If your AI provider uses these terms to explain delays, you should be on your guard.

Not because the terms are wrong. But because they are used systematically to obscure commercial accountability.

1. Hallucination

What it means: The model makes up answers that are factually wrong, with complete conviction and without any warning.

What your provider says: “That’s a known model problem, we’re working on it.”

What you should ask: “In which processes is the model already running in production, and how is hallucination detected and prevented there?”

If there is no answer: The system is not ready for production.

2. Fine-tuning

What it means: The base model is retrained on your own data to specialise it for your context.

What your provider says: “We need another 6 to 8 weeks for fine-tuning, then the quality will be much better.”

What you should ask: “What exactly will get better, and which measurable metric will we use to assess it?”

Fine-tuning is expensive, slow and often used as a justification when prompt engineering has failed. The difference: prompt engineering costs hours. Fine-tuning costs weeks and your data.

3. RAG Retrieval Augmented Generation

What it means: Before every answer, the model searches a knowledge base and adds relevant documents to its context.

What your provider says: “We’re implementing RAG so that the model knows your internal documents.”

What you should ask: “Which documents, of what quality, with what guarantee that they are up to date, and who maintains the knowledge base?”

RAG does not run by itself. An unmaintained knowledge base produces confidently wrong answers. The ongoing maintenance effort is never shown in a demo.

4. Embeddings / Vector Database

What it means: Texts are converted into mathematical vectors and stored in a special database to enable semantic search.

What your provider says: “We have to embed all your data first before we can start the actual development.”

What you should ask: “What does the initial embedding cost, what does the ongoing updating cost, and with which provider is the data stored?”

Embeddings create vendor lock-in. Migrating your data out of a particular provider’s vectors is complex and expensive.

5. Context Window

What it means: The maximum amount of text the model can hold “in memory” at the same time. Anything beyond that is forgotten.

What your provider says: “We’re using a model with a 128k context window, that’s more than enough.”

What you should ask: “What happens to information that is older than the context window, and how is that handled in our use case?”

A large context window doesn’t solve the forgetting problem, it only moves it. And a full context window costs many times more than an empty one.

What these five terms have in common

None of them is wrong. All of them are used to turn technical complexity into a shield.

A managing director who doesn’t ask questions is approving budgets on the basis of hope.

We translate the technology into commercial consequences. Before we build anything, we clarify with you what it costs, who maintains it and how you will measure success.