A small business signs up with an AI vendor to handle customer service, generate quotes, or produce content. Within months, the tool becomes part of daily operations — the team learns the right prompts, the process depends on it, reports come out with no extra effort. Almost nobody stops to ask what happens when that vendor doubles the price, discontinues the feature used every day, or retires the underlying model. That question usually arrives too late — after the change notice already landed in the inbox.
AI isn’t like switching spreadsheets: the technical reason
Switching spreadsheet or CRM vendors is painful but predictable: data exports in a standard format, and system logic doesn’t quietly change behavior on its own. With generative AI, the mechanics are different. The same instruction — “summarize this contract and flag the risks” — can produce a consistent result on one model and a vague or inaccurate one on another, because each model is trained on its own data, goes through its own behavior-tuning process (the step that calibrates tone, response format, and what the model will and won’t do), and operates within its own context window. There’s no “universal prompt” that behaves identically across every AI engine — the instruction is, in practice, tuned to the specific model answering it.
Add to that a factor most businesses don’t track: AI model providers retire model versions on a regular schedule, typically publishing notice months before shutting an old version down. When that happens, the tool a business uses every day can swap engines under the hood even without the business switching vendors — and the same instruction that worked well months ago starts responding differently, forcing a prompt rewrite nobody budgeted for.
The four layers of dependency a small business doesn’t see
Vendor dependency on AI rarely shows up as a single line in a contract — it builds up across four distinct layers, each more expensive to undo than the last:
1. Data. Conversation history, uploaded knowledge base, the output of every interaction — all of it usually lives only inside the vendor’s proprietary format, with no simple export for use anywhere else.
2. Prompts and business rules. Instructions refined over months — what the AI should approve, how it should respond, what it should never do without human review — typically exist only inside that one tool’s configuration, never documented anywhere outside it.
3. Integration. Vendor-specific API calls embedded in spreadsheet automations, the chatbot on the website, or the customer service system — each one has to be rewritten, not copied, in any future migration.
4. Team knowledge. People learned to operate that specific interface — its shortcuts, its limits, the usual workarounds — not necessarily the underlying business logic, which is what would actually need to survive a migration.
Three questions before deepening reliance on a single vendor
None of these four layers is a reason to avoid AI or split the budget across redundant vendors — that has its own cost and rarely pays off for a small operation. The point is knowing the real size of the dependency before deepening it. Three simple questions cover most of the diagnosis:
Can I export my own data today, in an open format, without opening a support ticket? If the answer requires waiting on the vendor, the data isn’t really yours — it’s a favor the vendor grants.
Are the business rules written down anywhere outside the tool? A simple document — what to approve, how to respond, what to escalate to a human — that survives independent of the interface is what makes it possible to rebuild the behavior on a different AI engine without starting from zero.
Have I tested running the same task with an alternative vendor, even with no intention of switching? That periodic “switch test” — quarterly, for instance — is the only reliable way to know what a real migration would actually cost, before one gets forced by a price hike or a discontinued feature.
The lock-in isn’t in the contract — it’s in what the company stopped keeping for itself
In the view of João Paulo Batistella, an innovation executive with a career spanning Ericsson, Telefônica/Vivo, and five years as CEO of EISA, AI only sustains real value once a business process is formalized as data — not left living inside one person’s head. The same logic applies, point for point, to the relationship with an AI vendor: the real risk isn’t the price-adjustment clause in the contract, it’s the business rule the company never wrote down anywhere outside the tool. A company that documents its own reasoning switches vendors with administrative pain. A company that let that reasoning live only inside a third party’s configuration switches vendors by rebuilding from scratch — and usually discovers that at the worst possible moment, under deadline pressure.
This also clarifies when it’s worth deepening reliance on a specific vendor — for example, by fine-tuning a model on the company’s own data, which increases dependency further, since that tuning stays locked to the vendor that processed it. Fine-tuning can be worth it when the quality gain is real and measurable; the problem is doing it before answering the three questions above, buying deeper lock-in without knowing the price.
Follow Eleva Tecnologia for more analysis on technology applied to business: follow @ElevaTechno on X, @elevatechnologies on Instagram or Eleva Technologies on LinkedIn, or learn more about the group at elevatec.net/about.
