# Out of the AI Lock-in: Companies Should Bet on Vendor Independence Now

> Tying your AI strategy to a single vendor is an unnecessary risk. A guest article on token costs, AI gateways and control over where sensitive data is processed.

- Date: 2026-10-22
- Source: https://romanoroth.com/en/publications/itdbusiness-ki-lock-in/
- Tags: AI, AI Strategy, Token Costs, AI Gateway, Digital Sovereignty


What matters now is not to become dependent on particular AI models. For every use case, pick the model and the vendor best suited to it. The most capable solution is not always the one that makes the most sense economically and strategically. Only then can companies really secure cost control, transparency, technological interchangeability, long-term scalability and, in the end, their own sovereignty.

OpenAI, Anthropic, xAI, and tomorrow perhaps a vendor that is still waiting at the starting line today: the market for generative AI is changing fast. New models appear almost every week, prices move, and the performance gaps are shrinking. For companies this has one important consequence: whoever ties their AI strategy firmly to a single vendor takes an unnecessary risk.

## AI costs make dependencies visible

Costs show this most clearly. As companies use more AI, it is no longer about a few chatbot licences. Once large language models (LLMs) are built into development processes, internal applications or automated workflows, token consumption can quickly become a major cost driver.

Vendors sometimes differ by a multiple. In an internal test, Zühlke ran an alternative model in parallel with its previous standard. The quality of the results was only about ten percent below the previous top model, at roughly a tenth of the cost per token. In many use cases careful prompting closes that gap. A more capable model is therefore not automatically the better economic choice for every use case. For many tasks a small or cheaper model is entirely sufficient; for others, quality, speed or special capabilities come first. Companies need the freedom to choose the right model for each use case and to change that decision when needed.

## An abstraction layer between the company and the AI vendor

Inside a company, requests to language models come from many places: the internal chatbot, the developers' coding assistant, automated workflows. An AI gateway bundles them. Every request goes to the gateway, and the gateway decides which model answers it.

That has two consequences. First, the model becomes interchangeable. Which model a request goes to is set in configuration, not in the code of the caller. Commercial models can be connected just as easily as models running in your own data centre or locally on a developer's machine. Switching becomes a configuration change instead of a rebuild. Second, consumption becomes visible. Because every request passes through the same layer, you can trace which model is used by which project and which team, and how many tokens that costs. AI costs can then be booked where they arise.

Not every company needs its own AI gateway right away. For organisations just getting started, three simple principles are often enough. First, measure token consumption from day one instead of reacting when the first big invoice arrives. Second, for every new use case, check deliberately whether the most expensive model available is really needed or whether a cheaper one does the job. Third, avoid contractual lock-in to a single vendor, or at least limit it in time, so that a switch does not fail on notice periods or migration effort. Whoever takes these three points into account early will find it much easier to retrofit once volume grows.

## Vendor independence is also a question of sovereignty

Freedom of choice between models is also about data sovereignty, not only cost. Not every request should automatically run through the same model and the same infrastructure. Depending on the data and the use case, it may be necessary to process information only in European or Austrian data centres, or within your own infrastructure. In this sense, vendor independence does not mean avoiding the big AI vendors on principle. It means companies keep the ability to decide.

Right now nobody can seriously predict which models and vendors will prevail in the long run, or how their pricing will develop. Companies should set up their AI infrastructure so that they can switch models at any time without rebuilding their entire architecture. Whoever wants to scale AI over the long term should keep three things in view: the technical interchangeability of models, transparency over usage and costs, and control over where sensitive data is processed.

---

*This guest article was originally published in German on [itdbusiness.at](https://www.itdbusiness.at/post/raus-aus-dem-ki-lock-in-unternehmen-sollten-jetzt-auf-vendor-unabhaengigkeit-setzen).*