Actionable insights — Warning: Investors Must Prepare For The Agentic Takeover

7:48 1. Score every holding on two axes: friction sold vs scarce execution left

1. Ask what share of the company's value comes from removing tedious steps for the customer — search, paperwork, comparison, form-filling. That share is the vulnerability.
2. Ask what scarce execution remains once those steps are automated — physical networks, regulated authority, owned content, manufacturing know-how, a real-world experience. That is the resilience.
3. Place the company in one of four boxes: high friction / low scarcity (maximum exposure), high friction / high scarcity (interface at risk, system survives), low friction / high scarcity (resilient), and owners of identity, memory and intent (beneficiaries).
4. Run it across the whole portfolio at once, then act on the names whose box changed since you bought them.

• Companies that own both sides (Meta has the agent and ad businesses an agent might route around); scarcity that is really a habit in disguise (see #2); and a box that only holds if agents actually get adopted — his own stated falsifier.

15:50 2. Test whether a "moat" is a network or just a habit

1. List why customers come back: installed app, saved address, subscription, familiarity, loyalty points, exclusive inventory.
2. For each reason, ask: does it survive a buyer who checks every alternative every time, for free? Habits that exist "only because comparison is annoying" do not.
3. Estimate the share of the value proposition that is "making comparison easier." Carlson's rule of thumb: around 70% and "it's in a lot of trouble."
4. Treat loyalty perks and subscriptions as a bundle the agent will price against alternatives — soft loyalty becomes "a consciously tested economic relationship."

• Authentication walls, exclusive inventory or loyalty economics strong enough that the agent chooses the incumbent anyway; incumbents blocking agents from their sites; and a company that looks habit-driven but actually has hidden execution (the next test).

21:37 3. Split aggregators into those that coordinate and those that execute

1. For any marketplace, ask what physically happens after the order: does the company do the work (drivers, dispatch, fraud controls, refunds, service recovery), or does it pass the booking to someone else?
2. Coordinators lose most of their value once an agent can reach supply directly. Executors lose the interface but keep the network, because the agent still has to route the job to whoever executes best.
3. For executors, look for evidence that operating quality drives the economics (logistics efficiency, fewer credits and refunds) — that's what an agent will reward.
4. Even for executors, discount some intermediary risk; losing the customer relationship still costs something.

• An agent layer that commoditizes the executors by forcing price competition on every order; a second network reaching similar density; and executors whose margins depend on in-app ads and promotions an agent would skip.

18:41 4. Use marketing intensity to size gatekeeper-swap risk

1. Pull marketing expense as a share of revenue. A high number means the company pays a gatekeeper (usually search) to reach demand it doesn't own.
2. Read the filing for signs of strain: shifts in direct vs paid traffic mix, falling performance-marketing ROI.
3. Model the bear case as a gatekeeper swap, not a volume collapse: the same demand arrives through a new interface, and the toll moves from search to the agent — as an explicit fee or as a lower net take rate.
4. Ask what the company can do alone (its own agent, first-party data, membership) versus what the agent owner controls.

• Agents preferring to call Booking's infrastructure rather than rebuild contracting and customer service (the bull case he states himself); Genius-style memberships that survive the agent's math; and take-rate changes in results as the first measurable sign.

34:14 5. For ad businesses, ask whether they capture intent or create it

1. Classify the ad revenue: does it monetize declared intent (a user searching to solve a problem) or created intent (discovery inside entertainment or social feeds)?
2. Declared-intent ads are exposed: an agent collapses a sequence of queries into one delegated request, so there are fewer paid clicks.
3. Created-intent ads are insulated and may gain value: the user discovers a product socially, then delegates comparison and checkout to an agent.
4. For the exposed type, list the defenses that let the company sell into the agent instead (distribution, intent data, a checkout protocol, a new ad unit).

• Search volume and paid-click trends as agent use grows; whether "sponsored agent results" become a disclosed ad unit; and whether social time holds up once agents handle more of daily life.

35:31 6. Find who gains when more machines act for people: the trust and bottleneck layers

1. List the new questions automation raises — authorization, spending limits, traceability, disputes, telling a real agent from fraud.
2. Find the companies whose existing product answers those questions and that never needed to own the interface.
3. Separately, list the physical inputs the new activity consumes (compute, and so chips and lithography) and look for a monopoly supplier.
4. Check that the new layer has no reason to replace them: "It's not even their goal."

• Agents paying over alternative rails (stablecoins, account-to-account) that bypass card tokens; network fees pressured by agent owners with bargaining power; and compute demand that shifts to chips not made with leading-edge lithography.

38:30 7. Separate generic analysis from authoritative, licensable inputs

1. When AI threatens an information business, ask: is its product a summary (cheap to reproduce) or an authoritative input (regulated, written into contracts, auditable)?
2. Ask who needs it: a small investor will accept a cheap summary; a bank lending billions needs "decision-grade data."
3. Check whether AI tools consume the data — then more AI means more licensing demand, not less.

• Regulators loosening rating mandates; AI vendors licensing alternative data sources; and the slice of revenue that really is summaries (research tools) rather than authority.

24:29 8. Re-test past sells (and a famous bear essay) against working technology

1. When a technology moves from forecast to product, use it yourself on real tasks before updating any view.
2. Re-read the influential bear case company by company: which calls held up once the product existed, and which ignored the network or authority behind the business?
3. Revisit names you already sold: is your concern now stronger or weaker than when you sold? A stronger concern confirms the sale and argues against buying back on a dip.
4. Write down what would prove the new view wrong, so you can update it "in real time."

• Adoption numbers once OpenAI and Anthropic ship their own agents; incumbents' own agent launches; and early-user excitement standing in for evidence of mass adoption.