AI lock-in: when vibe-coded apps are tied to one provider
Apps built quickly around one AI provider such as OpenAI or Anthropic are cheap to make and expensive to untangle. Switching models breaks the software, and the provider knows it.
Concepts around artificial intelligence and language models.
Apps built quickly around one AI provider such as OpenAI or Anthropic are cheap to make and expensive to untangle. Switching models breaks the software, and the provider knows it.
Many AI experiments get stuck in the pilot phase. That is rarely down to the technology, but to the choice of process. Start with one well-defined task with a clear return.
An AI agent can look up data, run calculations and make proposals on its own. By setting out precisely what it may and may not do, you get the convenience of AI without losing control.
Language models sometimes make up a source or an amount, and state it just as firmly as a correct answer. You cannot prevent that, but you can control it: with checks, comparison and a system that dares to admit doubt.
An AI assistant does not know the quirks of SAP. When the platform itself supplies the rules, pitfalls and working patterns, the AI does not have to guess.
The Dutch data protection authority received a report of such an incident at a GP practice. With AI that runs in your own environment, data does not leave the company.
A language model is strong at language and structure, but not at calculating, counting or looking things up. Leave that to ordinary code, and use AI to structure questions and explain results.
A large share of employees use AI tools whose use their employer does not know about or has not approved. Banning them does not help. A safe alternative that works better does.
A Two Swallows whitepaper for business and IT leaders. Who can still change an application when the original team or the original supplier leaves?
Converting speech to text can be done entirely on your own hardware, with open models such as Whisper and Parakeet. We measured how well and how fast that works.
With model escalation, a small AI model answers the question first, and a larger model only takes over when the answer scores too low in a check.
A conversation with an AI service in the cloud is not private by default. A case in Canada shows what a provider can see and share. What does that mean under the GDPR, in the United Kingdom and in Switzerland?
With RAG, a system first finds the relevant passages in your own documents and gives only those to the language model. When the system finds the right passages, the model can base its answer on them.
Generated ABAP code can look correct and still work differently from what was intended without any error message, because of limits in SAP that are not properly documented anywhere. If you know these pitfalls, you can catch them systematically.
With LoRA you teach an open AI model your own field without retraining it from scratch, and with quantisation you make it small enough to run on your own hardware.
A language model is an AI model trained on text that can write, summarise and translate text itself.
AI stands for artificial intelligence: software that performs tasks which normally require human thinking.
An MCP server gives an AI assistant live access to your systems and data, through an open standard: the Model Context Protocol.
Local AI runs on your own hardware instead of at a supplier in the cloud, so your data stays in-house.
Ollama is free open-source software that lets you run AI language models on your own computer or server, without the cloud and without costs per question.
The same question, the same model, the same settings and even the same seed, and still a different answer. In the trial we describe here, rounding in the compute kernels played a part. For testing, checks and accountability you need to know this.