When can you trust an AI's answer? The accuracy gate and source citation

The most misleading thing about AI is that it can speak fluently and confidently even when it is not sure. So “the model answered” and “this answer is trustworthy” are not the same thing. Two mechanisms build the gap between them.

The accuracy gate.

Every output carries a confidence measure. Output below the set threshold never reaches the user; instead it points somewhere — “I'm not sure, look here.” The error is turned openly into “I don't know” rather than being served quietly as if correct. A system knowing when to stay silent is as important as knowing when to speak.

Source citation.

When an assistant is grounded in the company's own documents, it answers only from them and shows which answer came from which document. No answer is produced without a basis. This makes the answer verifiable and, when it is wrong, shows where to correct it.

A third layer is where the arithmetic happens. Numeric work — price, cost, schedule — runs in a deterministic engine; the AI comes in for the interpretation and reading around that arithmetic. The model does not “make up” the number; the engine computes, the model explains.

Trust comes not from how large the model is, but from these three mechanisms: below the threshold is not shown, every answer cites its source, the arithmetic runs in a deterministic engine. At Miletus this layer is not an add-on sold later; it ships inside every solution. Trust is not claimed — it is built.