Will Claude Kill Your Startup in 3 Years?
What a $600M fund GP taught me over coffee about building AI startups that actually survive.
Hi I’m Swapnil. I am a serial entrepreneur, having started three startups and sold 2 of them. I am now the CEO of Zeni, a VC at Twin Ventures, and an investor in 40+ AI startups. I share insights from my 15-year entrepreneurial journey to help startup founders build scalable, fundable, and purpose-driven companies. Join me as I share real lessons from the founder trenches.
Most AI-native startups won’t survive Claude.
That came up over coffee with one of my board members this morning.
He runs a $600M+ fund. He’s backed multiple unicorns. He’s seen this movie play out in two prior cycles. So when he talks, I listen.
He sees AI-native startups growing revenue 5x or 10x per year.
He’s less impressed than you’d think.
Here’s why.
When growth is that easy, growth isn’t a signal. The wave is doing the work. The product underneath might not be doing much at all.
His real question isn’t “what’s your revenue this quarter?”
It’s “will Claude kill you in the next 1 to 3 years?”
That question hit me hard.
Because the same force that’s making you grow today is the same force that can erase you tomorrow.
The growth illusion
Here’s the picture.
In 2023, you could ship a product that wrapped GPT-4 and find paying customers. In 2024, you could ship a product that wrapped Claude and grow 10x. In 2025, agents were everywhere.
Every time the foundation model levels up, a layer of startups dies.
The model gets smarter. The wrapper gets thinner. The customer realizes they don’t need the middleman. And just like that, what looked like a high-growth startup becomes a feature inside someone else’s product.
This isn’t a future risk. It’s already happening every quarter.
Most founders don’t want to talk about it. They look at their growth chart and feel safe. The chart won’t tell you whether the next Claude release is your tailwind or your funeral.
The water and the plumbing
The way I think about it now, after that coffee:
Claude is becoming the water everyone has access to.
You don’t get rich selling water. You get rich owning the plumbing it runs through.
The plumbing is your complex workflows. It’s the trust you build by living inside critical operations. It’s the data and domain knowledge you bake in so deep that ripping you out hurts.
The plumbing is what doesn’t change when the water gets cleaner.
Service-as-Software vs Agent-as-a-Service
There are two kinds of AI startups right now, and most founders are confusing them.
Agent-as-a-Service is selling a smart prompt as a product. You take the customer’s input. You hand it to Claude. Claude returns an answer. You wrap it in a UI and charge a subscription.
This works until Claude gets smart enough that the customer can run that prompt themselves. Or until ten other startups offer the same wrapper at half the price. Whichever comes first.
Service-as-Software is delivering a real outcome inside a complex workflow. You don’t just route prompts. You own the data, the integrations, the human-in-the-loop, the audit trail, the customer’s pain.
Agents are commoditizing fast.
Managed deployments inside complex workflows are not.
That’s where the moat lives.
What this looks like at Zeni
I’ll use what I know.
Zeni is AI-powered finance for startups. We hit 90.81% automation on bookkeeping. We manage over $20 billion in finances across 1,000+ startups.
If we were just an agent calling Claude on every transaction, we’d be dead in 18 months. Claude would learn finance better than us. Customers would route around us.
What actually keeps us alive isn’t our prompt.
It’s the years of categorization rules we’ve built across thousands of real-world edge cases.
It’s the relationships our finance teams have with our customers. The late-night message about a wire that didn’t land. The nuance of how a Series B books deferred revenue.
It’s the audit trail and compliance posture we’ve earned with auditors, tax authorities, and the deep bank and payroll integrations we’ve built over years.
Claude can be brilliant. Claude cannot replicate any of that overnight.
That’s the difference between a workflow you own and a prompt you rent.
The test every founder should run
Here’s the test the investor’s question really points to.
Imagine Claude becomes twice as smart next year. Maybe four times as smart.
Do you get stronger? Or do you get weaker?
If you get stronger, your customers benefit from a more capable Claude inside your product. Your moat only gets more valuable.
If you get weaker, your customers start asking why they need you at all.
If you can’t answer that fast, you don’t have a startup. You have a Claude wrapper.
What I’d do if I were starting today
If I were starting an AI-native company today, here’s where I’d focus.
I’d pick a workflow no one wants to touch. Something boring. Something messy. Something where the data is hard to get, the regulations are real, and the customer needs to trust you over months before they’ll let you near their core operations.
I’d build human-in-the-loop into the product from day one. Not as a bandage. As a feature. The humans make Claude better. The loop becomes the product.
I’d treat data as a flywheel. Every transaction, every decision, every correction is something Claude alone can’t capture and you can.
I’d avoid anywhere a customer could replace me by writing a single prompt.
The new bar
Growth doesn’t equal survival anymore.
The startups that survive Claude are the ones that stop trying to compete with Claude and start sitting on top of it.
So ask yourself.
Are you building something that gets better because of foundation models?
Or something that gets replaced by them?
The difference is your survival.