Summary

AI is becoming a full software product cycle, not a standalone model cycle, because it compounds every prior layer—PC, internet, cloud, and mobile—and reaches billions of potential users through smartphones. Rampell says “the vast majority of net new revenue” in software is now coming from AI at both infrastructure and application layers, while capabilities advanced in two years from text, images, and basic reasoning to native audio and real-time interaction. The investor consequence is an application market growing on already-deployed distribution rather than waiting for a new device base.
Adoption evidence is moving from novelty to ROI: Ramp’s customer expense data inflected in January 2025, software companies are reaching $100 million of revenue from zero in one or two years, and roughly 15% of adults globally use ChatGPT weekly. Rampell’s behavioral shorthand is that people want to be “richer and lazier”; the “magic trick has actually gone into the enterprise” because it now saves time, lowers cost, or produces revenue, regardless of whether current valuations are rich or cheap.
AI-native replacements have their best opening at greenfield moments, while installed systems of record make brownfield displacement brutally difficult and let incumbents monetize captive workflows. Rillet can win when a 50-person company with three entities and two currencies must graduate from QuickBooks, but an “AI NetSuite” or Mailchimp clone faces switching friction. Rampell’s deliberately sharp maxim is “the best companies have hostages, not customers,” though he distinguishes durable moats from businesses users hate.
The largest new TAM comes from turning labor into software, but the compelling pitch is often revenue creation rather than headcount reduction. Salient reportedly helps auto lenders collect 50% more, speaks 21 languages, tracks legal requirements across all 50 states and sometimes counties, and automates work for a $50 million call center with 40%-70% annual employee churn. “We are going to make you more money, and it’s going to cost you less” is stronger than a savings-only story.
AI capability is differentiation, not defensibility; the moat is owning the end-to-end workflow and compounding private outcome data. EvenUp routes “literally 100%” of cases through intake, evidence gathering, medical chronologies, demand letters, and complaints, then learns which cases may be worth $50,000 versus $5 million—potentially lowering the viable case floor from $50,000 to $5,000. Haber calls that loop “showing up to a knife fight with a gun.”
Walled-garden data businesses can capture far more value by selling the finished answer instead of licensing raw information. OpenEvidence combines an exclusive medical-journal license with a ChatGPT-like interface reportedly used weekly by two-thirds of U.S. doctors; VLex’s AI layer reportedly quintupled revenue after 26 years of aggregating legal records, while Ask Leo uses otherwise unavailable contract history such as 50 Deloitte agreements. Rampell’s metaphor: own the rare “vegetables,” then sell the finished meal.
The startup opportunity survives strong incumbents, but selection shifts toward model aggregators, proprietary corpora, vertical operating systems, and acquisitions that buy distribution once—not endless services roll-ups. Acharya argues aggregators can offer a “single pane of glass” across specialized models, unlike labs tied to first-party models; Rampell prefers buying one shrinking collector with five blue-chip clients at three times EBITDA over integrating 200 accounting firms. Early enterprise retention is described as strong, with spending tilting toward forward-deployed engineering as customers ask startups where AI should be applied.

Deep dive

1. AI compounds prior product cycles rather than starting from zero

Rampell’s historical map starts with a recurring split: PCs, internet, cloud, and mobile each produced infrastructure suppliers and application companies, while “product cycles drive growth” through bubbles and Nasdaq drawdowns. Apple and Microsoft, Cisco and Akamai, AWS, then eBay, Amazon, Workday, Shopify, and Veeva illustrate how both layers can endure.
AI builds atop every previous layer. A $40 Android phone is more powerful than the ENIAC, and billions of people already possess smartphones connected to cloud infrastructure; without that installed base, Rampell argues, AI would be an impressive machine “you could go check out in a museum.”
The distance traveled is compressed into years: 2017’s Attention Is All You Need introduced the Transformer, while an early ChatGPT/GPT-2 demo still reminded Rampell of ELIZA’s question-reflecting imitation. Now systems can appear “fully sentient” by historical standards, forcing people to keep changing the AGI goalpost as applications enter what he calls a “golden age.”

2. The enterprise “magic trick” has turned into measurable demand

Rampell pushes directly against a paper claiming most enterprise AI deployments were not working: Ramp’s customer expense data showed a “giant tick up” in January 2025 among tech-forward companies with thousands of employees. His claim is not that every enterprise has transformed, but that actual purchasing is inflecting.
GPT-3.5 could generate a new Seinfeld episode and impress friends; GPT-4 felt remarkable. The break since then is economic: the “magic trick has actually gone into the enterprise,” and software companies can now go from zero to $100 million in one or two years because customers receive material value, not because they merely have excess budgets.
Roughly 15% of adults worldwide now use ChatGPT every week, according to Rampell, with U.S. minutes per user rising rapidly. His family example is intentionally mundane: his wife used ChatGPT in a school-bus dispute, including having it scan laws, before sending a polite complaint and receiving an apology. These “countably infinite” daily use cases—not demonstrations—are what should keep expanding engagement.

3. Greenfield systems of record are the cleanest AI-native opening

Rampell divides the investable application landscape into three shapes: traditional software rebuilt AI-native, software that performs labor rather than competing for software budgets, and walled gardens where proprietary data supports a uniquely valuable finished product. Across all three, the unresolved question is durability against labs, incumbents, and cheaply copied widgets.
Mercury is his canonical greenfield lesson: it built banking for new startups but stole no existing Silicon Valley Bank customer until the weekend SVB failed. Selling an AI-enhanced Mailchimp or NetSuite into an installed account is difficult; winning a newly formed company, or a buyer crossing an inflection or upgrade threshold, avoids that displacement fight.
Rillet illustrates the threshold. A business reaches 50 employees, three entities, and two currencies; QuickBooks struggles to handle its needs, and KPMG says it needs a stronger ERP. At that moment it can choose NetSuite or an AI-native system that “closes the books for you” and includes 50 AI features.
Incumbents should also become stronger. Workday could charge $500 to perform a reference check even if a rival charges $4.99 because the employee record already lives there; support software could shift from per-seat pricing toward payment per outcome when 99% of questions are automated. Rampell’s “hostages, not customers” line means switching-cost moats, not negative-100 NPS.

4. Software expands into labor where the value-cost equation flips

The labor market is “astronomically bigger” than software, but pricing is unsettled. Plaza Lane Optometry might spend only $500 annually on ordinary software while advertising roughly $47,000 for a receptionist; software handling five of eight job responsibilities might command $20,000—not the full wage, but far more than a conventional tool.
Rampell softens his own “software is eating labor” slogan: much of what he sees augments unavailable labor, including calls nobody would staff at 2:00 a.m. He invokes the shift from a country where 90% of Americans were farmers in 1789, while conceding that 3.5 million people who drive trucks might eventually have a better automated solution and that future occupations are unknowable.
Haber finds unusually strong incentive alignment in plaintiff law. A contingency attorney may accept only one of every 100 leads because each case consumes uncompensated labor until a win; making that attorney 5x more productive can increase revenue 5x or more. A corporate firm billing hourly, by contrast, may lose billable revenue when junior lawyers become 50x more productive.
EvenUp owns intake through litigation: its voice agent gathers evidence in 50 languages, sifts medical and employment records, estimates whether a prospect is worth $50,000 or $5 million, and drafts chronologies, demand letters, and complaints. With “literally 100%” of cases flowing through it, private outcomes sharpen future intake and could lower the minimum economical case from $50,000 to $5,000.

5. Salient shows why workflow and proprietary learning outrank AI features

Haber separates differentiation from defensibility. Speaking 50 languages or summarizing documents differentiates EvenUp from a human workflow, but the moat is contextual ownership of the entire case plus outcomes that public models cannot train on. As vibe coding accelerates imitation, Rampell warns that “your margin is my opportunity”; standalone primitives become easier to copy.
Salient automates auto-loan servicing and collections, including insurance follow-up and conversations borrowers often make hostile. Its first customer had a $50 million annual call center with 40%-70% employee churn. The decisive result was not cheaper calls but 50% higher collections: “I will collect 50% more revenue for you every single month.”
Its defense is operational depth accumulated over millions of calls: the system knows what may be said in Missouri, California, Iowa, all 50 states, and sometimes individual counties; ingests statutes while they are still proposed; and delivers conversations in 21 languages. That is Rampell’s answer to why Salient should beat hypothetical “Talient and Zalient.”
The required endpoint is a vertical operating system, not labor sold a penny cheaper. Toast survived skepticism about restaurant failures and low software spending because payments, lending, staffing, and DoorDash integration made it hard to displace; ServiceTitan and Mindbody offer similar evidence that narrowly vertical software can still produce large businesses.

6. Walled gardens turn public raw material into scarce historical data

Rampell’s governing metaphor has OpenAI operating a “vegetable farm” that sells tokens, then opening restaurants that compete with its application customers. One defense is to control a raw ingredient the farm does not possess, build a wall around it, and charge for access—or use it to deliver the meal directly.
FlightAware’s ingredient begins as public ADS-B transponder signals. Rampell says roughly 100 antennas collect aircraft locations, altitude, speed, and tail numbers worldwide; anyone can buy an antenna, but the assembled data becomes information ChatGPT does not have.
PitchBook’s archive of historical funding rounds and DomainTools’ old WHOIS records follow the same logic. A subscription to every legal-tech Series B since 1992 is useful, but a completed comparison memo could be worth $2,000 instead of $20 or $200 because it removes the analyst work between raw data and decision.
Time itself creates proprietariness. Today’s YouTube subscriber count is free, but MrBeast’s count on August 4, 2017 would need to have been recorded somewhere; county property records are public but scattered; old blender manuals can be bought cheaply on eBay and digitized. AI can make those previously marginal archives “10 or 100 times more valuable.”

7. Exclusive corpora become more valuable when paired with finished answers

OpenEvidence looks like ChatGPT but, Rampell says, has an exclusive license to the New England Journal of Medicine and other medical journals. Roughly two-thirds of U.S. doctors reportedly use it almost weekly because a question such as evidence-based care for a torn Achilles benefits from the source material general models lack.
VLex spent 26 years buying, aggregating, and digitizing Spanish and European legal records. Once it added AI, the CEO reportedly told Rampell that revenue quintupled: instead of selling individual articles or a modest subscription, it could produce a 7:00 a.m. client memo incorporating Spanish case law.
Ask Leo applies the pattern to procurement. When a company receives a Deloitte contract, the valuable question is not merely what the document says but what comparable customers rejected or what to push back on. A corpus of 50 prior Deloitte agreements supplies bargaining information that ChatGPT is unlikely to possess.
Kha’s challenge—why license the garden to an intermediary instead of serving the end customer—is exactly the strategic implication. Rampell says VLex should consume cheap model tokens, enrich them with its exclusive data, and sell the finished output directly. The “why now” is that AI can turn an obscure $20 million archive into a plausible $100 million product.

8. Strong incumbents and startup wedges can coexist

AI differs from cloud and mobile because incumbents and customers broadly agree that intelligence is valuable. On-premise vendors once dismissed cloud as unsafe, and many people initially dismissed iPhone, Uber, or Airbnb; today NetSuite, QuickBooks, SAP, Adobe, and Workday are actively searching for AI monetization rather than ignoring the transition.
Rampell is therefore bullish on incumbents and selective about disruption. Intuit may charge its installed QuickBooks base per collection, while NetSuite may find 15 new monetization paths. He is bearish on head-on brownfield replacements across the existing-software “bingo board,” but bullish on greenfield entrants, labor automation, and proprietary-data products that can attack established spending from a different axis.
White-collar services roll-ups are less naturally venture-shaped. Buying one accounting firm, integrating it for nine months, then repeating the process perhaps 200 times leaves a company competing with private-equity firms that already know the playbook; buying a dermatology clinic in San Carlos also does little to create Florida distribution.
A sharper strategy is to buy a sales channel once. An AI-enabled debt collector could acquire a declining company with five blue-chip customers for three times EBITDA, improve collections, and use those reference accounts to onboard 1,000 more without further acquisitions. Rampell sees a related opening in the roughly $100 billion MSP market because remote IT onboarding can scale digitally.

9. Consumer AI repeats the same three application patterns

Acharya maps Rampell’s framework directly onto consumer products. Krea is the AI-native Photoshop selected by young designers choosing their first tool; it is already over 18 months old and has AI primitives built in. ElevenLabs created a vertically integrated voice-and-audio category that barely existed five years earlier and now spans consumer and enterprise offerings.
Slingshot builds proprietary therapy data by providing working therapists with an AI scribe. Notes generated during counseling train a foundation model, which powers the direct-to-consumer therapist Ash. Acharya’s claim is that OpenAI and ChatGPT remain formidable, but they do not possess Slingshot’s specialized corpus, allowing a differentiated, higher-priced product.
Model aggregators have another structural opening. Acharya compares them with Kayak, where users prefer searching every airline over visiting only Delta or United: creative and vibe-coding models have different specializations rather than being perfect substitutes. An aggregator supplies the “single pane of glass,” while labs and Big Tech are, in his framing, constrained to their own first-party models.

10. Category expertise is a sourcing engine, and conviction uses two keys

Rampell describes venture work as “find, pick, and win deals,” then help without giving CEOs bad advice. The firm publishes category research, benchmarks, application rankings, and videos because being jokingly called “a media firm that monetizes with venture capital” captures a real method: content builds expertise and helps the firm find, pick, and win deals.
Rampell uses the phrase “adverse selection versus positive selection,” then says a cheap deal hanging around for six months is probably bad and that the team wants the best company every other strong venture firm wants. Competitive pricing is not the disqualifier; lack of competitive demand may itself be adverse information.
Investment approval is deliberately conviction-oriented rather than a committee vote followed by political trading. A two-key process checks whether the champion met every competitor and completed top-quality work, then often defers to the person “in the arena,” especially on smaller seed checks where a younger investor may understand an emerging consumer behavior better.
The organizational constraint is leverage in winning exceptional deals, not raw capacity to write checks. For a “superpower deal,” the whole firm participates—Rampell jokingly calls Marc Andreessen the “Air Force” and an F-35. Senior operators with board gravitas can matter because choosing the wrong company creates both a loss of invested capital and the much larger omission of the category winner.

11. Retention looks healthy, while enterprise delivery shifts toward engineers

Acharya says the portfolio has not yet shown widespread price-shopping or switching. Retention is strongest when a company surrounds the AI primitive with a rich software ecosystem and acts as the customer’s broader AI solutions provider, continuously translating new primitives into top-line gains; he also reports encouraging consumer retention signals.
Acharya reports unusually strong inbound demand—EvenUp had not needed an outbound motion despite its scale—but expects substantial enterprise sales eventually. The near-term investment is more often forward-deployed engineering: large corporations need startups to identify where AI applies and to adapt products to those workflows, not merely another salesperson carrying a generic pitch.
Rampell closes with the cultural version of the thesis: before adding headcount, many startups now ask, “Can you use AI for this job?” AI-native vendors cannot credibly transform customers while operating internally through untouched legacy processes; they must apply the technology to both their own cost base and revenue engine.

E221 | CES and Chinese Brands Going Global: Do We Really Need Humanoid Robots?Harvey CEO Winston Weinberg: How to Make Mega Deals | Lessons from Rabois, Halligan & Grady