Your SaaS Provider Is Now Your Drug Dealer

Yesterday I wrote about AI pricing. How the tokens feel cheap right now because OpenAI and Anthropic are burning venture capital to acquire users. It's not a market price. It's a customer acquisition cost.

That post got people thinking about what happens when the subsidies end. Fair question.

But here's the part nobody is talking about.

The pricing trap is the small play. While you're watching per-token costs, something much more dangerous is happening inside your SaaS tools. They're building a dependency you cannot break. Not with contracts. Not with data exports. Not with money.

Your SaaS provider is becoming your drug dealer.

The High Is Real

Let me be clear: the product is genuinely good. That's what makes this dangerous.

AI inside your CRM does not just save you time. It learns that your VP of Sales hates bullet points and wants narrative summaries. It learns the Johnson account always escalates through legal before procurement. It learns your support team is casual on chat but formal on email, and it mirrors both perfectly.

After three months, your team is faster than they have ever been. After six months, the AI knows your business better than most of your new hires. After a year, removing it would feel like firing your most capable employee. The one who trained everyone else. Who remembers every customer preference. Who never takes a sick day.

The high is real. Your team is superhuman.

And that's exactly when the dependency is complete.

What You Actually Lose

When a traditional SaaS vendor holds your data hostage, it's annoying but manageable. You export your contacts as a CSV. You pull your transaction history through their API. You hire a migration consultant. Two weeks and $15K later, you're on the new platform. Expensive, painful, done.

AI lock-in is different. It's not about your data. It's about the intelligence layer that has been growing on top of your data for the past year.

That layer includes:

• Every customization the AI made to match your team's communication style
• Every preference it learned about how you like reports structured
• Every customer interaction where it adjusted its tone, depth, and urgency based on context
• Every workflow shortcut it developed because it watched your team do the same thing 300 times
• Every prediction it makes about what your customer needs before they ask

You can export your data. You cannot export the intelligence that has been trained on it.

The new system will have your old emails, your old tickets, your old reports. But it will not know that Sarah in procurement responds faster to two-sentence emails. It will not know the quarterly review always needs the regional numbers before the national summary. It will not know to escalate the Johnson account.

It will have the files. It will not have the file memory.

The Export Lie

SaaS companies are getting ahead of this. They're adding "AI data portability" to enterprise plans. They're publishing whitepapers about how you own your model outputs. They're building export buttons.

It's theater.

What they will export: your chat logs, your generated content, your configuration settings. A zip file of everything the AI produced.

What they will not export: the model weights that learned your business. The fine-tuning that happened implicitly through a year of your team's usage. The context embeddings that make the AI feel like a colleague instead of a tool. The reinforcement learning from human feedback your team unknowingly provided every time they accepted or rejected a suggestion.

That's not in the export. It cannot be. The learning is baked into their proprietary model architecture. It's not a file they can hand you. It's a capability they built using your data as fuel. And they own the engine.

You paid for the AI. You trained it with your team's expertise.

You do not get to keep the trained version when you leave.

The Real Cost of Switching

Let's put numbers on this.

A mid-market company with 50 employees using an AI-augmented SaaS platform for 12 months. The AI has handled roughly 15,000 customer interactions. It has learned from approximately 8,000 internal decisions. It has built contextual awareness across 12 departments.

Here's what switching actually costs:

Data migration: $25K, 3 weeks. The easy part.

Retraining the new AI: Your team spends the next 6 to 8 months correcting the new system, re-teaching it every preference, every workflow nuance, every customer quirk. Lost productivity: $150K to $300K, conservatively.

The invisible cost: For those 6 to 8 months, your team is slower. They are frustrated. The new AI makes mistakes the old one stopped making months ago. Customer satisfaction dips. Employee satisfaction dips. You don't measure this line item. You feel it.

And here's the worst part. While you're retraining the new system, your competitor who stayed on the original platform is pulling further ahead. Their AI has 18 months of learning. Yours has zero.

This is not like switching CRMs. This is like firing your COO and replacing them with a college graduate who has never seen your industry. Same title. Same access to files. Completely different capability.

What To Do About It

The answer is not "avoid AI." That ship sailed. The answer is controlling where the intelligence lives.

Three things you can do now:

First, separate the AI layer from the application layer. Use AI tools that run on your infrastructure, your hardware, your terms. If the AI is a capability you own, not a feature inside someone else's product, the learning stays with you when you switch applications.

Second, demand model portability. Before signing with any AI-augmented SaaS vendor, ask one question: "If we leave in 18 months, what exactly do we take with us?" If the answer is "your data exports," walk. If the answer includes a credible path to exporting the model's learned behavior: fine-tuning checkpoints, embedding stores, preference profiles. Stay and negotiate.

Third, build internal AI capability now, while the field is still level. Your team does not need to become machine learning engineers. They need to understand how to run open-weight models on a server you control. One person, one machine, an afternoon of setup. The cost of building this is dropping every quarter. The cost of not building it compounds.

The SaaS drug dealer model works because the high is real and the withdrawal is terrifying.

The only way out is owning your own supply.