Why I Don't Sell AI SaaS (And Why That's a Feature, Not a Bug)

Remember Columbia House? You’d get 12 CDs for a penny, feel like an absolute genius, and then spend the next two years trying to figure out how to escape the “membership” you technically agreed to in six-point font. The business model wasn’t selling music. It was selling the illusion of a deal while locking you into a revenue stream you couldn’t easily exit.

I think about Columbia House a lot when I look at the AI SaaS landscape.

Most AI platforms operate on the same principle: low friction to start, high friction to leave — and maximum vendor lock-in once you’re deep enough that leaving hurts. You get an attractive demo, a cheap pilot tier, and a smooth onboarding experience. Then you start building workflows on top of their API, training your team on their interface, and embedding their proprietary data formats into your business processes. By the time you realize you’re paying $4,000/month for something that originally cost $200, leaving feels harder than staying.

That’s a choice. And it’s a choice I made deliberately not to participate in.

The Business Model Most AI Vendors Don’t Want to Talk About

Here’s the quiet truth about SaaS AI platforms: their revenue model is predicated on you never leaving. Not because the product is so good you couldn’t imagine life without it — though they’d love for you to feel that way — but because leaving is expensive.

Think about what “leaving” actually means. You’ve got your data in their format, in their database, on their infrastructure. Your automations are built on their specific API conventions. Your team knows their UI, their terminology, their quirks. You have integrations to rebuild, data to export and re-import, workflows to re-design from scratch. The switching cost isn’t part of the marketing deck, but it’s absolutely part of the business model.

This isn’t cynical speculation. It’s venture-backed logic. High switching costs = low churn = high lifetime value. That’s a great deal for the vendor. It’s a terrible deal for you.

And vendor lock-in isn’t just a cost problem. It’s a control problem. When your AI runs on someone else’s infrastructure:

• You don’t control your data. You hope their privacy policy stays favorable.
• You don’t control uptime. You’re at the mercy of their status page.
• You don’t control pricing. They can (and do) raise rates after the moat is built.
• You don’t control the roadmap. Features you need may never ship.

For an SMB running lean, that’s an enormous amount of strategic exposure to hand to a vendor whose incentives aren’t aligned with yours.

What Client-Owned AI Infrastructure Actually Looks Like

My standard setup for a client AI deployment looks something like this: a VPS (usually Hetzner or Linode — Hetzner in particular is embarrassingly affordable), Docker for container management, PostgreSQL for storage, and pgvector when we need vector search for RAG applications. The orchestration layer is LangGraph or Hatchet depending on complexity. Everything is deployed via Ansible, so the entire infrastructure is defined in code and fully reproducible.

That’s it. No proprietary platform. No vendor APIs that could change pricing on you. No data sitting in someone else’s database that you can’t easily export. That’s what self-hosted AI infrastructure looks like — less glamorous than a SaaS dashboard, but entirely yours.

The bill for this kind of setup is usually somewhere between $30 and $150/month depending on scale. I’ve had clients running meaningful AI workloads for less than their Netflix subscription.

The tradeoff — and I’ll be direct about it — is that it requires more setup work upfront and more technical sophistication to maintain. A managed SaaS platform handles infrastructure complexity for you. That’s genuinely valuable, especially if you don’t have engineering resources.

But here’s the thing: that’s exactly what I’m providing. I do the setup. I write the Ansible playbooks. I configure the monitoring. I document how everything works and hand you a system that you could have another engineer maintain, modify, or migrate without my involvement. The goal is to be thoroughly unnecessary by the time we’re done.

Your Data in Your Hands

The deeper issue isn’t really about infrastructure costs. It’s about data sovereignty.

When I migrate a client’s workflows to an AI system I’ve built, their data stays on a server they control — one they can SSH into, back up, move, or shut down whenever they want. If they decide in two years to switch to a different LLM provider, we swap an API key. If they want to bring the whole thing in-house, we copy a Docker image. If they want to hand it to a different engineer, we have Ansible configs and documentation.

Compare that to migrating away from a well-entrenched AI SaaS platform. You’re looking at data export negotiations, integration rebuilds, retraining your team on new tooling, and probably a few months of running both systems in parallel. That migration cost is where vendor lock-in actually lives — not in the monthly fee, but in the moment you decide you’ve had enough and realize how expensive “enough” is going to be.

I once worked with a company that had built a year’s worth of custom workflows on top of a mid-tier AI automation platform. Nothing exotic — just solid business process automation. When the platform was acquired and the new owners announced a 3x price increase and a feature sunset, they were stuck. The workflows were undocumented, tightly coupled to platform-specific features, and would take six months to rebuild. They paid the increase.

Own your infrastructure. It’s boring advice. It’s also the right advice.

The Business Case for Ownership

Let’s run this practically. Suppose you have an AI workflow doing customer support triage — reading incoming tickets, categorizing them, drafting suggested responses for your team to approve.

SaaS path: You use a platform that offers this out of the box. Setup is fast. Monthly cost starts at $500 and scales with volume. All your ticket data lives in their system. Works great until pricing changes, the feature set drifts away from what you need, or you want to add a capability they don’t support.

Client-owned path: We build the same workflow on infrastructure you control. Setup takes a few weeks rather than days. Monthly infrastructure cost is $40-80. The LLM API cost (OpenAI, Anthropic, or a hosted open-source model) scales with actual usage. You own the data, the workflow, the prompts, and the model configuration. When GPT-5 or Claude-next comes out and is better and cheaper, you swap it in.

The second option costs more to build. It costs less to run, less to change, and costs nothing to exit. That’s a different risk profile. For a business planning to run AI for five-plus years, the math isn’t close.

Why I Made This a Principle, Not a Feature

I don’t avoid the SaaS model because I’m ideologically opposed to managed services. I avoid it because I watched too many clients get burned by the pattern, and I decided I wasn’t going to be part of what burned them.

There’s enough dependency in the world. I’d rather be the person who builds you something you actually own, hands you the keys, and leaves you in a better position than when we started. That’s the goal every time.

---

If you’re evaluating AI infrastructure options for your business — whether to buy into a platform or build something you control — see how I approach it or just let’s talk it through. I’m happy to help you think through the tradeoffs, even if I’m not the right fit for the work.