I've worked at 23 different companies. I don't take weeks to ramp up at any of them. I start contributing in the first few hours, and the way I do it is boring: I ask who's struggling, what's broken, where the code lives, what's on fire right now. That's it. Not arrogance, just math — there's only so many ways you can break a Flask app or misconfigure a MariaDB server. Ask the right four questions and you know more about a company by lunch than most people learn in their first month.
Here's the thing nobody tells you: none of that speed came from a tool. It came from knowing which four questions actually matter and being willing to walk up to a stranger and ask them. AI hasn't changed a word of that math.
There's this recent trend of founders bragging about how they shipped a landing page, support docs, onboarding email sequence, a demo script, a SQL dashboard, a Python scraper, a positioning matrix, and an investor-style market summary before lunch.
Cool. Now go talk to five potential customers and see if any of them care.
Builder mode is usually just avoidance. It feels productive, and it keeps you from doing the hard thing where a real person can say "no" to your face.
The augmented founder is still a founder
AI can make people faster. It can also make them louder and more obnoxious.
That doesn't change the core job of a founder.
The job is still solving a problem for a customer you can actually reach. Killing your favorite idea when it keeps not working. Understanding that speed does not save you from physics, market timing, manufacturing constraints, customer indifference, or the fact that someone still has to want the thing you're building.
AI can help you produce more artifacts. That does not mean it produces more evidence. A landing page inundated by bots is not demand. A dashboard with feel-good numbers is not insight. A market summary is not a market.
Asking "what's on fire" on day one at company number 23 gets you real signal in an hour because a person answers you, badly or honestly or somewhere in between, and you can read which one it is. Ask a model the same question about your own market and it will answer confidently every time. That's a different kind of signal. Or, more accurately, no signal at all — just a very well-dressed guess.
Schrödinger didn't want us to start boxing up cats
People love misreading Schrödinger's cat as "observation changes the outcome." Schrödinger was pointing at the absurdity of the interpretation, not asking everyone to start a cat-box physics simulation.
Startups are doing the same thing with AI.
"AI changes the outcome."
No. AI changes the speed of the work you do to answer the only question that matters: will someone pay for this? If you use AI to generate a market summary and it convinces you the market exists, that's wrong. Step away from the AI and go validate. Or better, use the AI's market summary to get to validation faster.
What AI is actually good for
AI is good for churning out ideas based on previous information and turning reams of unstructured data into something I can actually react to.
-
First drafts
Landing copy, support macros, onboarding emails, rough docs, internal runbooks. Stuff that used to take a day of staring at a blank page or copy-pasting from other documentation. -
Compression
Turning "we need to review 40 customer calls" into "here are the themes, here are the quotes, here are the contradictions." You still listen to a few calls yourself. You just don't have to drown in them. -
Breadth
Ten variations of positioning, ten experiment ideas, ten ways to explain the same feature to different buyers. Most will be garbage. That's fine. You're buying options. -
Internal tooling
Scripts, QA checklists, triage flows, "if X then Y" support logic. The unsexy stuff that keeps a tiny team from tripping over itself.
One caveat: smaller, specialized models often beat big general models in narrow product problems. If you have specific data, specific failure cases, and known examples of what good and bad look like, the model with the better leaderboard screenshot may not be the model that actually works.
Where you should not trust AI
LLMs are bad at counting, staying consistent, making the mundane interesting, and are biased to make you feel like a genius. It's the unfortunate consequence of capitalism. Would you keep throwing cash at an AI that called you an idiot every day?
I learned this the hard way, early. I was building an AI survey-response tool where generated personas answered questions based on attributes I'd assigned them. One persona had a job title of "janitor" and a salary of $15,000 a year. I asked it, "what's your job title?" It said "CEO." I asked, "what's your salary?" It said "$35,000 a year." That's where I first ran into the word "hallucination" — watching it happen live, in my own output, back when ChatGPT 3.5 was barely out the door and context windows held a few thousand tokens. The model wasn't lying to me. It was just predicting the most likely next token given the context, and "janitor" and "CEO" both show up plenty in the context of "job." Useful tool. Not a creative one, and definitely not an honest one in the way people mean when they say honest.
I keep AI away from:
-
Analytics and dashboards
I want deterministic queries and reproducible charts. An LLM will "estimate" churn right off a cliff. -
Accounting and finance
Same issue, except now the mistake becomes a tax problem, a payroll problem, an investor problem, or a "why is our Stripe account frozen?" problem. -
Compliance and legal
AI can summarize and flag issues. It can't replace source documents, deterministic workflows, or a lawyer who can be meaningfully blamed and held accountable... kind of. -
Production systems that need guarantees
Auth, billing, permissions, retention, logging, migrations. These belong in software you can test and inspect.
If your product needs humans in a back room to make the numbers come out right, you don't have AI. You have Excel with extra steps. There is nothing wrong with Excel with extra steps if that is what you are honestly selling. There is a lot wrong with calling it intelligence because the spreadsheet got shy.
The actual job
The founder still has to do the uncomfortable human parts:
-
Customer selection
The model can suggest segments. It can't decide which pain is worth building a company around. -
Judgment
Which metric matters. Which customer is lying politely. Which feature request is a distraction. Which "urgent" thing is just someone anxious. -
Taste
The difference between useful and slop. Clear versus clever. A real workflow versus a demo. -
Accountability
A model did not build the wrong product. You did.
This is the part AI can make worse if you're not careful.
If you're already prone to believing your own narrative, a model that generates plausible support for it will happily hand you gasoline. It will produce the market analysis. It will produce the persona. It will produce the investor memo. It will produce the confident explanation for why the customer who said "no" was not really your target customer anyway.
Very convenient.
Also how you end up with a beautiful strategy for a company that should not exist.
The real upgrade
Founders don't need to grind every task. They need to design the work, pick the right tool for the right kind of truth, and inspect outputs like a paranoid adult.
Before you generate another artifact, write down the one human behavior that would prove someone cares. A reply. A call. A signup. A payment. A second login. Something observable.
Then use AI to get to that test faster, not around it.
-Sethers
