Mar 24, 2025 · 52m · startup-ideas
If I wanted to build $1M+ AI startup in 2025, I'd do this
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
Entrepreneurs Andrew Wilkinson and Greg Isenberg examine the rapid commoditization of AI software and outline actionable business blueprints for 2025. They discuss tactical operational workflows, evaluate high-potential startup concepts, and explain why media distribution and personal monopolies represent the only durable moats in an automated economy.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Greg holds 36.2% of the talking time here. How this is scored →
Greg as informed peer 6.3 Guest teaching 4.7 Guest disagreement 2.8 Greg pushing back 3.0
speaking balance: gold is Greg, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 19:01 Wilkinson rejects Cal AI community moat comparison
Wilkinson bluntly shuts down Isenberg's Weight Watchers comparison, pointing out that Weight Watchers had physical in-person meetups and advising any wrapper app founder to cash out immediately before copycats reduce pricing to zero.
Hardest push from Greg ▶ 4:00 Isenberg challenges generic moat commentaryIsenberg interrupts the abstract discussion about moats and distribution, telling Wilkinson that everyone online already agrees with the theory and demanding concrete, actionable business moves.
Biggest teaching moment ▶ 17:02 Wilkinson explains frontier models eating proprietary dataWilkinson explains why proprietary data moats are largely an illusion because frontier model providers scrape and assimilate nearly all niche web data, rendering basic data sets non-defensible.
Greg holds their own ▶ 9:16 Isenberg maps three-stage media-to-software flywheelIsenberg clearly lays out a distinct strategic framework for bootstrapping sustainable AI businesses by acquiring attention first, capturing workflow data second, and deploying specialized AI copilots third.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Greg as informed peer | Guest teaching | Guest disagreement | Greg pushing back | Why |
|---|---|---|---|---|---|---|
| The Moat Dilemma and AI as Sustaining Innovation | 6 | 5 | 4 | 4 | Wilkinson takes a sober, contrarian stance arguing that AI is merely sustaining innovation that eliminates arbitrage margins quickly. Isenberg pushes Wilkinson to move past high-level moat platitudes and get tactical on what actually works. | |
| Sponsor Segment: Startup Empire Membership | 7 | 4 | 2 | 3 | Isenberg articulates his strategic playbook of buying distressed media assets and high-value communities to build distribution and data moats before vibe-coding tools. Wilkinson shares his insider knowledge about TED being for sale while validating the concept. | |
| Market Saturation, Ephemeral Arbitrage, and Data Realities | 5 | 7 | 6 | 4 | Wilkinson challenges optimistic AI app valuations, calling apps like Cal AI temporary arbitrage plays whose equity value will be wiped out by competition and platform incumbents. Isenberg counters by arguing Cal AI could build a modern Weight Watchers community moat, but Wilkinson dismisses this by citing physical vs digital moats. | |
| Tactical AI Workflows: Lead Enrichment to Candidate Screening | 7 | 4 | 1 | 2 | Both founders share concrete, hands-on automation workflows using Gumloop, Zapier, Lindy, and Manus AI. Isenberg details how he enriches inbound sales leads while defending the necessity of human touchpoints for high-tier prospects. | |
| Startup Idea: Secure Model Context Protocol (smcps.com) | 7 | 5 | 2 | 3 | Wilkinson breaks down how Model Context Protocol enabled him to run live financial analysis across holding companies, while Isenberg quickly productizes the insight by proposing smcps.com and building a security-focused brand position. | |
| Startup Idea: Autonomous Web Design Agency via v0 | 6 | 2 | 1 | 2 | Isenberg pitches an autonomous agency concept leveraging v0 to redesign local business websites and sell recurring hosting. Wilkinson enthusiastically validates the unit economics and the defensibility of low-churn hosting models. | |
| Startup Idea: Message Maxing and Local Relationship Intelligence | 6 | 5 | 3 | 3 | Wilkinson pitches message-maxing and private email synthesis using local databases to avoid privacy leaks, and Isenberg categorizes these viral concepts as 'format businesses' similar to Grammarly for relationships. | |
| Startup Idea: Autonomous Vertical Lending and Vibe Capital | 7 | 4 | 1 | 2 | Wilkinson pitches vertical automated lending to bypass slow bank bureaucracy, and Isenberg immediately synthesizes it with Sahil Lavingia's viral proposal into 'Bank of Vibecoding'. | |
| Macro Realities: Competition, Survival Rates, and Personal Monopolies | 6 | 6 | 5 | 4 | Wilkinson doubles down on his gloomy macro thesis that AI collapses startup success rates to 3-5%, while Isenberg emphasizes personal monopolies and creator distribution as the ultimate defensible hedge. |
Statements from this episode (23)
Wilkinson: AI is a sustaining innovation that drops costs, not a moat
“In a lot of instances, it's like a sustaining innovation, right? And what I mean by that is it's a new technology that will drop the cost.”
Wilkinson: Building $1M startups and losing equity has never been easier
“It's never been easier to build a million dollar business and it's never been easier to lose all your equity value.”
Wilkinson: ChatGPT will eliminate software middlemen by building tools on demand
“For all tools basically all middlemen businesses are gone. And I believe fundamentally that we are all going to go to the search bar and the new search bar is ChatGPT and you're going to query it and ChatGPT will just build whatever you need or it'll hook into…”
Isenberg: Defensible AI businesses require proprietary data or network effects
“If you want to build an AI business, you need either a data advantage or a network effect advantage.”
Isenberg: Media audience and proprietary data must precede AI software monetization
“Your first step is how do I get your, the attention? Next step is how do I get your data? And then the third step is, okay, how can I build assistance co-pilots? AI software to actually monetize and build a sustainable business.”
Wilkinson: Chris Anderson is currently shopping TED for sale
“You know, Ted is for sale right now. Maybe you should buy that. Is that for real? Yeah, it's for, I think it's public knowledge right now. They're shopping it around and he's going to give it, I think he's going to give it away to whoever is going to be the be…”
Isenberg: Apps like Cal AI are currently generating $20M ARR
“For example, the one, the flavor right now is Cal AI, you know, apps like that. Like those companies are doing like twenty million dollars ARR on, on downloading an app.”
Wilkinson: Standalone apps like Cal AI will face 100 clones within months
“There's tons of those, but there's going to be like a hundred of them in the app store in like two months. And then Cal AI will not be able to charge what they are.”
Wilkinson built an automated local newsletter with Lindy AI in an hour
“I recently, I own like a local newsletter business and I was goofing around with Lindy and I was able in about an hour to build a Lindy agent that could go through. It would basically like choose this. I chose a city like a city near me. And I said, every day …”
Wilkinson: Device controllers like Apple will capture most consumer AI value
“The people that will ultimately capture a lot of this value are going to be whoever controls the devices. Because I do believe like a lot of these LLM a lot of this functionality will ultimately be implemented.”
Isenberg: Human touch still outperforms full AI automation for top leads
“People are saying that agents are going to automate everything, twenty-four-seven. There's some things that the human touch is going to be better, right? Like, me recording a video, if you know, signed up to one of our forums, maybe me actually recording a vid…”
Wilkinson: Manus AI accurately ranked 50 job applicants in 30 minutes
“I took all those emails, I exported them out of Gmail as a zip and I uploaded 50 candidates to Manus. And then it just spent 30 minutes scoring them all based on what I needed. And it was like, these are the best three people. And then I just focused on those …”
Wilkinson built a Lindy agent to triage email and draft rejections
“I built a Lindy agent that kind of does what you're talking about. So it ingests every single email that I get and then it labels it or archives it based on context. And I have all these examples of like different kinds of emails and whatever. And then if it's…”
Wilkinson: Building around Anthropic's Model Context Protocol is a massive opportunity
“I actually think that might be like my first million kind of opportunity for somebody to really do MCP right.”
Wilkinson: Claude found $100,000 in tax savings from my Xero data
“Like at one point I said how could you optimize my taxes? And it said, oh, you have a credit line over here and you have a credit line in your company. If you use the company money to pay down the personal credit line, you can just immediately write off. Like,…”
Wilkinson: 1Password's moat is trust and uncompromised security, not software complexity
“Like think about like one password is not hard. Like anyone can vibe code one password. What matters about one password is everybody knows one password has not been compromised and they trust.”
Isenberg: Branding an MCP product around security will win software deals
“So if you could basically own, I'm the secure MCP guy. Everywhere else is not secure and create that positioning dichotomy. I think that, that itself is like going to help you win deals.”
Wilkinson: Apple will analyze messages more, but never flag toxic contacts
“I think like some of these things Apple will do over time. Like I think their message summary is, is like an indication that eventually they're going to start digesting more of your messages. But I don't think that Apple would ever tell you who's toxic or narc…”
Wilkinson: An AI text message coaching app will easily make $1M
“I think this is a one year business. I think this is easy to copy. I think someone will copy this idea when I share it, they'll make a million dollars. It's not a longterm business, but it's a cool idea.”
Wilkinson: AI agents make administratively heavy business models commercially viable
“Previously someone would say this to me and I would just say like, no, because it's an administrative nightmare and you need to have an army of accountants and lawyers to do it all. And it just wouldn't make sense. But I think now, It totally makes sense to th…”
Wilkinson: Almost all current AI businesses will be zeros in five years
“To be clear, I'm still saying almost all these businesses will be zeros in five years.”
Wilkinson: I spent $30,000 hiring an agency to build Vibe AI
“And so I hired a agency that I know. And for, you know, 30 grand, we threw together this app called vibe it's in the app store.”
Wilkinson: Vibe is a short-term cash-flow opportunity, not a durable business
“A hundred percent. This is a Cal AI. This is like your, I want like some college kid who's like, I want to get in and get out. I want to come in. I want to get it to a couple million dollars of ARR. I want to cashflow it. And then maybe it's a thing in two yea…”