We walked away from our AI SaaS: here’s why bootstrapping didn’t work

When we started our first AI SaaS, we thought we were building a clever e-commerce tool: a conversational sales agent where shoppers could just describe what they wanted, instead of relying on rigid keywords or annoying filters. It felt obvious, a natural evolution of search. Smarter recommendations, happier shoppers, and recovered revenue for brands.

But the deeper we got, the more we realized something uncomfortable: building an AI startup is way harder than it looks from the outside. The hype makes it seem like all it takes is calling a large language model and suddenly you have a product. The reality is messier.

Here’s what we learnt from this experience.

AI is expensive

One of the first things we realized, even before going live, was how expensive AI becomes at scale. We ran numbers based on our test environment and some realistic usage estimates. When we looked at how many visitors typical e-commerce sites get and estimated how many of them might actually use the chatbot, the projected costs jumped quickly. Every conversation turned into a noticeable cost on the model side. As a small, bootstrapped product, those costs would have hit long before the revenue had any chance to catch up. To keep a margin that would actually sustain us, the prices had to be high and the usage limits had to be low.

The AI ecosystem is unstable

We also saw how unstable the AI ecosystem is, even while still in development. Models changed, pricing changed, limits changed. Sometimes the exact same prompt started behaving differently after an update. New models and tools kept appearing, and every few weeks we felt like we “should” rebuild parts of the system to keep up. On top of that, the platforms we wanted to integrate with started moving in the same direction as us. At one point, Shopify announced its own AI features that overlapped with what we were trying to build. That made the whole thing feel even more fragile: not only was the underlying tech unstable, but there was also a real risk that the main platform in our niche would just ship a similar solution and make our product irrelevant. Long-term stability started to look like a constant struggle, not something we could assume or plan around.

Chatbots create security and control problems

Even while trying it out ourselves, it became clear that a chatbot interface creates its own problems. Just by testing real scenarios and trying different types of inputs, we started to see how many things needed to be locked down properly. That raised a long list of questions about data protection, access control, logging, and how to protect from things like prompt injection. On top of that, we had to plan strict limits around usage (message caps, token budgets, throttling) just to keep projected costs under control. Those limits made sense from a business and security point of view, but they already made the experience feel worse than the original vision we had in mind.

Conclusion

Looking back, none of these problems were impossible to solve. With enough time, money, and a larger team, every issue could be managed or engineered around. However, it wasn’t something we could realistically bootstrap. The costs, the instability, and the security requirements were all too high for a small, self-funded project. This isn’t the kind of SaaS you build on evenings and weekends, and it’s definitely not the kind of product that quietly replaces your job. For us, the smart move was to walk away before we spent even more time trying to force something that didn’t fit the way we wanted to build.

So I was wondering, are people actually making bootstrapping AI SaaS?