-
20 MOST FREQUENT MISTAKES STARTUPS MAKE (ESPECIALLY AI STARTUPS)
Paul Graham has spent the better part of thirty years writing things down. He built a startup, sold it, then spent two decades watching thousands more try to do the same thing through Y Combinator, turning what he saw into essays that read like advice from your smartest friend, not a consultant’s slide deck. I went back through his full archive on paulgraham.com – the early essays on founders and location, the growth writing, the piece on staying “default alive,” the one on the eighteen mistakes that kill startups – and lined it up against what roughly twenty-five other venture investors, across enterprise software, consumer, fintech, healthcare and deep tech, have been saying about the current wave of AI-branded startups. The pattern that jumped out: almost none of the new mistakes are actually new. AI just gives founders a faster, shinier way to make the same old errors, plus a handful of new ones stacked on top. Here are the twenty that show up again and again, whether the startup in question sells software, insurance, legal services, or diagnostics.
1. Nobody actually wants it. Graham’s oldest line still holds: there’s really only one mistake that kills a company, and that’s not making something people want. Everything else here is downstream of that. Founders fall for a model’s raw capability instead of a customer’s pain, and mistake applause at a demo for willingness to pay.
2. One founder, no partner. A solo founder isn’t automatically doomed, but Graham has long flagged it as a warning sign, because it usually means nobody who understood the space well enough was convinced to join. Investors watching AI teams describe the same pattern: a technical founder building alone, with nobody around to argue about whether the model is solving the right problem.
3. Chasing a safe, tiny niche. Picking an obscure corner of a market to dodge competition. Graham’s rule still holds: you can only avoid competition by avoiding good ideas. Because it’s cheap to spin up a narrow AI tool fast, founders chase micro-niches that only look defensible because nobody else has bothered yet.
4. Copying someone else’s idea. A derivative pitch – “it’s like that other company, but with AI” – usually fails because the founder has no first-hand insight into the problem. Investors describe a flood of near-identical products, built by teams who noticed a hot category rather than lived the problem.
5. Taking forever to ship. Graham has always pushed founders to launch fast and rough and learn from real use. AI teams still fall into the old trap of holding a product back for another round of polish instead of getting it in front of paying customers.
6. Betting the whole company on one model. Graham warned early founders about picking the wrong platform to build on. Today that means building an entire business on top of a single model provider’s API, with no plan for what happens when that provider changes its pricing or simply ships your feature for free.
7. A thin wrapper with no moat. This is the AI-era mistake investors talk about most. A founder puts a friendly interface on someone else’s model, charges a markup, and calls it a company. It works until the model provider ships that exact feature natively, and the business disappears in a release note. The ones built to last invest in owned data and real switching costs, not just clever prompts.
8. Hiding from customers behind the product. Graham’s advice to do things that don’t scale – manually recruiting and talking to your first users – gets skipped by founders who assume the model will do the selling for them. It won’t. Somebody still has to sit with a real buyer and watch them use the thing.
9. Raising far more than needed. A big round feels like validation, but Graham has pointed out the real cost: pressure to spend it, a harder time pivoting, and a company optimizing for its investors instead of its customers. In AI, oversized rounds too often get burned training a custom model before anyone proves the idea works with an off-the-shelf one.
10. Spending like you’ve already won. Compute bills alone can sink a company that mistakes usage for revenue. Graham’s simplest survival test still applies: assuming flat expenses and current growth, does the cash on hand actually reach profitability, or just the next round? Too many AI startups can’t honestly answer that.
11. Letting investors run the company. Founders who can’t manage their own board end up building the company their investors want instead of the one their customers need. Graham’s writing on this hasn’t aged: good investors want founders who push back, not founders who default to whoever’s loudest in the room.
12. Chasing margin over users. Sacrificing product quality for short-term profit – throttling usage, degrading output, adding friction to save on inference costs – kills the very thing that made people want the product. Graham flagged this decades before compute cost was a line item on anyone’s income statement.
13. Founders who won’t touch the unglamorous work. Founders who see themselves as purely visionary and won’t sit in on support tickets, sales calls, or the actual outputs their customers are seeing. Graham’s view has always been that the founders who get their hands dirty early are the ones who learn what’s actually broken.
14. Unresolved founder conflict. Cofounder blowups still quietly end more startups than funding problems do. It has nothing to do with AI, and yet it shows up just as often in this wave, because speed and hype put pressure on partnerships that were never that solid to begin with.
15. A half-committed founder. Graham’s closing point in his essay on mistakes was the simplest one: a startup takes full obsession, not a side project with a day job still attached. That’s just as true when the side project happens to be in a very fundable category right now.
16. Mistaking a pilot for a business. This one is newer, and it is everywhere in enterprise AI right now: a flashy proof-of-concept in one department that never turns into a signed, budgeted, company-wide deployment. Investors and operators increasingly describe this as pilot purgatory – dozens of promising demos, almost none of them converting into real recurring revenue.
17. Bolting AI onto an old workflow. Adding a chatbot to an existing product instead of asking what the whole process should look like once intelligence is cheap. The companies actually winning right now are rebuilding the workflow from scratch, not decorating the old one.
18. Ignoring the rules of the industry you’re entering. This mistake barely existed for a typical software startup, and it’s now one of the most common ways AI companies die. A tool that scores loan risk, drafts a treatment note, or writes a contract clause trips regulation built for a different era, and “we’re just software” stopped being a defense once real money, health, or liberty was on the line.
19. Confusing benchmark scores for customer love. Optimizing the product to look good on a public leaderboard or a slick demo instead of on the number that actually matters – whether a paying customer renews. It’s a seductive trap in this category because the leaderboards are public and the renewal data isn’t.
20. No answer for what happens when the model improves. Every AI company should explain, in one sentence, why it still exists once the underlying models are ten times more capable and cheaper. Some investors are uneasy enough about the pace of change that they’re telling founders of even successful AI companies to consider selling within a year or two rather than assume today’s edge holds.
None of this is really about AI, and that’s sort of the point. Strip away the model and the funding round and it’s the same list Graham has been writing about since before most of these founders could vote: build something people actually want, badly enough that you’ll do the unscalable, unglamorous work of getting it in front of them, spend like the cash is finite because it is, and keep your own team and investors pointed at the customer instead of each other. AI just handed everyone the fastest, shiniest way ever invented to skip that work. Don’t skip it.
-
Marketing Plan for BidBus (random company): Insights from a Simulated 100-Person Focus Group
Looked into BidBus, a company we have no affiliation with, just wondering how a marketing plan would look for them in their vertical. A longer report is available upon request. 100 fictitious people cooked up a BidBus marketing plan. See the 1 page report.
-
Marketing Plan for WithCoverage: Insights from a Simulated 100-Person Focus Group
AI, prompted correctly, can help generate marketing plans that hold up in practice. The method: simulate a 100-person focus group split into 20 breakout rooms of 5, gather their deliberations, then merge everything into unified recommendations – distilled into a one-pager and a longer 5-20 page report.
See a sample – WithCoverage (unaffiliated demo). Message me for the full report.
-
Marketing Plan for Energy Substantiation: Insights from a Simulated 100-Person Focus Group
With the right prompting, AI can help produce marketing plans that actually hold up. My approach: simulate a 100-person focus group discussing the company, split into 20 breakout rooms of 5 for deliberation, then merge everyone’s input into unified recommendations. That gets distilled into a full report (5-20 pages) plus a one-pager.
As a demo, here’s an unaffiliated example – Energy Substantiation to test the prompts. Contact me for the longer version. -
Marketing Plan for Reevo: Insights from a Simulated 100-Person Focus Group
Model 100 fictitious people in AI, each with a variety of skills. Break them into 20 random groups of 5. Have them independently discuss a random unaffiliated company (Reevo) twenty times. Collect those comments and create a marketing plan out of it.