.youtube.com/watch?v=HKskNSEb5_c&t=343s) How AI Agents Reshape Business: From Automation to Demand Gatekeeper

Now that working agentic technology is here, investors must confront how a frictionless world will impact companies across the board—from consumer platforms like DoorDash, Uber, Booking Holdings, and OpenTable, to task-based software like Intuit TurboTax, to enterprise giants like Salesforce. To find out where investors should be concerned and where they can feel confident, Joseph mapped the impact of agentic technology across his entire portfolio and several other popular stocks.

His core conclusion is that the market is currently pricing in the wrong primary risk. Investors are fixated on **automation**—fearing that automated processes will replace human labor and cause widespread job losses. While labor automation matters, Joseph argues that we have entered a different category entirely:

> "Automation is not the biggest risk today. Losing your customer is the risk. Losing contact with your customer."

The deeper economic transformation happens when software moves from merely answering a question to completing an end-to-end transaction. When an AI agent compares options, negotiates, books, pays, monitors the results, and learns the user's preferences over time, `"the agent becomes the demand gatekeeper."`

```mermaid flowchart TD accTitle: Agentic Commerce Shift and the Friction Framework accDescr: Illustrates how AI agents become the demand gatekeeper between users and businesses, and how companies are evaluated on friction removal versus scarce execution.

subgraph Gatekeeper["The Agent as Demand Gatekeeper"] U["Human User"] -->|"Delegates intent"| A["AI Agent (e.g., Meta Muse)"] A -->|"Compares, negotiates, books & pays"| M["Underlying Merchants & Services"] end

subgraph Framework["The Two-Axis Evaluation Framework"] direction LR V["High Friction Removal (Search costs, paperwork, inertia)"] -->|"Creates"| Vuln["High Vulnerability (Agent absorbs friction)"] S["Scarce Execution Remaining (Physical or unique capability)"] -->|"Creates"| Res["High Resilience (Defensible moat)"] end

A -.->|"Bypasses UI, ads & promotions"| V ```

[07:26](https://www.youtube.com/watch?v=HKskNSEb5_c&t=446s) The Friction Framework: Evaluating Vulnerability vs. Resilience

Citrini Research proposed a valuable stress test for any business: ask whether a company's real product is, `"I will navigate complexity that you find tedious."` If the answer is yes, a capable agent attacks the very reason that intermediary exists. Many companies exist solely to take a high-friction process and make it slightly less frictional. As Joseph puts it, `"Friction is the product until a machine absorbs the friction."` The more a business monetizes search costs, paperwork, confusion, inertia, and comparison fatigue, the more directly an AI agent attacks its profit pool.

To determine which stocks are safe and which are endangered, Joseph evaluates every company along two distinct axes:

1. **Vulnerability (Friction Removal):** How much of the company's value comes from removing tedious steps? 2. **Resilience (Scarce Execution):** How much scarce execution or capability remains after those tedious steps are automated?

Companies characterized by **high friction and low scarcity** sit at maximum exposure in the crosshairs of agentic commerce. At the top of this vulnerable list are routine travel comparison platforms—specifically **Booking Holdings**, a stock Joseph used to own and says he is even more concerned about today than he was when he sold it.

[09:00](https://www.youtube.com/watch?v=HKskNSEb5_c&t=540s) The Friction Framework: Mapping Disruption vs. Resilient Moats

When analyzing businesses against the incoming wave of autonomous agents, the core question is whether a company's business model relies on smoothing out friction or on owning genuinely scarce physical or technical assets. Businesses offering maximum friction reduction with low underlying scarcity are in the absolute blast zone: finding the best plane ticket, booking travel, generating templated legal forms, lead reselling, commodity freelance services, basic customer support, and simple tax filings (such as Intuit's TurboTax). In those domains, "they're taking something that's hard to navigate and they're removing friction"—an exact job description for a generalized AI agent that can both perform the labor and shop the underlying provider directly.

```mermaid flowchart TD accTitle: The Friction and Scarcity Exposure Matrix accDescr: A quadrant mapping business types based on their friction removal and scarcity levels to determine AI agent vulnerability.

subgraph HighFriction["High Friction Removal"] MaxEx["Maximum Exposure\n(Low Scarcity)\n• Online Travel Agencies\n• Intuit TurboTax\n• Lead Reselling & Basic Support"] InterfaceRisk["Interface at Risk, Network Survives\n(High Scarcity)\n• Uber\n• DoorDash\n(Physical logistics & driver networks)"] end

subgraph LowFriction["Low Friction Removal"] Tollgate["New Demand Drivers / Toll Collectors\n(Identity & Intent Owners)\n• Meta (Muse)\n• OS / Platform Ecosystems"] Resilient["Maximum Resilience\n(High Scarcity)\n• ASML (Semiconductor Lithography)\n• Visa & Mastercard\n• S&P Global & Moody's\n• Texas Roadhouse & Disney"] end

Moving down the vulnerability curve, companies with high friction and high scarcity—such as DoorDash and Uber—face interface risk, but their underlying systems survive. While autonomous bots can easily interact with customer-facing booking portals and bypass their direct app sessions, they cannot magically replicate physical driver and delivery networks. Losing that direct customer relationship is still a real headwind, but the back-end infrastructure prevents total disintermediation.

At the most resilient end of the spectrum sit high-scarcity, low-friction businesses: - **Semiconductor Lithography (ASML):** They do not profit from navigational headaches; they manufacture indispensable, highly complex physical equipment that the AI wave actively requires. - **Payment Rails (Visa & Mastercard):** They do not exist to simplify clunky UI; they operate mission-critical payment networks that transactions must traverse. - **Credit Rating Agencies (S&P Global, Moody's):** Regulated, deeply embedded institutional standards that cannot be substituted away by software. - **Physical Experiences & Scaled Dining:** The Cheesecake Factory, Texas Roadhouse, live entertainment venues, and Disney cannot be consumed digitally. - **Aggregators of Identity, Memory, and Intent (Meta):** By building tools like Muse, Meta aims to become the new gatekeeper. If a consumer shops across Nike, Target, Amazon, and local grocery stores via a single Meta prompt, Meta becomes the ultimate toll collector, monetizing sponsored rankings or taking a slice of transactions just as Amazon did with third-party sellers.

---

[11:57](https://www.youtube.com/watch?v=HKskNSEb5_c&t=717s) The Four Stages of the Agent Economy

The distinction between passive chatbot assistants like ChatGPT or Gemini and true agentic commerce must not be conflated. The transition unfolds across four progressive architectural phases, shifting the dynamic from informational queries to machine-governed markets:

```mermaid flowchart LR accTitle: The Four Stages of the Agent Economy accDescr: Progression from passive chatbots to autonomous machine-to-machine commerce.

S1["Stage 1: Conversational Assist\n• ChatGPT / Gemini era\n• User still executes on-site"] S2["Stage 2: Autonomous Action\n• Single-instruction execution\n• Traditional UI bypassed"] S3["Stage 3: Persistent Memory\n• Compounding preference lock-in\n• Durable demand profile"] S4["Stage 4: Machine-to-Machine\n• Dynamic pricing vs. agent negotiation\n• Standardized commerce protocols"]

S1 --> S2 --> S3 --> S4 ```

- **Stage 1 (Informational Assistance):** Where the market sits today. Chatbots write summaries, commoditize how-to guides, and recommend vacation itineraries, but execution remains strictly with incumbents: "A chatbot can recommend a flight while the user still books through an airline or an online travel agency." - **Stage 2 (Authorized Autonomous Execution):** Software acts on behalf of the user. An instruction like *"Find me a good hotel near the conference for under $350 with a quiet room and free cancellation"* is resolved entirely by the agent. It queries inventory, parses contractual terms, books the reservation, stores confirmation receipts, and updates the calendar without the user ever touching an online travel portal. - **Stage 3 (Persistent Memory & Distribution Lock-in):** Memory transforms convenience into an unassailable distribution advantage. The agent remembers the user prioritizes arrival reliability over saving $3 on delivery fees, despises red-eye flights, prefers aisle seats, avoids loud restaurants, and requires specific food allergy exclusions for their children. By eliminating the friction of setting up filters on future tasks, persistent memory builds a "durable demand profile" that permanently locks the user into the agent's ecosystem. - **Stage 4 (Machine-to-Machine Commerce):** The entire transaction layer shifts to autonomous negotiation. While airlines roll out dynamic pricing engines to capture maximum consumer surplus, humans will fight back by unleashing consumer agents to negotiate and arbitrage prices in real time: "We will put their agents on them and they will battle it out." The underlying infrastructure is already being engineered in plain sight through initiatives like Google's universal commerce protocol, Microsoft's 365 multi-agent workflows, and Visa and Mastercard's tokenized agent authorization frameworks.

---

[15:48](https://www.youtube.com/watch?v=HKskNSEb5_c&t=948s) The Death of Habitual Intermediation

Many dominant digital marketplaces masquerade as indispensable network effects when, in economic reality, they merely prey on cognitive fatigue. Consumers do not repeatedly use DoorDash, Uber Eats, or specific travel aggregators because of brand devotion; they use them because checking competing platforms, cross-referencing restaurant websites, researching pick-up discounts, and testing credit card promos takes too much time. The cost of manual comparison consistently outstrips the expected cash savings.

AI agents completely break this dynamic of "habitual intermediation": - **Human Defaults:** Consumers open whichever app is installed on their home screen, browse ranked listings, tolerate hidden service fees, accept dynamic surcharges, and check out with saved credentials. - **Agent Defaults:** The agent queries every authorized merchant simultaneously, normalizes opaque service fees, objectively balances speed versus quality, and routes orders strictly to the best-fit economic provider. - **Erosion of Fake Loyalty:** While rewards points and subscription bundles (like DashPass or Amazon Prime) provide temporary buffers, an agent evaluates mathematical realities rather than psychological branding: "It turns soft loyalty into a consciously tested economic relationship." Incumbents relying on UI convenience, friction-heavy comparisons, and sponsored search ad listings will find their pricing power systematically dissolved.

[18:10](https://www.youtube.com/watch?v=HKskNSEb5_c&t=1090s) The Elimination of Consumer Inertia and the "Gatekeeper Swap"

A massive portion of consumer commerce is dictated purely by inertia. People repeatedly choose the same services simply because switching to an alternative feels slightly too annoying. When autonomous agents enter the picture, that friction disappears entirely. Backend products and service providers will no longer benefit from consumer laziness; instead, they will be forced to compete on a continuous, transaction-by-transaction basis to retain every customer relationship. Moving past theoretical mechanics, this structural shift provides a much more rigorous framework for evaluating corporate winners and losers than the superficial commentary put out by short-sellers like Citron Research.

The first major casualty in this transition could be Booking Holdings (parent company of Booking.com and Kayak). Travel serves as the ultimate litmus test for agentic commerce because options are highly structured, comparison shopping is genuinely valuable, and consumers are already conditioned to rely on intermediaries to assemble flights, lodging, and car rentals. The fragility of Booking’s business model is visible directly in its financials: in 2025, the company reported $8.1 billion in marketing expense, representing roughly 30% of its total revenue. For every $10 Booking brings in, $3 is immediately handed over to customer acquisition channels. While Booking enjoys global supply, deep conversion data, payment rails, and customer loyalty through its Genius program, it remains dependent on spending billions just to capture travel intent.

```mermaid flowchart TD accTitle: The Travel Gatekeeper Swap accDescr: Diagram illustrating how consumer travel booking shifts from Google search and OTAs to personal AI agents.

subgraph AgenticModel["Agentic Commerce Funnel"] E["Traveler"] -->|"Natural Language Prompt"| F["Personal AI Agent"] F -->|"Direct Query / Best Terms"| G["Direct Supplier / Booking API"] G --> D end ```

The existential hazard for Booking Holdings is a profound "gatekeeper swap." In the legacy internet model, travel economics flowed through search engines: whenever a user expressed travel intent on Google, Booking bid aggressively to purchase that user's click. In an agentic paradigm, a traveler states their intent just once to a personal assistant. The assistant asks clarifying follow-up questions, evaluates cancellation preferences, factors in loyalty memberships, and handles booking end-to-end. Even though travel has notoriously "ugly expectations"—canceled flights, misleading room photos, broken payment gateways, conflicting local tax laws, and stubborn refund policies—an AI agent might treat Booking as nothing more than a back-end utility API. If an agent can route around Booking or force it to compete purely inside an agentic interface, Booking loses control over the marginal traveler. Its billions in marketing spend may not evaporate, but it will shift into sponsored recommendation slots within AI platforms or manifest as sharply compressed net take rates.

- **Booking Holdings 2025 Marketing Spend:** $8.1 billion, or approximately 30% of gross revenue ($3 out of every $10 earned). - **The Core Vulnerability:** Lack of physical logistics infrastructure; value is tied almost entirely to demand aggregation and interface control. - **The Bear Case Mechanism:** Not an immediate collapse in gross booking volume, but the loss of interface ownership, reducing the company from a high-margin consumer destination to a low-margin backend fulfillment API.

---

[21:45](https://www.youtube.com/watch?v=HKskNSEb5_c&t=1305s) Physical Execution vs. Coordination: DoorDash vs. OpenTable

The contrast between demand aggregators highlights why physical infrastructure is the ultimate moat against agentic disruption. OpenTable and DoorDash both aggregate consumer demand, but their structural resilience could not be more divergent. OpenTable operates as a pure coordination and scheduling database; it owns no physical layer, making it extremely vulnerable to an agent that can easily contact, reserve, and negotiate dining slots across competing platforms or directly with restaurants.

DoorDash, by contrast, is anchored to the messy, physical world. While an AI agent can certainly reach outside of the DoorDash interface to cross-check menu pricing against Uber Eats, restaurant direct ordering, or local pickup discounts—chipping away at habitual in-app browsing—the actual delivery cannot be solved by a database lookup. Fulfilling an order requires local driver density, intelligent dispatch algorithms, order batching, route optimization, fraud prevention, point-of-sale merchant integrations, prep instruction tracking, refund mitigation, and live customer service recovery. DoorDash’s ongoing margin improvements have been driven directly by operational gains, including logistics efficiencies, high-margin advertising, and lower error/refund rates.

While losing exclusive interface control introduces real intermediary risk—a dynamic reflected in recent stock price softness as investors price in platform risk—DoorDash remains significantly more defensible than asset-light booking portals because someone still has to pick up the food and bring it to the front door.

- **OpenTable Exposure:** High vulnerability due to operating as a friction-removal coordination layer with no proprietary physical delivery asset. - **DoorDash Defense:** High resilience rooted in dense physical driver networks, batching economics, and merchant integrations that cannot be replicated by pure software. - **Intermediary Headwinds:** Agents can bypass habitual app opening to compare prices across delivery platforms and direct restaurant ordering, pressuring top-of-funnel customer lock-in.

---

[23:30](https://www.youtube.com/watch?v=HKskNSEb5_c&t=1410s) Intuit and the Shift to Ambient Tax Preparation

Tax preparation is an industry primed for radical compression as agentic workflows transition from episodic tasks to ambient, continuous background processes. For standard filers, the primary pain point has always been gathering tax forms, answering repetitive questionnaire prompts, mapping life circumstances to arcane regulatory rules, and filing standard documents. TurboTax has historically monetized this anxiety by charging $100 to $200 for what is fundamentally a glorified, animated digital questionnaire backed by seasonal support.

When a persistent financial AI agent maintains continuous access to bank accounts, payroll platforms, and expense receipts, tax compliance shifts from a dreaded annual ordeal into an ambient background calculation. The agent tracks deductible events in real time and compiles forms automatically, destroying the perceived value of guided form-filling software.

Intuit is attempting to defend its franchise by rolling out proprietary AI capabilities, but the software premium is evaporating quickly. Having previously exited a position in Intuit due to early doubts about TurboTax's longevity, the progression of autonomous financial agents makes the outlook for Intuit shareholders far more alarming today. Both TurboTax and consumer lead-generation properties like Credit Karma face severe margin compression as autonomous tools make financial optimization accessible without high-fee software suites.

- **The TurboTax Model:** Charging consumers $100–$200 for guided questionnaire completion and automated form generation. - **Ambient Preparation:** Persistent AI agents eliminate seasonal friction by reconciling financial events, receipts, and deductions continuously throughout the calendar year. - **Portfolio Assessment:** Intuit’s core consumer cash cows (TurboTax and Credit Karma) face deep structural pricing compression as generalist agents match their functional capabilities at negligible cost.

---

[25:01](https://www.youtube.com/watch?v=HKskNSEb5_c&t=1501s) LegalZoom, Continuous Insurance Shopping, and the Threat to Zillow

The deflationary impact of agentic workflows spreads rapidly across legal templates, programmatic paperwork, and lead-generation brokers. Companies like LegalZoom, which monetize basic legal formations, boilerplate contracts, standard wills, and compliance filing reminders, are exceptionally vulnerable. Large language models and agents already excel at synthesizing legal prose, spotting drafting risks, and drafting bespoke variations of corporate paperwork. Routine white-collar outputs—such as basic contract review, logo variations, first-pass copywriting, transcription summaries, spreadsheet cleaning, and boilerplate code translations—can now be delegated directly to an agent with minimal oversight.

```mermaid flowchart LR accTitle: Friction Removal in Insurance and Real Estate accDescr: Comparison of legacy customer retention through friction versus agentic continuous optimization.

subgraph LegacyFriction["Legacy Market: Consumer Inertia"] A1["High Complexity / Tedious Shopping"] --> B1["Consumer Signs Policy / Locks In"] B1 --> C1["Years of Inaction (Inertia Margin)"] end

subgraph AgenticContinuous["Agentic Market: Continuous Arbitrage"] A2["Persistent AI Agent"] --> B2["Real-Time Market Scrapes"] B2 --> C2["Automatic Switch at Lowest Rate"] end ```

Lead generation marketplaces centered around insurance—including SelectQuote, GoHealth, and EverQuote—are built around information asymmetry, convoluted policy terms, and the reality that consumers hate shopping for insurance once a policy is active. That pricing power collapses under continuous monitoring. In one real-world test shared on X, a user uploaded an existing insurance policy to an AI agent, and within 20 minutes the system scanned the market and surfaced an identical policy saving $600 per year. When comparison shopping requires zero human labor, insurers can no longer count on customer apathy to protect high-margin renewals.

Real estate marketplaces like Zillow face a parallel crisis. A capable multimodal housing agent—such as Meta Muse—can monitor MLS feeds around the clock, weigh complex personal lifestyle trade-offs, automatically schedule property viewings, analyze neighborhood school metrics, track historical price adjustments, and parse municipal disclosures. Because an integrated assistant can query multiple listing services and manage communications directly, portals that function primarily as search front-ends and lead-generation funnels for real estate agents will struggle to justify their place in the value chain.

- **LegalZoom & Document Assembly:** Highly susceptible as entity formations, simple wills, compliance reminders, and contract reviews become natively handled by generalist agents. - **Insurance Lead Gen (SelectQuote, GoHealth, EverQuote):** Business models reliant on customer exhaustion fall apart when agents can parse complex insurance terms and extract hundreds of dollars in annual savings in minutes. - **Real Estate Portals (Zillow):** Threatened by dedicated housing agents that autonomously filter listings, weigh trade-offs, gather disclosure documents, and coordinate property tours directly with listing brokers.

[27:05](https://www.youtube.com/watch?v=HKskNSEb5_c&t=1625s) Alphabet: Defending the Search Moat in an Agentic Paradigm

Moving from smaller software players to mega-cap portfolio cornerstones, Joseph Carlson addresses his single largest holding: Alphabet (Google), sitting as a $24,000 position in his portfolio with roughly $110,000 in unrealized gains. Evaluating how Alphabet fares in an agent-driven world reveals a nuanced mix of major headwinds and underappreciated defensive strengths. Alphabet reported $224 billion in Google Search and other revenues in 2025 (up from $198 billion in 2024), meaning that even a minor fractional disruption to commercial intent search volumes carries massive dollar consequences.

The primary threat lies in the collapsing query funnel: traditional search monetizes an extended user session—clicking multiple links, browsing comparative options, and hitting successive ad impressions for high-value terms like "best hotel in Chicago," "cheaper car insurance," or "tax software for freelancers." When a user delegates that workflow to an agent like Meta's Muse or Google's own Gemini, the system gathers criteria once, pings providers directly through APIs, and returns one or two direct recommendations. Monopolizing consumer attention through browser friction declines, slashing paid ad clicks.

```mermaid flowchart TD accTitle: The Search Funnel Shift accDescr: Comparison between legacy search ad journeys and collapsed agentic query workflows. subgraph Legacy["Traditional Search Model"] A["User Query"] --> B["10 Blue Links & Search Ads"] B --> C["Multiple Ad Impressions & Clicks"] C --> D["Manual Site Exploration & Purchase"] end subgraph Agentic["Agentic Query Flow"] E["Delegated Prompt"] --> F["Agent Parses Requirements"] F --> G["Structured Backend API Queries"] G --> H["1-2 Curated Recommendations / Direct Buy"] end ```

Despite this threat to traditional blue-link search ads, Alphabet holds critical structural moats that make it significantly more insulated than specialized point-solution software: - **Massive Distribution:** Ubiquitous touchpoints across Android, Chrome, Google Search, Gmail, Google Maps, and YouTube. - **Deep Intent Signals:** Unmatched contextual understanding mined from location data, video consumption, emails, and global queries to ground agent actions. - **Full-Stack Technical Control:** Foundation models like Gemini, hyperscale cloud infrastructure, enterprise developer toolsets, and integration pipelines. - **The Universal Commercial Protocol:** Direct infrastructure designed to bridge discovery, merchant checkout, and post-purchase customer servicing seamlessly. - **Reinvented Ad Units:** Ads can evolve from traditional sponsored blue links into machine-readable bids for an agent's consideration set, sponsored recommendations, and preferred fulfillment incentives.

---

[29:47](https://www.youtube.com/watch?v=HKskNSEb5_c&t=1787s) Amazon, Uber, and Duolingo: Execution Moats vs. Friction Vulnerabilities

Turning to an eclectic trio—Amazon, Uber, and Duolingo—Carlson analyzes how each company loses a specific traditional advantage, yet preserves distinct operational moats against automated software agents.

```mermaid flowchart LR accTitle: Platform Vulnerability vs Core Defensibility accDescr: How Amazon, Uber, and Duolingo balance front-end agent risks against operational back-end moats. subgraph Amazon A1["Front-End Search & Ads At Risk"] --> A2["Moat: Prime, Logistics, AWS, Fulfillment"] end subgraph Uber U1["Aggregator Comparison Risk"] --> U2["Moat: Real-Time Fleet Density & Physical Dispatch"] end subgraph Duolingo D1["Functional Translation Redundant"] --> D2["Moat: Gamification, Streaks, Personal Identity"] end ```

Amazon faces genuine margin pressure at the top of its funnel. When shopping agents bypass Amazon's on-site search bar to programmatically compare prices across brand storefronts, Walmart, niche merchants, resale hubs, and local inventories, Amazon's role as the default starting point for online commerce weakens. This threatens its high-margin sponsored product ad listings. However, agents prioritize guaranteed transaction success. Because Amazon possesses unmatched third-party seller selection, Prime member loyalty, ultra-fast delivery, effortless return logistics, integrated payment infrastructure, reliable customer reviews, and AWS hosting, autonomous agents will often still choose Amazon to fulfill orders. While search ad monetization faces headwind pressure, back-end volume gains ground through superior execution.

Uber sits in an even safer position. While an autonomous assistant can scan Uber, Lyft, Waymo, and Tesla simultaneously to arbitrate ride prices, fulfillment requires physical assets. The agent will ultimately pick the optimal blend of vehicle arrival time, safety, rider preferences, and cost. Uber’s immense marketplace liquidity and driver density guarantee that it remains the most dependable option in the vast majority of scenarios.

Duolingo presents a unique dynamic: an agent fluent in every spoken language reduces the purely practical, transactional need to learn a foreign language. However, Duolingo's primary business is not resolving a translation utility problem—it monetizes habit-forming entertainment, personal identity, and intellectual achievement. Users engage with streaks, leagues, avatar designs, and curated curricula because "users often want to become a person who speaks a language, not merely solve a translation problem." Consequently, Duolingo remains far more durable than pure utility platforms.

---

[32:15](https://www.youtube.com/watch?v=HKskNSEb5_c&t=1935s) Microsoft and Meta: Enterprise Operating Systems vs. Attention Engines

Carlson highlights Microsoft and Meta as major holdings in his portfolio that act as net beneficiaries of the agentic era rather than victims, drawing a sharp contrast with friction-reliant platforms like Intuit and Booking Holdings.

Microsoft operates as the foundational "agentic operating system." Under Satya Nadella's forward-looking stewardship, the company has pre-empted disruption by embedding multi-agent workflows into its core enterprise ecosystem: - **Foundry Agent Architecture:** Hosting multi-agent orchestration, persistent long-term memory, and deep Microsoft 365 workflow automation. - **Permissioned Context Moat:** The critical enterprise barrier determining identity, document permissions, compliance controls, corporate spending authority, and audit tracking across teams. - **Cross-Application Integration:** Coordinating tasks seamlessly across work identities, Outlook emails, Word documents, Teams meetings, Windows, GitHub developer tools, and Azure cloud infrastructure.

Meta approaches the agent ecosystem from the opposite pole: social identity, creator culture, messaging, and consumer attention. While Google monetizes declared commercial intent when a user actively solves a problem, Meta monetizes organic discovery within an entertainment environment. An autonomous agent that removes purchase friction actually *boosts* the conversion value of feed-based ads: a user discovers an item casually on Instagram Reels, taps to purchase, and hands the checkout and delivery management over to an agent like Meta's Muse. Because browsing social video feeds is driven by entertainment rather than friction, Meta's ad impression inventory remains intact while agent interactions provide richer behavioral data to train ad-targeting algorithms.

---

[35:14](https://www.youtube.com/watch?v=HKskNSEb5_c&t=2114s) Visa and Mastercard: The Critical Need for Transaction Trust

Addressing the payment rails, Carlson pushes back forcefully against short-seller commentary (such as Citron Research) claiming that Visa and Mastercard will be disintermediated by artificial intelligence. Holding an active $189,000 position in Mastercard, Carlson argues that an explosion of autonomous software actors dramatically increases—rather than decreases—the systemic necessity for centralized, trusted transaction infrastructure.

When machines begin buying goods on behalf of humans without real-time manual sign-off, trust architecture becomes paramount: - Determining whether an automated agent had legitimate authorization to initiate a payment. - Enforcing dynamic spend thresholds and user-defined budget rules. - Controlling tokenized credential delivery to prevent merchants or malicious bots from storing raw payment data. - Establishing rigorous auditability to track, dispute, or reverse unauthorized automated charges. - Helping merchants differentiate legitimate automated consumer agents from automated fraud rings.

Visa and Mastercard are positioning themselves directly at this chokepoint through programmatic spending controls, identity tokenization, and dedicated frameworks like Mastercard's "Agent Pay." By turning automated agents into reliable economic actors, the payment giants stand to extract tolls on an expanding universe of autonomous micro-transactions.

[36:12](https://www.youtube.com/watch?v=HKskNSEb5_c&t=2172s) Payment Rails and Hardware Bottlenecks: Mastercard and ASML

Mastercard and Visa operate deep within the financial plumbing, using tokenization, passkey integrations, and infrastructure partnerships with companies like Microsoft, IBM, Braintree, and Checkout.com. These capabilities are natural extensions of their established network roles. Crucially, neither payment network cares about winning the consumer-facing user interface; their business is solely focused on owning the underlying transaction layer. As AI agents step in between consumers and purchases, human supervision becomes more detached, which actually heightens the necessity for institutional trust and secure execution. Because autonomous agent purchases introduce higher transaction volumes and an even greater need for fraud prevention and settlement integrity, the upside potential for Mastercard meaningfully outweighs any downside threat, making it an easy holding to maintain.

```mermaid flowchart LR accTitle: Agentic Commerce Value Chain accDescr: Diagram illustrating how agents drive demand into hardware bottlenecks and trusted payment rails.

subgraph Demand ["Consumer & Agent Layer"] A["User Intent"] --> B["AI Agent Interface"] end

subgraph Infrastructure ["Enabling Bottlenecks"] B -->|"Compute & Inference"| C["ASML Lithography"] B -->|"Autonomous Settlement"| D["Visa & Mastercard Rails"] end ```

Moving into physical hardware, ASML represents an impenetrable bottleneck. Agents cannot write software routines that route around leading-edge extreme ultraviolet (EUV) lithography. Autonomous software inherently consumes massive training and inference capacity, creating relentless demand for cutting-edge logic and high-bandwidth memory chips. ASML's competitive moat is cemented by decades of accumulated physical engineering, intricate supplier coordination, global service networks, and deep customer manufacturing integrations. A general-purpose AI model cannot synthesize physical fabrication capabilities out of thin air. ASML remains the uncontested "apex predator" at the very top of the semiconductor food chain—impervious to agent disruption, while directly profiting from the surging computational demand that agents require.

- **Visa & Mastercard:** Secure the transaction layer without needing to own the consumer assistant's brand or front-end interface. - **ASML:** Enables the physical computation underpinning AI; completely protected by physical engineering moats that software models cannot replicate.

---

[38:11](https://www.youtube.com/watch?v=HKskNSEb5_c&t=2291s) The Untouchables: S&P Global, Costco, Netflix, and Texas Roadhouse

When evaluating companies like S&P Global, Moody's, Costco, Netflix, and Texas Roadhouse, AI agents pose minimal competitive threat because these businesses do not derive their core value from mere friction removal. In the financial data sector, credit ratings are not just simple summaries of publicly available information. They are highly regulated benchmarks mandated by debt covenants, contractual frameworks, institutional surveillance, and decades of legal precedent. While generic market summaries have become cheap commodities (the host jokes that he offers one via his own platform for $10 a month), multi-billion-dollar credit funds and Tier 1 lending institutions cannot rely on unverified scrapes. They legally require "decision-grade," licensable, auditable data backed by recognized institutional authority. S&P Global's subscription model—spanning proprietary indices, valuation metrics, and specialized risk data—experiences rising demand as autonomous models query their archives for reliable ground truth.

```mermaid flowchart TD accTitle: Moat Archetypes in the Agentic Economy accDescr: Categorization of business models based on physical, contractual, or intellectual moats.

subgraph Regulatory ["Contractual & Benchmark Moat"] SP["S&P Global / Moody's"] FICO["FICO"] end

subgraph Physical ["Operational & Physical Supply"] COST["Costco Wholesale"] TXRH["Texas Roadhouse"] end

subgraph IP ["Proprietary Licensing"] NFLX["Netflix Catalog"] end ```

Costco, Netflix, and experiential businesses enjoy a similar degree of insulation:

* **Costco:** An AI shopping assistant can readily compare prices, but it cannot replicate Costco's physical procurement volume, curated SKU selection, Kirkland Signature private-label margin advantages, or warehouse membership economics. An agent simply acts as a household replenishment tool, identifying durable bulk value and routing repeat orders directly into Costco's infrastructure. * **Netflix:** While a cross-platform assistant can recommend what to stream (slightly blunting Netflix’s native home-screen discovery advantage), it cannot legally deliver copyrighted media. The agent must still route through an active commercial subscription to access Netflix's proprietary, globally licensed entertainment catalog. * **Texas Roadhouse and Physical Real Estate:** A dining concierge bot can check reservation wait times, secure a booth, or highlight menu items, but software cannot grill a hand-cut steak, foster restaurant hospitality, or recreate an energetic dining room atmosphere. The same protective barrier shields gyms, live sporting events, theme parks, and concert venues. Agents optimize booking logistics, which ultimately drives higher physical foot traffic and table utilization to top-performing operators.

---

[41:14](https://www.youtube.com/watch?v=HKskNSEb5_c&t=2474s) The Exposure Spectrum: Winners, Losers, and the Future of Interface Gatekeepers

The impending transition to agentic commerce presents vastly divergent risk profiles depending on whether a company's primary asset is a defensible physical or legal moat, or simply an intermediary user interface. Pure software middlemen that monetize search discovery and basic data aggregation face the highest structural risk. However, nuances exist within high-exposure categories: DoorDash has physical driver networks and merchant fleet integration, making it considerably more resilient than pure reservation directories like OpenTable. Similarly, tax preparation services like Intuit face existential risk if autonomous agents turn annual tax filing into an ambient, automated background task that parses receipts in real time.

| Category | Representative Companies | Vulnerability & Growth Thesis | |:--- |:--- |:--- | | **High Exposure** | OpenTable, Intuit, Booking Holdings | Interfaces are easily bypassed; software agents capture intent and automate administrative tasks directly. | | **Moderate / Mixed Exposure** | DoorDash, Uber, Duolingo, Alphabet (Google Search), Amazon (Ad Business) | Facing ad discovery compression and UI displacement, but backed by physical delivery logistics, network effects, or core platform utilities. | | **Resilient Moats** | FICO, S&P Global, Moody's, Costco, Netflix, Texas Roadhouse, The Cheesecake Factory | Protected by statutory mandates, decision-grade authority, proprietary content libraries, or offline physical dining and retail experiences. | | **Direct Beneficiaries** | Microsoft, Meta, Shopify, ASML, Visa, Mastercard | Providing model compute, hardware manufacturing, agent runtime infrastructure, merchant tools, or global transaction rails. |

The timeline for these market shifts is measured in months rather than decades, as millions of consumers begin adopting agentic shopping workflows. The ultimate test of this thesis rests on convenience: if agent technology underperforms, consumers will stick with classic direct web apps. Yet historical consumer behavior overwhelmingly demonstrates that whichever channel removes friction and delivers the fastest outcome always wins. Because autonomous software makes executing multi-step tasks easier than navigating cluttered commercial portals, agentic commerce represents a permanent, structural shift in digital economics.

## Notable quotes > [00:00](https://www.youtube.com/watch?v=HKskNSEb5_c&t=0s) friction going to zero — Joseph Carlson

> [05:43](https://www.youtube.com/watch?v=HKskNSEb5_c&t=343s) Automation is not the biggest risk today. Losing your customer is the risk. Losing contact with your customer. — Joseph Carlson

> [05:43](https://www.youtube.com/watch?v=HKskNSEb5_c&t=343s) the agent becomes the demand gatekeeper. — Joseph Carlson

> [07:26](https://www.youtube.com/watch?v=HKskNSEb5_c&t=446s) I will navigate complexity that you find tedious. — Joseph Carlson

> [07:26](https://www.youtube.com/watch?v=HKskNSEb5_c&t=446s) Friction is the product until a machine absorbs the friction. — Joseph Carlson

> [11:57](https://www.youtube.com/watch?v=HKskNSEb5_c&t=717s) A chatbot can recommend a flight while the user still books through an airline or an online travel agency. — Joseph Carlson

> [11:57](https://www.youtube.com/watch?v=HKskNSEb5_c&t=717s) We will put their agents on them and they will battle it out. — Joseph Carlson

> [15:48](https://www.youtube.com/watch?v=HKskNSEb5_c&t=948s) It turns soft loyalty into a consciously tested economic relationship. — Joseph Carlson

> [29:47](https://www.youtube.com/watch?v=HKskNSEb5_c&t=1787s) users often want to become a person who speaks a language, not merely solve a translation problem. — Joseph Carlson

## Mentioned AI, Alphabet, Amazon, Amazon Prime, Android, Anthropic, ASML, AWS, Azure, Booking Holdings, Booking.com, Braintree, ChatGPT, Checkout.com, Chrome, Citrini Research, Citron Research, Claude, Costco Wholesale, Credit Karma, DashPass, Disney, DoorDash, Duolingo, EverQuote, Expedia, FICO, Gemini, GitHub, Gmail, GoHealth, Google, Google Maps, Google Search, Grok, IBM, Instagram Reels, Intuit, Joseph Carlson, Kayak, Kirkland Signature, LegalZoom, Lyft, Mastercard, Mastercard Agent Pay, Meta, Meta Muse, Microsoft, Microsoft 365, Moody's, Netflix, Nike, OpenAI, OpenTable, Outlook, Qualtrim, S&P Global, Salesforce, Satya Nadella, SelectQuote, ServiceNow, Shopify, Smith's, Teams, Tesla, Texas Roadhouse, The 2028 Global Intelligence Crisis, The Cheesecake Factory, TurboTax, Uber, Uber Eats, Visa, Walmart, Waymo, Windows, Word, YouTube, Zillow

## Video description AI agents are moving from simply answering questions to actually doing things for us — shopping, booking hotels, making reservations, comparing prices, managing transactions, and eventually negotiating with other machines on our behalf.

That shift could have enormous consequences for some of the world’s biggest companies.

In this episode, I look at the rise of agentic commerce and a simple framework for determining which businesses are most exposed: How much of their value comes from removing friction, and how much scarce infrastructure or capability remains once that friction disappears?

We break down what this could mean for Booking Holdings, DoorDash, Intuit, Google, Amazon, Uber, Duolingo, Microsoft, Meta, Visa, Mastercard, ASML, S&P Global, Costco, Netflix, Texas Roadhouse, and more.

Some businesses may lose their direct relationship with the customer. Others could become even more valuable as AI agents increase transactions, computing demand, and the need for trusted infrastructure.

The question may no longer be, “Can AI automate this business?”

It may be: “What happens when an AI agent becomes the customer’s gatekeeper?”

Join Qualtrim, the stock analysis platform I built and use, and join over 13,000 other paying members: https://www.qualtrim.com/

0:00 The Agentic Commerce Thesis 2:03 Meta Muse: The First Real Consumer Agent 5:43 How AI Agents Could Reshape Business 7:26 The Friction Framework: Who Is Most Exposed? 11:57 The Four Stages of the Agent Economy 18:28 Booking Holdings, DoorDash & Intuit 27:12 Google, Amazon, Uber & Duolingo 32:15 Microsoft, Meta, Visa & Mastercard 36:50 ASML, S&P Global, Costco, Netflix & Texas Roadhouse 41:14 Which Companies Are Most Exposed to AI Agents?

-Disclaimer Some of the links below are affiliate links, I can earn money from them at no cost to you.

This content is not a solicitation, is not endorsed by M1, and was not reviewed by M1; the opinions expressed are solely those of the authors and do not reflect M1's views. Information presented is accurate as of the video posting date; for the most up-to-date information, please refer to m1.com. Before making any investment decisions, consult your personal investment, legal, and tax advisors, as this content is for informational purposes only and not intended as investment recommendations.

▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀ Join Here: https://www.qualtrim.com/

📚 My favorite Investing Books: https://amzn.to/3KwyIhG 📷 All the tech I use to record videos: https://www.amazon.com/shop/josephcarlson/list/2L5YMP4DFYA8O?ref_=cm_sw_r_cp_ud_aipsflist_aipsfjosephcarlson_FAKKKMYKF486ZR6H2DSQ

🚀 Growth Portfolio: https://click.linksynergy.com/deeplink?id=5mNcifgFllM&mid=50362&murl=https://dashboard.m1.com/share?token=39d17ae6-4a4a-3fc4-bb15-d86abe3fbb67 💵 Dividend Portfolio: https://click.linksynergy.com/deeplink?id=5mNcifgFllM&mid=50362&murl=https://dashboard.m1.com/share?token=3b283352-1bff-31d9-8576-f5770c2bfdfd

SOCIAL MEDIA ▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀ 🎥 More free content: https://www.youtube.com/channel/UCbta0n8i6Rljh0obO7HzG9A 🐦 I post random thoughts on Twitter too: https://twitter.com/joecarlsonshow

DISCLAIMER ▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀ I am not a professional investor and have never claimed to be. I'm an amateur investor sharing my experience of what I've learned, where I have had success, and where I've had failures. I share my thoughts on investing and performance with transparency. My approach and goal to investing is to buy high-quality long-term investments in world-class businesses that I call "compounders". I view my investments as businesses, not as stocks. Before creating content on YouTube full time I worked as a senior-level programmer for 8 years. Over the years as a programmer, I compounded my knowledge of development. I take the same iterative learning approach to my study of investing. I study investing as a craft in the continual pursuit of being better. I will make mistakes in investment decisions from time to time. Results are not guaranteed. Please do not blindly follow me into any investments, and make sure your portfolio and investments are built around your specific income, risk tolerance, personality, timeline, and overall circumstances.

[Source](https://www.youtube.com/watch?v=HKskNSEb5_c)

中文简报

AI代理绕过消费者门户，Booking与Intuit等平台面临威胁随着自主代理成为需求把关人，靠交易摩擦获利的企业面临利润率崩溃，定价权正转向支付通道与实体物流网络。 Carlson分析了向代理商业的转型，并通过Meta Muse等工具展示了自主AI代理正超越被动聊天，开始执行端到端的采购、预订和比价。他指出，对商业模式的真正威胁并非劳动力自动化，而是直接界面所有权的丧失，因为代理成为了消费者把关人，直接绕过了传统搜索广告、赞助列表以及依赖交易摩擦的聚合平台。在评估企业究竟是消除摩擦还是具备稀缺的实体或监管执行力时，他指出Booking Holdings、OpenTable和Intuit TurboTax等轻资产中介面临严重冲击。相反，依托实体物流、监管强制要求、计算硬件或关键任务支付通道的企业——例如DoorDash、Uber、标普全球、ASML、Visa和Mastercard——则依然具备防御力，或有望捕获不断扩大的交易量。 Meta Muse等实用型消费级AI代理的部署，实现了直接在商家网站上自主执行多步骤操作（包括食品杂货订购、行程规划和预订），完全绕过了中介应用程序和搜索界面。 (2:03) Booking Holdings面临消费者把关人更替带来的严峻结构性风险，对其轻资产商业模式构成威胁；该公司在2025年仅为了捕获旅游搜索意图就花费了81亿美元（约占总营收的30%）。 (18:10) 纯软件协调平台与实体履约网络之间存在关键分水岭：OpenTable等平台极易被代理绕过，而DoorDash和Uber则凭借线下的司机密度和物流履约能力维持着可防御的护城河。 (21:45) Visa、Mastercard和ASML构成了核心结构性受益方，它们分别提供了机器对机器支付必不可少的身份与授权通道，以及支持代理计算需求所需的EUV光刻硬件。 (36:12)

Reading summary

Joseph Carlson argues that the primary market disruption from artificial intelligence agents is not labor automation, but the loss of direct customer contact as autonomous software becomes the primary demand gatekeeper. Under this thesis, businesses that profit primarily by absorbing human friction face severe disintermediation, whereas companies anchored to scarce physical infrastructure, regulatory mandates, or transaction settlement remain highly defensible. The Demand Gatekeeper and the Friction Framework (5:43) Carlson observes that consumer artificial intelligence is advancing from informational chatbots to autonomous agents capable of executing multi-step workflows. In his evaluation of Meta's Muse app, the agent demonstrated the ability to create allergy-compliant weekly meal plans, apply digital coupons, and reserve restaurant tables directly on merchant websites without requiring the user to open native applications. This shift transforms software from an advice tool into an intermediary that directly negotiates, purchases, and bypasses sponsored app interfaces. To gauge investment risk, Carlson introduces a two-axis framework assessing friction removal against scarce execution capability. Citrini Research's viral thesis that 'friction is going to zero' highlighted the vulnerability of companies whose sole product is navigating tedious interfaces. When AI agents absorb routine administrative tasks, form-filling, and price comparisons, the primary danger to businesses shifts away from labor automation toward losing customer touchpoints entirely. The Erosion of Pure Digital Middlemen (18:10) Companies that monetize search costs and administrative complexity without owning scarce assets sit at maximum exposure. Carlson highlights Booking Holdings as an example of an impending 'gatekeeper swap.' In 2025, Booking spent $8.1 billion—roughly 30% of its revenue—acquiring search traffic. If personal agents bypass commercial search funnels to query providers directly, Booking risks being demoted from a high-margin travel portal to a low-margin backend fulfillment utility. Administrative and lead-generation software faces comparable compression. Intuit's TurboTax has historically charged $100 to $200 for guided digital questionnaires, but Carlson projects persistent financial agents will turn tax compliance into an ambient background process that tracks deductions in real time. Similarly, platforms like LegalZoom for basic legal filings, and insurance lead brokers like SelectQuote, lose pricing power when autonomous agents can continuously draft contracts or scan policies to eliminate the margins previously protected by consumer inertia. Interface Risk Versus Physical Execution Moats (21:45) Organizations combining high friction removal with physical assets encounter interface compression, but retain defensible operational moats. In contrast to OpenTable's pure reservation directory, DoorDash and Uber rely on physical driver networks, route optimization, and localized logistics that software agents cannot replicate. While agents can bypass apps to compare delivery fees across platforms, the physical fulfillment network prevents total disintermediation. Similar dynamics appear among mega-cap tech platforms. Amazon faces pressure on its high-margin sponsored search advertisements as shopping bots bypass on-site search, but its fulfillment footprint, Prime loyalty, and return logistics make it the agent's preferred execution partner. Alphabet confronts risks to Google Search—which generated $224 billion in 2025—as conversational prompts collapse multi-click journeys, yet Carlson maintains that Alphabet remains insulated by its Android distribution, Gemini foundation models, and universal commerce protocols. Bottlenecks, Regulatory Standards, and Settlement Rails (35:14) At the most resilient end of Carlson's framework are businesses shielded by physical bottlenecks, institutional authority, and payment rails. ASML operates as an indispensable hardware choke point, protected by extreme ultraviolet lithography engineering that software cannot bypass while benefiting from rising agent compute demand. Carlson also pushes back against short-seller claims that Visa and Mastercard will be displaced, arguing that machine-to-machine commerce escalates the need for tokenized credentials, spending limits, fraud prevention, and auditability through frameworks like Mastercard's 'Agent Pay.' Regulatory and physical moats offer equivalent protection. S&P Global and Moody's supply credit ratings and financial benchmarks mandated by debt covenants and institutional standards, which cannot be substituted by general web scrapes. Concurrently, physical experiential businesses like Costco, Texas Roadhouse, live entertainment venues, and proprietary media libraries like Netflix remain insulated because autonomous software can optimize reservations or replenishment, but cannot digitize the underlying physical product. Carlson concludes that the transition toward agentic commerce is accelerating rapidly because consumers consistently adopt the path of lowest friction. While platforms backed by physical networks, specialized hardware, and regulatory mandates are well positioned to capture value, pure software intermediaries that rely on interface convenience and customer fatigue face permanent structural compression.

Source quotes

[0:00] Joseph Carlson: friction going to zero [5:43] Joseph Carlson: Automation is not the biggest risk today. Losing your customer is the risk. Losing contact with your customer. [5:43] Joseph Carlson: the agent becomes the demand gatekeeper. [7:26] Joseph Carlson: I will navigate complexity that you find tedious. [7:26] Joseph Carlson: Friction is the product until a machine absorbs the friction. [11:57] Joseph Carlson: A chatbot can recommend a flight while the user still books through an airline or an online travel agency. [11:57] Joseph Carlson: We will put their agents on them and they will battle it out. [15:48] Joseph Carlson: It turns soft loyalty into a consciously tested economic relationship. [29:47] Joseph Carlson: users often want to become a person who speaks a language, not merely solve a translation problem.

Original transcript

[0:00] In February of this year, an article was [0:01] released that shook the markets. It [0:04] caused a widespread panic and it caused [0:05] many stocks to sell down over 7 to 10% [0:08] and they continue to go down just by the [0:11] release of this article and it was a [0:13] narrative about a possible future, a [0:15] situation where agents have took over [0:17] the world. This is the article from [0:20] Citrory. It was published February 22nd. [0:22] Ever since then, many companies [0:24] including SAS companies went through [0:26] what's called the SAS apocalypse. [0:28] Citrony Research was already the biggest [0:30] financial substack in the world with [0:32] millions of readers and this also made [0:34] it even bigger. This went hyper viral. [0:36] The article title was the 2028 global [0:39] intelligence crisis. Intelligence is [0:42] abundant. We have AI that knows about [0:44] everything and so everybody knows about [0:46] everything. But that wasn't even their [0:48] major point. Their biggest point of this [0:50] article is friction going to zero. Now, [0:53] how does friction go to zero? It goes to [0:55] zero through agents. Agents are an [0:57] extension of humans. They're robots that [0:59] can go out and do things that humans [1:00] normally do. Humans normally go and they [1:02] fill out forms. They look at user [1:04] interfaces. They purchase things. They [1:06] shop for things. They compare things [1:08] together. All of that stuff is friction. [1:10] And in a world where agents do all of [1:12] that for you, it erases that friction. [1:14] Now, Citroen didn't just talk about [1:15] agentic technology and the potential of [1:17] it. They also listed specific companies [1:20] and how they think it might work out. [1:21] For example, they talked about Service [1:23] Now decelerating. They made fake [1:25] headlines like Mastercard's revenue [1:27] decelerating down to 6% year-over-year [1:29] growth. They said it would destroy the [1:31] job market, that software companies were [1:32] getting obliterated, and they [1:33] highlighted Door Dash as the poster [1:35] child of the type of company that would [1:37] be disrupted by Agentic technology. Now, [1:40] while Citrony's article was very [1:42] thoughtprovoking, it was well written. [1:44] It was convincing at the time, it was [1:46] also full of flaws. Many of these [1:47] companies aren't going to be affected in [1:49] the way that they've illustrated, and [1:51] I'll give them the benefit of the doubt. [1:52] We didn't really have agentic technology [1:54] then like we do now. Times changed [1:56] quickly. In just a number of months, [1:59] we've had technology released that is [2:01] truly agentic technology. The first [2:03] example of this in a big consumer way is [2:06] Meta's Muse. Meta just recently released [2:09] an app called Muse. I've mentioned it [2:10] before, but the more that I've used this [2:13] app, the more that I realize that it has [2:15] a lot of these type of things in it. It [2:18] is reducing friction dramatically. And [2:20] for those of you that think that Meta's [2:22] Muse is like another ChachiBT, another [2:25] Grock, another anthropics claude, it's [2:28] not. It is entirely different. It feels [2:30] very different. It's the first time [2:32] since the very beginning of using AI [2:34] where I thought it was incredible AI [2:36] could code that I was wowed again. I [2:39] opened up the Metamuse app and it was [2:41] just immediately different than ChatBT. [2:44] It wasn't an empty chat box that spits [2:46] out information. It was this helpful [2:49] assistant that will help you accomplish [2:50] anything you want to do. It said it [2:52] could do things like book hotel [2:54] reservations. It can book restaurants [2:55] for you. It can join your email and [2:57] clear out your email box and unsubscribe [2:59] from unwanted emails. It can organize [3:02] trips. It can plan things for you. Or it [3:04] can shop on your behalf. It can just [3:06] shop for you. I thought, can it really [3:09] book restaurants? And it said yes, it [3:11] can. It can link up to Open Table or it [3:13] can verify directly on the restaurant's [3:14] website. And that was impressive in and [3:17] of itself. But it got me thinking, when [3:20] I can book a restaurant through [3:22] OpenTable without visiting OpenTable's [3:24] app, that changes things. I'm working [3:26] through an agent that's talking to [3:28] OpenTable and I don't have to see [3:30] OpenTable's sponsored listings. I don't [3:32] have to deal with their little [3:33] promotions. I don't even have to have [3:35] the Open Table app on my phone. It can [3:37] go in and just book things on my behalf [3:40] right from this text box. Another thing [3:42] that was interesting is it doesn't just [3:44] do restaurant reservations, but it also [3:46] does the commerce in a serious way. I [3:48] asked it to look up Nike shoes, very [3:50] specific ones that my brother-in-law was [3:52] looking for, and it searched the web far [3:54] and wide. It brought up its own little [3:56] browser. It search through all different [3:58] websites, looking at reviews, trying to [3:59] find the best prices in the exact shoe [4:01] match. It was like having a secretary or [4:04] an executive assistant spending the next [4:06] 30 minutes of their time specifically [4:08] trying to accomplish that task. And then [4:10] just today, I decided that I wanted to [4:13] try out more of the shopping and see if [4:16] it could tackle a harder task. I have [4:18] four kids, one of them is a newborn. So, [4:20] I have three kids that were feeding [4:21] normally that we have to shop at the [4:22] grocery store. And for two of them, they [4:24] have specific allergies like a dairy and [4:25] nut allergies. And I plugged all of this [4:28] in to Muse. I said, "Here's my kids' [4:30] ages. Here's the allergies that they [4:32] have. Could you make a week plan, a meal [4:34] plan for them based on this and make [4:36] sure you check the allergies of each [4:38] food?" said, "No problem. It'll help out [4:40] with that." It recommended Smith because [4:42] Smith has a lot of allergy alternative [4:44] foods. They have lots of substitutes. It [4:46] went on to Smith's website. It made an [4:48] entire meal plan for them. Added every [4:50] single item. And while it did it, you [4:52] can actually watch it live. And you'll [4:54] notice as it goes through these, not [4:56] only does it verify that it's the right [4:57] item for me, it's what I want, but it [4:59] also picks the one with the lowest [5:00] price. It attaches any coupons to it, [5:02] any way to save. It'll actually search [5:04] online for any any applicable way to [5:06] lower the price for you as well. And [5:08] this is just a preview of the new [5:09] agentic world that we've entered into. [5:11] Muse is just the beginning. We know that [5:13] chatbt, we know OpenAI is going to be [5:15] coming out with their version of this [5:17] very soon. Same with Anthropic and same [5:19] with all the rest. We'll have multiple [5:21] agents to be able to extend what we want [5:23] to do to be able to go out in the world [5:25] and do things for us. And it's not just [5:27] shopping for food or groceries or goods [5:29] and services. It's also booking things, [5:31] planning trips, booking hotel [5:34] reservations, going on to booking [5:36] holdings or Expedia may not be quite as [5:38] necessary when you have an agent that [5:40] can do it for you. All of these type of [5:42] things are going to change. And the [5:44] question is, how does this impact all [5:45] these different companies? How does it [5:47] impact companies like Door Dash or Uber [5:49] or Booking Holdings or Open Table? How [5:52] is this going to impact doing different [5:54] tasks like your taxes, filing out your [5:57] taxes with Inuit Turbo Tax? How will it [5:59] impact bigger companies, ones like [6:01] Salesforce that work with sales teams? [6:03] This is going to impact how everything [6:05] is sold and everything is purchased. It [6:07] raises a lot of questions. But now that [6:09] we see the technology working, we can [6:11] make some assumptions of which companies [6:13] are more exposed than others where [6:15] investors should be concerned and where [6:17] they should feel a little bit more [6:18] certain. What I've done is I've looked [6:20] at this technology and plotted it [6:22] through my entire portfolio. I wanted to [6:25] see how Agentic Technology truly living [6:28] in a frictionless world impacts my [6:30] portfolio plus a number of other popular [6:32] stocks. So, let's go ahead and jump in. [6:34] The summary is that the largest risk [6:36] right now being priced into the market [6:38] is automation. That's what everybody's [6:41] focused on. Oh, no. Jobs are being [6:43] automated, processes are being [6:45] automated. It's becoming more efficient [6:47] and there'll be job loss because of [6:48] automation. But we have to change gears. [6:51] Automation is what's been happening, but [6:53] we're moving into a different category. [6:55] Automation is not the biggest risk [6:57] today. Losing your customer is the risk. [7:00] Losing contact with your customer. The [7:03] popular framing is that artificial [7:04] intelligence replaces labor. That [7:07] matters, but it's not the most [7:08] consequential economic change. The [7:11] deeper change arrives when software [7:12] moves from answering a question to [7:15] completing a transaction. It compares [7:17] options, negotiates, books, pays, [7:20] monitors the results, and learns the [7:22] user's preference. At that point, the [7:24] agent becomes the demand gatekeeper. In [7:26] Citrony's research, they said as a [7:28] stress test to ask whether a company's [7:30] real product is quote, I will navigate [7:32] complexity that you find tedious. And if [7:34] the answer is yes, a capable agent [7:37] attacks the reason the intermediary [7:39] exists. Stay with me for a minute [7:41] because we're going to get very [7:42] specific, but I want to cover a few more [7:44] generalities before we jump into [7:46] specific stocks. The other thing that I [7:48] I think is important to understand here [7:50] is that the companies that are the most [7:52] at risk are the ones that sell friction. [7:55] And there's a lot of companies that only [7:57] exist to make something that has a lot [7:59] of friction have less friction. Friction [8:02] is the product until a machine absorbs [8:05] the friction. The more a business [8:07] monetizes search costs, paperwork, [8:10] confusion, inertia, and comparison [8:13] fatigue, the more directly an agent [8:15] attacks its profit pool. I would [8:17] evaluate every single company on two [8:19] axes. How much of its value comes from [8:22] removing tedious steps and how much [8:24] scarce execution remains after those [8:26] steps are automated. The first creates [8:28] vulnerability. The second creates [8:30] resilience, high friction and low [8:33] scarcity. This is the maximum exposure. [8:36] These are the companies that I believe [8:37] are truly in the crosshairs of Agentic [8:40] Commerce. And there's one that we're [8:42] going to get into. It's a company that I [8:43] used to own called Booking Holdings. I [8:45] am becoming increasingly concerned about [8:47] this company. I am more concerned about [8:49] it today than I was when I sold it. So, [8:52] I'll be going through that. But the ones [8:54] that have by far the most exposure are [8:56] routine travel comparisons. When you're [8:59] looking at where to travel, where to [9:00] book, where to get the best best plane [9:02] ticket, those are maximum exposure. [9:05] Templated legal documents, simple tax [9:08] filings. We have init Turboax and simple [9:12] tax filings. The most most friction, [9:14] they're they're taking something that's [9:16] hard to navigate and they're removing [9:17] friction. We have lead reselling, [9:20] commodity, freelance tasks, and basic [9:22] customer support. A general agent can [9:24] perform the work and shop the provider. [9:27] Then we have the ones that are a little [9:29] bit less exposed. High friction and high [9:31] scarcity. Interface at risk, but the [9:34] system survives. So this is the category [9:36] that I would put Door Dash and Uber in. [9:38] Door Dash has an interface and Uber has [9:40] an interface. The bots can talk with the [9:42] interface, but they can't replace the [9:44] networks. That's a whole different [9:45] layer. Now, even in these situations, [9:47] losing the connection with a customer is [9:49] something that should be taken into [9:50] account. But having that back-end [9:52] network is very important for these [9:53] companies. Then we get into the most [9:55] resilient ones. These are the low [9:58] friction and high scarcity ones. So we [10:00] have semiconductor lithography. There's [10:03] really no friction with those companies. [10:04] They just create a product that's really [10:06] needed. They're not going to be [10:07] disrupted by this. They're going to be [10:08] very resilient. We have payment [10:10] acceptance. So companies like Mastercard [10:12] and Visa, they're not taking a friction [10:14] system and making it less friction. They [10:17] have a network that you have to use. So [10:19] that's one that I don't believe will be [10:21] disrupted. We [snorts] also have credit [10:23] ratings. companies like SMB Global and [10:24] Moody's exclusive entertainment. We have [10:26] Netflix and Disney. We have procurement [10:29] scale. We have restaurants, live events, [10:31] and physical experiences. So, you have [10:33] Cheesecake Factory and Texas Roadhouse. [10:35] These are companies with already very [10:37] low friction and very high scarcity. [10:39] Therefore, they're more resilient. Then [10:41] we get into the companies that will [10:42] actually benefit from this. The ones [10:43] that own the identity, they own the [10:45] memory and the intent. the platform that [10:47] knows a user, receives standing [10:49] instructions, controls permissions, and [10:51] can complete transactions that become a [10:54] new demand driver. Obviously, Meta is [10:57] one of these companies. They have Muse. [10:59] They are the ones that are owning the [11:01] identity, the memory, and the intent. [11:03] But Meta also has a lot of other [11:04] businesses that that may actually be [11:07] exposed. So, think about this. If Meta's [11:10] Muse starts to aggregate all the the [11:12] consumer habits onto its one chat where [11:15] now you're shopping for Nike products, [11:16] you're shopping for Amazon products, [11:18] you're shopping for your Smiths and [11:19] Walmart and Target all through this [11:21] single app. All of a sudden, Meta is the [11:24] aggregator. Meta can start putting these [11:26] sponsored agent results in your feed or [11:29] they can take a sliver of every [11:31] transaction with your permission. [11:32] There's a million ways that they can [11:33] monetize because they are now the new [11:36] toll collector. You have to go through [11:38] them to get the product you want because [11:40] they're the ones that reduce all the [11:41] friction, know all of your behavior, [11:43] know your intent. They know what you [11:45] want to buy. We've seen examples of this [11:47] done before. This is exactly what Amazon [11:49] did with their online store to many [11:51] sellers. Amazon is the gatekeeper. They [11:53] have Amazon.com. Everybody goes through [11:55] them to get to all the millions of [11:57] sellers. Now, before we get into [11:58] specific companies, I still think that [12:00] there's some things we need to outline [12:01] specifically with how powerful the [12:03] agents are and how it's different than [12:05] Chat GBT or Gemini. A lot of people [12:08] still might conflate these two things [12:10] together because there's some [12:11] similarities, but I want to go through [12:13] some of the distinctions and why agents [12:15] are really different. So, we have stage [12:17] one, and this is what we're already in. [12:19] We're already in the chatbt era of uh [12:22] review pages being pressured, how-to [12:24] content being commoditized, search [12:27] results and clicks, and basic customer [12:29] service. But it often leaves the [12:31] incumbents in control of execution. A [12:33] chatbot can recommend a flight while the [12:35] user still books through an airline or [12:38] an online travel agency. That's what [12:40] people are doing right now. They're [12:41] going to catch and saying, "Hey, plan [12:43] out my vacation." Then they learn from [12:45] that, but then they still use all the [12:47] services in the way that they're using [12:48] them before. This is where things [12:51] diverge with Agentic Commerce. Stage [12:53] two, once the software can take [12:55] authorized action, the incumbent may [12:58] lose the session entirely. Find me a [13:00] good hotel near the conference for under [13:02] $350 with a quiet room and free [13:05] cancellation. That becomes one [13:07] instruction. The agent query supply [13:10] interprets terms, chooses, books, and [13:13] stores a confirmation, and updates the [13:15] calendar. The traditional comparison [13:17] interface is no longer mandatory. You [13:20] don't actually have to visit any of [13:21] these companies websites to accomplish [13:23] what you want to accomplish. That's a [13:25] huge divergent from what we have today. [13:27] Then we have stage three, memory. Memory [13:30] converts convenience into distribution. [13:33] Persistent memory creates a compounding [13:35] advantage. The agent knows that the user [13:37] values arrival reliability over the [13:39] lowest delivery fee, dislikes redeye [13:41] flights, prefers aisle seats, avoids [13:44] restaurants with loud dining rooms, and [13:46] wants tax documents filed automatically. [13:49] This is not simply personalization. It [13:51] is a durable demand profile. So, if you [13:54] look at this again, it's very valuable [13:56] to just know and learn about a person [13:57] because it makes the future instruction [13:59] set even have less friction than the [14:01] previous one. The first time I do a [14:03] shopping list, I have to go through my [14:05] kids allergies. Every single shopping [14:07] list after that, I don't unless their [14:09] allergy profile changes, unless my [14:11] family profile changes, unless one of [14:12] them moves out or something. But until [14:14] then, now every single time I shop, it's [14:16] easier than the first. That makes it so [14:19] that you have distribution, the people [14:21] that control the memory profile, this [14:23] advanced learning about the user [14:25] themselves. When you enter in the [14:27] credentials of Smiths or Target or [14:29] Walmart, it'll save that securely [14:30] forever. You don't have to enter it in [14:32] again. It already knows that and that [14:34] creates that lock in and distribution. [14:37] Then we have stage four. Commerce [14:39] becomes machine to machine. This is a [14:42] conversation I was recently having with [14:44] my brother. He said airlines are [14:45] starting to price things by the minute [14:47] based on the user. The airlines are [14:49] trying to determine exactly how much [14:50] you're willing to pay by increasing [14:52] prices to make more money. That's true, [14:54] but that's only part of the story. We [14:56] aren't living in a world where the [14:57] airlines will have dynamic or [14:59] personalized pricing and humans will [15:01] just book airline tickets by themselves. [15:04] We live in a world where airlines will [15:05] have dynamic and personalized pricing [15:08] and humans will have agents to fight [15:10] back. We will put their agents on them [15:12] and they will battle it out. This is [15:14] where we get to commerce becomes [15:16] machine-to-achine. [15:17] Google's universal commerce protocol is [15:19] designed to connect agents, merchants, [15:21] and payment providers across discovery, [15:23] checkout, post-purchase support. [15:25] Microsoft describes hosted agents, [15:27] multi-agent workflows, persistent [15:29] memory, and Microsoft 365 integration. [15:32] Visa, Mastercard are building tokenized [15:34] credentials, authorization controls, and [15:36] ways to identify legitimate agents. The [15:38] plumbing for this is being built in [15:40] public. And how do we know that Meta's [15:42] Muse or similar type of agents will [15:44] aggregate all of this demand? Well, [15:45] we've already outlined some of the [15:46] advantages, but I want to go through a [15:48] few more. Many marketplaces look like [15:50] networks, but behave economically like [15:52] habits. Agents are unusually good at [15:55] breaking habits that exist only because [15:57] comparison is annoying. So, again, we go [16:00] back to that first thing. If the product [16:02] only exists because of comparison, you [16:06] think about a company. If the only [16:07] reason the company really exists is to [16:09] make comparison easier, if that's like [16:11] 70% of the value proposition, it's in a [16:13] lot of trouble. A person orders from an [16:16] app already installed. They may have a [16:18] subscription, saved addresses, familiar [16:20] interfaces, and a learn expectation that [16:22] the order will probably work. They do [16:24] not check the restaurant website, [16:26] competing delivery apps, pickup prices, [16:29] credit card offers, estimated arrival [16:31] quality, and current promoting uh [16:33] promotions every night. The cost of [16:35] checking exceeds the expected savings. [16:37] So, the reason people don't do all this [16:39] extra work to save money is it just [16:41] takes too long. You're not going to [16:42] scavenge the internet to try to save [16:44] some money on your booking. They're not [16:45] going to try other booking apps, even if [16:47] it saves a little bit of money. An [16:49] agent's cost structure is different. It [16:51] can check every route each time. It does [16:54] not get tired, forget a coupon, or [16:56] decide that seven tabs are too many. [16:59] Citrony calls this weakness habitual [17:01] intermediation. demand appeals loyal to [17:04] the intermediary because human default [17:07] to the familiar path not because the [17:09] intermediary owns an irreplaceable [17:11] capability. So a lot of companies that [17:13] we think have a huge moat because they [17:15] seemingly have a lot of demand. That [17:17] moat only exists because it's annoying [17:19] to do something else and agents take [17:21] away the annoyingness of it. Humans by [17:24] default open familiar apps, browse [17:26] ranked inventory, accept displayed [17:29] prices and use saved payments. Hope [17:31] fulfillment is good. Agent default is [17:34] they query all authorized channels, [17:37] normalize fees and terms, score price, [17:39] quality, speed and reliability and [17:42] policy and execute the best fit that [17:45] weakens app habits, default status, [17:48] opaque fees, renewal inertia, and [17:50] sponsored rankings. These advantages do [17:52] not disappear instantly. Authentication, [17:55] loyalty rewards, subscriptions, and [17:57] exclusive inventory can preserve [17:59] preference, but the agent can calculate [18:02] whether the bundle is actually worth it. [18:04] It turns soft loyalty into a consciously [18:07] tested economic relationship. [18:10] So, a lot of things people do just [18:12] because they've been doing it just [18:13] because the other options are a little [18:14] bit more annoying, they're no longer [18:16] going to be doing that just because [18:17] other options are slightly more [18:19] annoying. these type of products and [18:21] services on the back end will constantly [18:24] be competing to keep the same [18:25] relationship. It removes a lot of that [18:28] inertia. So, now that we've gone over a [18:30] bit of the technology, what I expect to [18:32] happen, let's go ahead and look at some [18:33] examples. And again, these aren't 100% [18:36] guaranteed, but I believe these are far [18:38] more accurate than what Citroen gave us [18:39] a few months ago. We'll start off [18:41] looking at the specific company Booking [18:43] Holdings. We have Booking and Kayak. [18:45] Travel is the cleanest early test [18:47] because comparison is valuable, terms [18:50] are structured, and the consumer already [18:52] expects an intermediary to assemble [18:54] options. Booking holdings reported $8.1 [18:57] billion of marketing expense in 2025, [19:00] equal to about 30% of revenue. So every [19:04] $10 that they gain in revenue, $3 of [19:05] that goes to to marketing. It also [19:08] discussed a higher direct room night [19:09] mix, lower performance marketing return [19:12] on investment, and changes in paid [19:14] traffic mix. Those facts show both the [19:16] strength and fragility of the model. [19:18] Booking has global supply, conversion [19:21] data, payments, customer support, and [19:23] brand. Yet, it still spends heavily to [19:25] acquire the demand. The threat to [19:27] booking holdings, and the reason that [19:29] I'm so worried about this company is a [19:31] gatekeeper swap. Historically, a [19:33] meaningful part of the online travel [19:34] economics flowed through search engines. [19:36] Booking paid to appear wherever travels [19:38] expressed intent. In an agentic market, [19:41] the traveler may express that intent [19:43] once to a personal assistant. The agent [19:46] asks clarifying questions, searches [19:48] broadly, applies loyalty and [19:50] cancellation preferences, and then [19:52] executes. The paid search auction does [19:55] not necessarily vanish. It can reappear [19:57] as a sponsored recommendation or [19:59] preferred supplier slot inside the [20:01] agent. Travel has ugly expectations. [20:04] Flights are canceled all the time. Rooms [20:05] do not match descriptions. Payment [20:07] methods fail. Local taxes differ. [20:09] Loyalty terms conflict and refunds [20:11] require persistence. A general agent may [20:13] prefer to call Bookings's infrastructure [20:15] rather than rebuild global contracting [20:18] and customer service. Booking can also [20:20] build its own trip agent, use its [20:23] firstparty data and increase direct [20:24] engagement, package flights on the [20:26] ground transport, and make Genius [20:28] membership an explicit user experience. [20:30] So, they do have something they can do. [20:32] Booking can try to build its own agentic [20:35] layer as well and use all of its data. [20:36] They're certainly going to do that. But [20:38] again, they're competing in a very [20:40] uphill battle here. They are the most [20:42] exposed because they don't have any real [20:44] internal structure like Door Dash or an [20:47] Uber where they have thousands of [20:49] drivers or riders going around picking [20:50] up food. The big thing that they offer [20:53] is putting together this trip, making [20:55] sure that they have all the terms [20:57] correctly, and compiling all of that as [20:59] a gatekeeper. And this is exactly the [21:00] type of thing that agents are good at. [21:02] The bare case does not require bookings [21:05] volume to collapse. And that's an [21:07] important thing here. It requires the [21:09] marginal traveler to become owned by a [21:11] different interface. So even if booking [21:14] holdings is being routed all the [21:16] customers losing that direct interface [21:18] control where now they're just an API a [21:21] backend being routed is a big change in [21:23] their economics. If the agent can [21:25] credibly route around booking booking's [21:28] marketing bill may move from Google to [21:30] the agent layer or appear as a lower net [21:32] take rate rather than an explicit [21:34] expense. The agent would have a huge [21:37] part of the control, the consumer [21:39] preference, the ability to advertise, to [21:41] do sponsored listings or take a pay if [21:44] they control the consumer demand. Both [21:46] Open Table and Door Dash are aggregate [21:48] demand companies. But the resilience [21:50] differs because Door Dash physically [21:52] executes the transaction while Open [21:54] primarily coordinates it. Open table is [21:57] super vulnerable. Here we have Door [21:59] Dash, which does have a level of [22:01] vulnerability, but it also has some part [22:03] of it that's more defensible. The agent [22:04] can compare Door Dash to Uber Eats, [22:06] restaurant direct ordering, pickup, or [22:08] subscription benefits and promotions. [22:10] This type of comparison, reaching [22:12] outside of Door Dash's interface when [22:13] you open up the app, attacks that habit [22:15] forming aspect. But the delivery part of [22:17] this is not a database lookup. It [22:19] requires local carrier density, [22:20] dispatching, batching, routing, fraud [22:22] controls, merchant integrations, [22:24] accurate prep instructions, customer [22:26] support, refund, and service recovery. [22:28] Door Dash has reported that revenue [22:30] margin improvements was supported by [22:32] list logistics efficiencies, advertising [22:35] and fewer credits and refunds. A [22:37] reminder that operating quality material [22:39] sha materially shapes the economics. So [22:42] when we look at these two type of [22:43] companies, this is what we should be [22:44] thinking about. Companies where you're [22:46] just the interface and you're just a a [22:48] comparison tool or a booking tool like [22:50] OpenT, you're super exposed to agents. [22:53] Door Dash has exposure because it can [22:55] reach outside of the interface, but you [22:57] still have to go with whatever network [22:58] has the most density and the best [22:59] execution, which is likely Door Dash. [23:02] Now, even with something that has as [23:03] defensible of an infrastructure and [23:05] network and logistics as Door Dash, even [23:08] taking away the interface is still some [23:10] intermediary risk. It's something to [23:12] take into account. When I look at my [23:14] holding in Door Dash, I can see that [23:15] it's dropped down a little bit. And I [23:17] believe it's because investors are [23:18] becoming a little concerned about these [23:20] type of things, which they should be [23:21] factoring in now. Now, I'm keeping Door [23:23] Dash as a position because I think it [23:25] will be more defensible than the other [23:27] companies like Booking Holdings or into [23:29] it. Speaking of into it and taxes, [23:32] that's the next one that I want to [23:33] address into it with tax preparation [23:35] becomes ambient. For a simple filer, the [23:38] enduring friction is collecting [23:40] documents, answering repetitive [23:42] questions, mapping facts and rules, and [23:45] submitting forms. A persistent financial [23:47] agent with permissions to access can [23:49] make much of this continuous rather than [23:52] seasonal. That pressures the value of a [23:54] guided form completion. So if you look [23:56] at what Turboax does, it's basically a [23:58] gigantic form that you have to fill out. [24:00] Like every step of the way, it has [24:01] little animations. It guides you through [24:03] it. It can have like tax people that [24:05] zoom in or whatever. Uh but it really is [24:07] just a big form to fill out. Now, why do [24:09] you pay $100 or $200 for Turbo Tax? [24:12] Because taxes are important. You don't [24:13] want to screw it up. But that's where [24:15] this gets a lot of pressure. If you can [24:17] have an agent that already knows a lot [24:19] about you, your situation, and you feed [24:20] it some of the documents, and it can [24:22] piece it together with reliability, that [24:24] makes it much more difficult to justify [24:26] paying separately for in it. Now, in it [24:29] is starting to fight this. Of course, [24:30] they're not going to just stand still [24:31] and do nothing. All of these companies [24:33] have their own plan. They're trying to [24:35] implement this technology as well. When [24:36] I sold into it, I was concerned about [24:39] Turboax. But after really understanding [24:41] what this technology is capable of, I'm [24:43] far more concerned about in it IT [24:45] shareholders today than I was when I [24:47] sold I believe a lot of the value that [24:49] in it offers with Turboax is being [24:52] compressed and a lot of the tools even [24:54] like Credit Karma are being compressed [24:56] as well. Now again these don't happen [24:58] overnight but the technology is [25:00] progressing dramatically. We have other [25:02] companies ones like Legal Zoom. [25:04] Templates are exposed. Execution and [25:05] judgment remain. Entity formation, [25:08] standard contracts, wills, and [25:10] compliance reminders are structured [25:12] enough for agents to absorb much of the [25:14] explanation and document assembly. I [25:16] know people that work at jobs as [25:18] accountants and lawyers. And a lot of [25:20] what they do is summarize documents and [25:22] they try to understand the nuances of [25:23] contracts. Agents are already doing [25:25] that. AI is already summarizing these [25:27] documents, flagging red flag things. [25:30] they already do much of that work. So [25:31] any type of software that you sell to do [25:34] that is going to be pressured by agents. [25:36] Doing stuff like routine logo [25:37] variations, basic copy, transcription, [25:40] research summaries, spreadsheet [25:42] cleanups, simple coding translation, and [25:44] first pass support can be completely [25:46] directed by agents. All of this is [25:49] completely replaceable. Home selling [25:51] companies, lead generation companies are [25:53] going to be dramatically influenced by [25:55] Agentic technology. insurance, housing, [25:58] travel portals, they're going to weaken [26:00] because comparison is continuous. Select [26:03] quote, Go Health, and Ever Quote. [26:05] Insurance distribution is built around [26:07] complexity, limited consumer attention, [26:09] and periodic shopping. Once people sign [26:11] up for their insurance, they just don't [26:12] want to look anymore. It's annoying. [26:15] It's not fun to look at competing [26:16] offers. But there's one person on X just [26:20] yesterday that he said, "Here's my [26:22] insurance document. Can you find me [26:23] anything better?" In 20 minutes, the [26:25] agent found insurance for $600 per year [26:28] less than what they're paying with the [26:29] very same terms. So, people are going to [26:32] do shopping that they previously weren't [26:34] doing because now it doesn't have the [26:35] friction. We have Zillow. Think about [26:37] how at risk Zillow is. A housing agent, [26:41] one like Muse, could search listings, [26:44] model trade-offs, schedule tours, [26:46] compare neighborhoods, monitor price [26:48] changes, and gather disclosures. They're [26:50] already capable of doing all of this. [26:51] You can have it keep track of prices of [26:53] items or multiple things. You can have [26:54] it remind you of things. You can have it [26:56] watch and scour real websites. How are [26:59] these companies going to compete when an [27:01] agent baked in already does all of this? [27:03] All of these companies are trying to [27:05] insulate and become more defensible and [27:06] build these technologies in themselves. [27:08] But it's going to be very difficult, a [27:10] challenge for all of these companies. [27:12] Now, we move on to much bigger [27:13] companies, ones that are in my portfolio [27:15] and likely in yours. My biggest holding [27:17] today is Google, a $24,000 position. [27:21] $110,000 in the green on this one. So, [27:23] it's been a great position. Let's go [27:26] ahead and take a look at how Alphabet [27:27] fares in a world of agents. There's some [27:29] good here and there's some bad. Alphabet [27:31] reported $224 billion of Google search [27:34] and other revenues in 2025, up from 198 [27:37] billion in 2024. This scale matters [27:40] because even a gradual shift in [27:42] commercial intent can have large [27:44] consequences. Traditional search [27:46] monetizes a sequence of user queries and [27:48] choices. An agent can collapse that [27:50] sequence into one delegate de delegated [27:52] request. The person searching quote best [27:55] hotel in Chicago or cheaper car [27:57] insurance or tax software for [27:59] freelancers creates multiple ad [28:01] opportunities. An agent may gather [28:04] requirements once, query providers [28:06] through structured connections, and [28:08] return one or two recommendations. Fewer [28:10] visible options can mean fewer paid [28:12] clicks and less opportunities to [28:14] monetize browser friction. So this is [28:17] the downside is that basically if you [28:19] search things through like a Muse or [28:21] even a a Gemini and it it just finds the [28:24] thing that you want without giving you a [28:25] 100 results, there's less opportunities [28:27] to advertise. But there are a lot of [28:30] ways that Alphabet's moving that are [28:32] going to protect it. One of them is just [28:33] raw distribution. They have Android, [28:36] Chrome, search, Gmails, map, YouTube, [28:38] all these different things. They have [28:39] intent data, queries, location, video, [28:42] emails, map, commercial signals can help [28:44] authorize agents understand context. [28:46] That's very valuable. They have a [28:48] technical stack. We have Gemini, cloud [28:50] infrastructure, developer tools, and [28:51] enterprise integrations. And then we [28:53] have commercial protocol. The universal [28:56] commercial protocol aims to connect [28:57] discovery, checkout merchants, and post [29:00] purchase service. So, Google has a lot [29:02] of ways to deal with this and adapt that [29:04] many other companies that are exposed [29:06] don't. Google can also reinvent [29:08] sponsored placements. The relevant ad [29:10] unit may stop being a blue link and [29:12] become a clearly disclosed [29:14] recommendation, merchant incentive, or [29:17] preferred fulfillment option or bid for [29:19] an agent's consideration set. The core [29:21] capability matching commercial intent to [29:23] supply and running an auction remains [29:25] valuable if the user trusts the result. [29:27] So basically, even if an agent skips [29:30] Google's search, Google has so much [29:32] information that they could work with [29:33] the agent to still present a very [29:35] valuable backend. They could really [29:37] offer a lot of intent and things that [29:39] the agent will need to make a good [29:41] decision. And Google could monetize that [29:43] intent. This is much more defensible [29:45] than other companies that are exposed. [29:47] Next, we go into an odd combination. We [29:49] have Amazon, Uber, and Dualingo. Each of [29:51] these companies lose one thing, but it's [29:53] a different thing. Amazon's product [29:55] search is exposed. A shopping agent can [29:57] compare Amazon with brand sites, [29:59] Walmart, niche retailers, resale [30:01] markets, and local inventory. That [30:03] weakens Amazon's role as a default place [30:05] to begin a product search and may [30:07] pressure sponsored listings if the agent [30:09] filters aggressively. Yet, Amazon owns [30:12] enormous seller selection, prime [30:14] loyalty, fast fulfillment, returns, [30:17] payments, reviews, advertising demand, [30:19] and Amazon Web Services. Agents may [30:21] prefer Amazon because the full [30:23] transaction is reliable. Amazon can also [30:26] supply the infrastructure that runs [30:28] agents and creates its own shopping [30:30] assistant. They're already clearly doing [30:32] that. Amazon's trying to build in their [30:33] own agentic shopping system with Alexa. [30:37] It's the new Alexa in the Amazon. But [30:39] still, this is a potential problem [30:41] because if you look at Muse, it'll shop [30:43] on Amazon plus a million other websites. [30:45] The risk is to some of the search and [30:47] advertising economics. The opportunity [30:49] is more transactions through superior [30:51] execution. Again, the agents will gather [30:54] all that demand on the front end. So the [30:56] companies with the best back-end [30:58] infrastructure, the best quality, the [31:00] most reliability, the best customer [31:02] support, they'll aggregate the demand on [31:04] the back end. With Uber, this is far [31:06] less exposed because while the agent can [31:07] compare Uber against Lyft, Whimo, Tesla, [31:10] it can compare all of those together. [31:13] Ultimately, it's going to go with the [31:14] best combination of reliability, all [31:17] your preferences, safety, travel time, [31:20] all of that, plus price. And Uber [31:22] already has massive demand density which [31:24] will make it in most cases the best [31:26] choice. We have Dualingo. Now this is [31:28] the same type of case we've had for a [31:30] while with this one. The utility of it [31:32] is exposed but the motivational aspects [31:34] are not. When you shop with an agent the [31:36] agent knows every language. So it can [31:38] navigate any language of any product of [31:40] any website and shop on your behalf. In [31:43] some ways that reduces the practical [31:45] need to learn a language. A lot of [31:47] people, in fact, the majority of people [31:48] learning in Dualingo aren't doing it [31:51] because they want to be able to just [31:52] transact easier. They really want to be [31:54] able to communicate and be part of a [31:55] different culture. So, I would not [31:57] extend that into Dualingo is highly [31:58] exposed overall. Language learning is [32:00] also a hobby, an identity, a project, a [32:03] social signal, and a long-term [32:04] achievement. Streaks and leagues and [32:06] character and curriculum design, [32:08] progress tracking, habit formation [32:09] matter. Users often want to become a [32:11] person who speaks a language, not merely [32:14] solve a translation problem. Next we get [32:16] into how these agents will impact [32:17] Microsoft and Meta. Microsoft is the [32:20] agentic operating system. Microsoft sits [32:23] inside work identity, emails, documents, [32:26] meetings, operating systems, developer [32:27] tools, databases, cloud infrastructure, [32:29] and security. Its foundry strategy [32:32] describes hosted agents, multi-agent [32:34] workflows, persistent memory, Microsoft [32:36] 365 integration. The architecture puts [32:39] Microsoft in a position to coordinate [32:40] tasks across applications rather than [32:43] merely add a chat box to each one. This [32:46] is something that Sachin Adella has done [32:48] incredibly well. He is super forward [32:50] thinking. He's been thinking about this [32:51] for a while. He's not going to let [32:53] Microsoft be disrupted by agents even [32:55] when they don't have their leading AI [32:57] model lab. The economic advantage is [33:00] permission context. An enterprise agent [33:02] needs to know which employees have [33:04] access to which customer records, which [33:06] policy governs an action, who approves [33:09] spending, and how activity is audited. [33:11] Microsoft is fairly insulated here. And [33:13] ironically, while it's a software [33:14] company that hasn't performed too well, [33:16] it's likely far more insulated from [33:18] agentic threat than many other [33:20] companies. With Meta, we have a whole [33:22] different ballgame because Meta sits on [33:24] both sides of things. Meta has Muse, its [33:26] own agentic technology, but Meta also [33:29] has a massive advertising business. [33:31] Advertising relies on shopping. Meta's [33:34] core product monetizes attention and [33:36] discovery, not primarily the user's need [33:39] to navigate a tedious commercial task. [33:41] An agent that removes friction from [33:43] buying may actually make the feed ad [33:45] more valuable. The user discovers a [33:47] product socially, then delegates [33:49] comparison and checkout. Meta has said [33:51] AI interactions can become another [33:53] signal used to personalize content and [33:56] ads. This creates a direct feedback [33:58] loop. Conversations reveal interest and [34:00] intentions. The feed creates demand. [34:02] Messaging connects people and business [34:04] and agents can help complete the action. [34:07] When we actually look at this, meta is [34:08] probably more insulated from a gentic [34:10] threat than Google is because of the way [34:12] that their advertising works. Search ads [34:14] often monetize declared intent at the [34:17] moment a user is solving a problem. Feed [34:19] ads often create or shape intent inside [34:22] an entertainment or social environment. [34:24] Agents compress the first more directly [34:26] than the second. So, if you think about [34:27] it, with an agent, you're far more [34:29] likely to not run into a Google ad [34:31] because you're not needing to search [34:32] different things than you are to run [34:34] into a meta ad. Because even with [34:36] agents, you still want to have Instagram [34:38] res and you want to check up on Facebook [34:40] or whatever. You're still using those [34:42] services because those services aren't [34:44] based on friction. They're based on [34:45] entertainment. Both of these companies [34:48] have to become the gatekeepers. [34:50] Microsoft's route begins with work and [34:52] permissions. Meta begins with social [34:54] identity, messaging, creators, and [34:56] attention. The reason that I have both [34:58] of these companies still in the [34:59] portfolio today, and I've been buying [35:01] Meta like crazy, because not only I [35:03] believe the valuation is good, but [35:05] Microsoft and Meta are likely well [35:07] insulated and even beneficiaries of [35:10] Agent technology, just the opposite of [35:12] in it and booking holdings. Now, next [35:14] up, we get to the big payment [35:15] processors, Visa and Mastercard. This is [35:18] where Citrony basically said that these [35:19] companies were going to go way down. [35:21] They're going to get hurt somehow, and I [35:22] never agreed with that. I continue to [35:24] invest in Mastercard aggressively, and I [35:26] still have a huge position in Mastercard [35:28] today. It is a $189,000 [35:31] position. There's actually a reason to [35:32] believe these companies will do much [35:34] better in this new agentic commercial [35:36] world. Visa and Mastercard represent [35:38] trust on the transaction level and more [35:41] actors and more complexity creates more [35:43] need more need for trust. An autonomous [35:46] transaction introduces new questions. [35:48] Was the agent authorized? Was the amount [35:50] within the user's limit? Which merchant [35:52] received the credential? Can the payment [35:53] be traced, disputed or revoked? How does [35:56] the merchant distinguish a trusted agent [35:58] from an automated fraud? Visa's position [36:01] is Agentic Commerce works around [36:03] tokenization, authentication, spend [36:06] controls, payment, and infrastructure. [36:08] Mastercard's agent pay introduces [36:10] Agentic tokens built on existing [36:12] tokenization. Pasi capabilities with [36:15] partnerships spanning from Microsoft, [36:17] IBM, Brainree, and Checkout.com. These [36:20] are extensions of network roles. These [36:22] companies are not trying to win the [36:24] interface. They don't care. Visa [36:26] Mastercard never really care about being [36:27] the interface. Visa, Mastercard just [36:29] care about being on the transaction [36:31] layer and this arguably introduces more [36:34] transactions and more need to trust [36:36] those transactions as humans are further [36:38] away from the transaction. There's one [36:40] step happening in between. So, there's a [36:42] bigger need for trust. So, I'm keeping [36:44] my holding a Mastercard because I [36:45] believe the actual upside potential is [36:47] bigger than any type of downside [36:49] potential with that company. Next, we [36:51] get to ASML. Agent growth increases [36:54] demand for physical bottlenecks. This [36:56] should be pretty basic, but agents don't [36:59] route around leading edge lithography. [37:01] There's just no going around ASML. If [37:04] you want to do anything that requires a [37:05] chip, you're going through ASML. Not [37:08] only do they not route around ASML, but [37:10] they rely on ASML. They consume [37:12] inference and training capacity, which [37:14] increases demand for advanced logic and [37:16] memory. ASML's moat comes from physical [37:19] accumulated engineering, supplier [37:21] coordination, service infrastructure, [37:23] and customer process integration. A [37:25] generalurpose model cannot synthesize [37:28] that capability through a better [37:29] interface. ASML continues to be the apex [37:32] predator. Nobody hunts it. It's just at [37:35] the top of the food chain. Nobody can [37:37] touch it. Agents can't either. And not [37:39] only can agents not threaten ASML, but [37:42] they create further demand for ASML's [37:44] incredibly monopolistic technology. So [37:46] if we look at these companies in [37:48] combination, Visa Mastercard secure the [37:50] transaction, but ASML enables the [37:53] computation. Both of these companies [37:55] participate in the agentic activity [37:57] without needing to own the assistance [37:58] brand. Not only are these companies not [38:00] being circumvented and they're actually [38:02] benefiting from AI, but the Agentic [38:04] layer also has no need to replace them. [38:06] It's not even their goal. It's not where [38:07] they're planning on monetizing to begin [38:09] with. Now, let's move on to the next [38:11] batch of companies. We have to ask how [38:12] companies like S&P Global, Netflix, [38:16] Costco, and Texas Roadhouse. These are [38:18] four companies. How do agents really [38:20] affect them? Well, not to any big [38:22] degree. SM Global and Moody's is one of [38:24] the questions people have. But the value [38:26] of these companies is not from less [38:28] friction. It's from authoritative [38:30] output. Credit ratings are not merely [38:32] summaries of public information. They [38:34] are regulated, embedded, and mandated in [38:36] contracts backed by methodology and [38:38] surveillance and issued by institutions [38:40] who opinions market participants [38:42] recognize. Proprietary data, benchmarks, [38:45] workflows, and long histories become [38:46] more useful when an agent can query [38:48] them. S&P Global describes subscription [38:51] revenue built from data, valuation [38:54] services, analytics, research, and [38:56] rating information alongside ratings, [38:58] and index linked economics. So, a lot of [39:00] investors have this wrong. They think [39:02] that because AI can put together good [39:04] research summaries that it will pressure [39:06] SMB Global, but that's just the [39:08] opposite. They are getting more demand [39:10] because AI is using their data to put [39:12] together research summaries. Generic [39:14] analysis may become cheaper. trusted, [39:16] licensable, auditable inputs do not. So, [39:20] when you look at it, it is cheap to find [39:22] generic analysis websites. I have one. I [39:24] sell it for $10 a month. It's not that [39:26] expensive. But when you're running a [39:28] fund that manages 10 billion dollars or [39:30] you're a bank and you're lending out [39:31] hundreds of billions of dollars, you [39:33] can't go with just whatever makes [39:34] economic sense for your small portfolio. [39:36] You need to go with what makes economic [39:38] sense for your bank or institution. In [39:40] that case, you have to have perfectly [39:42] accurate decision-grade data. And that's [39:44] where these companies come in. The [39:46] reason that the ratings are so important [39:48] is not because they could find that [39:50] information somewhere else. It's because [39:51] of the authority and trust and brand [39:53] recognition behind them. We also have [39:55] Costco. Agents can compare prices. They [39:58] cannot duplicate the bundle. Costco's [40:01] value is in procurement scale, limited [40:04] assortment, private label quality, [40:06] membership economics, trusted pricing, [40:09] and physical operating model. An agent [40:11] may help a household decide what to buy [40:14] and when, but it can also reinforce [40:16] Costco by recognizing durable value and [40:19] routing recurring baskets there. Costco [40:21] is more of a supplier than a search [40:23] interface. With Netflix, a general [40:25] assistant can tell a user what to watch [40:27] across services that may weaken [40:29] Netflix's home screen discovery [40:31] advantage. But Netflix owns and licenses [40:34] its own entertainment. The agent can [40:36] point to a show. It cannot legally [40:38] stream the catalog without a commercial [40:40] relationship. I find that Netflix is [40:42] very insulated from agentic technology. [40:44] A restaurant agent can reserve a table, [40:46] compare weight times, and suggest [40:48] dishes. It cannot cook the steak, create [40:50] the service culture, or reproduce the [40:52] atmosphere. The same is true for theme [40:54] parks, gyms, hotels, concerts, travel [40:56] experiences, and live sports. Asians may [40:58] reshape discovery and booking while [41:01] increasing utilization of the underlying [41:03] physical product. I actually believe [41:05] that Texas Roadhouse is not only fine, [41:07] but because all their operations, the [41:09] actual restaurant so good, they may [41:11] actually get more increased demand as a [41:13] result of Agentic Commerce. Now, again, [41:15] I don't know if this will play out [41:16] perfectly this way, but I believe this [41:18] is a more accurate look at the type of [41:20] impact Agent technology will have than [41:22] what Citriny came out in February. And [41:24] we'll be able to update this and see how [41:25] it's going in real time because again, [41:27] in months time, we're going to have [41:29] millions of people using agents to shop. [41:31] It is right around the corner. We look [41:33] at the companies that have the highest [41:34] exposure today. These are the ones where [41:36] the agent can take over the user [41:38] interface. Now amongst this, I also [41:40] believe there's a high variability. For [41:42] example, I think that Door Dash is is [41:44] much less exposed than Open Table or [41:48] into it or booking holdings because Door [41:50] Dash has that physical infrastructure [41:52] behind. We have booking having some [41:54] physical infrastructure but not nearly [41:56] as much. We have intu tax prep. If [41:59] filing becomes ambient, if it's much [42:01] easier to organize documents, what does [42:03] that do to in it? And we have all of [42:05] these type of companies where the agent [42:07] will already know a lot of the intent. [42:09] We get to the companies where they're a [42:10] little bit more moderate or they may be [42:12] beneficiaries. We have companies that [42:14] are moderate like Uber, Dualingo, Google [42:17] with their search ads, Amazon with their [42:19] ads. In many of these, it's mixed cases. [42:21] Part of the business is exposed. Part of [42:23] it will actually do really well. And [42:24] then we have the companies that are [42:25] resilient. FICO is an authoritative [42:27] brand is more resilient. We have S&P [42:29] Global and Moody's very resilient. [42:31] Costco and Netflix very resilient. Texas [42:33] Roadhouse, Cheesecake Factory, all of [42:35] those type of companies are very [42:36] resilient. And then we have some that [42:38] should actually benefit. Microsoft and [42:40] Meta should benefit from Agentic [42:42] Commerce. Shopify should benefit. ASML [42:44] should benefit as well. So overall, [42:46] that's my thoughts on how this could [42:47] play out. Now, if the agentic threat [42:50] turns out to not really be real, if [42:52] people don't really use the agents, [42:54] it'll just be because the agents weren't [42:56] good enough to make it so that it was [42:57] better than using the actual interface. [42:59] If these companies can really create a [43:01] much better experience and people just [43:02] want to go directly to that experience, [43:05] then the agentic commerce won't play [43:07] out. We won't have this thesis play out. [43:09] But I believe there's a very solid [43:11] chance that agentic commerce is going to [43:13] be a very real thing because people like [43:15] reducing friction. They like the easier [43:17] option. And in many of these cases, the [43:20] agent is easier. I know this was an [43:22] unusually long episode and a little bit [43:23] more dense than most, but hopefully this [43:25] was informative. That's going to be it [43:26] for this episode. Have a good one.