Every pitch deck I’ve seen in the last three months has an AI slide. Every single one. I sat through a batch of startup pitches last month – EF alumni network, so the bar is usually decent – and eight out of ten were some variation of “we put GPT-4 in front of [industry] data.”
I’ve founded two startups. I’ve been through Entrepreneur First and Google for Startups. I know what a defensible business looks like and what a science fair project looks like. Most of what’s being built right now is the latter.
The wrapper problem
Here’s the test I apply: if OpenAI adds your feature to ChatGPT next Tuesday, do you still have a company? If the answer requires more than three seconds of thought, you’re a wrapper.
Wrappers are easy to build, impressive in demos, and worthless when the platform moves. I saw this exact pattern in mobile (2012-2013), chatbots (2016), and crypto (2021). The playbook is always the same: new technology enables easy prototypes, hundreds of startups launch, the platform or an incumbent absorbs the feature, and the startups die.
The AI version is faster because the prototyping is faster. You can build a convincing demo in a weekend. That’s the trap.
Where the actual value is
The startups that will survive have at least one of these:
Proprietary data loops. If every user interaction makes the product better in a way competitors can’t replicate, you have something. This is boring and slow to build. Good.
Deep workflow integration. If ripping you out requires a migration project, switching costs protect you. This means being embedded in existing processes, not sitting as a standalone tool people can ignore.
Domain expertise that reduces error rates. In regulated industries – finance, healthcare, legal – the model being 90% right is a liability, not a feature. The value is in the last 10%, and that requires domain knowledge the model doesn’t have.
Distribution. If you already have the customers and you’re adding AI to an existing product, you win against the startup that’s trying to acquire customers and build AI simultaneously.
Notice what’s not on the list: “better prompts.” Prompt engineering isn’t a moat. It’s barely a speed bump.
What I’d actually build
If I were starting a company today – and I think about this more than I should – I’d focus on the infrastructure layer. The tooling for evaluation, observability, and cost management. The picks-and-shovels play during a gold rush is a cliche for a reason: it actually works.
Or I’d go deep vertical in a domain I know. Fintech, specifically. (I’m biased – I’ve spent years in the space.) The financial services industry has massive data complexity, real compliance requirements, and budgets. An AI product that can navigate GAAP, regulatory reporting, and multi-currency reconciliation isn’t getting replaced by ChatGPT anytime soon.
The honest take
Most AI startups being funded right now won’t exist in 18 months. Not because AI isn’t real – it absolutely is. But because having access to an API isn’t a business. The companies that survive will be the ones that did something hard with the technology, not something easy.
If your pitch starts with “we use GPT-4 to…” you’ve already lost. Start with the problem. The model is a detail.