I've run a small SaaS on the side, and I've used Claude Code and LLM agents to automate everything from boilerplate CRUD to A/B test copy. It works. The output is decent. But the moment I stopped giving it precise direction, it started generating solutions for problems nobody had. That's the gap.
In the startup world, the "Stanford grads in an incubator" thing looks like an efficiency problem. Why pay people to think when a data center thinks faster and cheaper? But incubators do something AI can't: they force founders to talk to real customers, face rejection, and pivot before the runway runs out. That's embodied learning, not just token prediction. LLM agents can simulate a customer interview, but they can't feel the sting of a lost deal or notice the hesitation in a prospect's voice. Those signals shape a founder's judgment in ways that don't reduce to data.
Resource Acquisition vs. Innovation #
The broader worry you raise is real. If every corporate plan starts with "assemble GPUs and data centers," the competitive dynamic becomes buying scale before anyone else can. That's resource acquisition, not innovation. Truly new stuff still comes from a person asking "what if we looked at this differently?" — often because they're not bound by the incumbent's metrics. A data center can optimize a known objective. It can't decide which objective is worth optimizing.
For anyone working through this, here's a practical workflow that's been serving me: use AI for the mechanical parts — prompt engineering for customer support, LLM agents for outreach sequencing, Claude Code for the MVP's boring endpoints. Keep the human in the loop for judgment-heavy calls: picking a niche, setting pricing, and deciding what not to build. This isn't a complete guide to deploying agents from scratch; it's more of a real-world boundary. The AI is the accelerant, not the fuel. The "keep all the profit" framing also misses how startups actually create value. Profit from an automated workforce is just arbitrage — it lasts until the next person automates the same thing cheaper. Innovation, on the other hand, compounds through relationships, brand trust, and proprietary insight into a customer segment. Those are built by humans who ship, support, and iterate. No model can do that alone.
If AI can build anything, the point of your startup is to figure out the one thing that shouldn't be built — and then find the people who need it. That's not a step you can outsource.
[Next Google Earth's AI Fabricates Satellite Images →](/en/news/4559/)