AIGIS NEWS
Articles · Use cases · White papers
Edition 2026
AI Analysis & Orchestration
Describe your case →

← All articles

Article · 9 November 2026

Nobody trusts their AI. And they are not wrong.

Why we should only listen to an AI once it has to say where it got its information!

### The distrust is justified

Anyone who works with AI today has experienced it: you ask about a contract clause, and the answer comes immediately, neatly structured, in the right tone. Except that the clause does not appear anywhere in that form. A section number is mixed up, a deadline is invented, an obligation does not exist at all.

This is not a slip by individual vendors. Language models are built to produce probable sentences, not true ones. Most of the time the two coincide. When they do not, you cannot tell from the answer. It sounds just as certain.

For a love letter, that does not matter. For a compliance review, a tender or a contract package, it is the end of trust. Anyone who has to check an AI answer sentence by sentence has saved nothing. They have only added a second source of errors.

So anyone who distrusts AI is not wrong. They have been paying attention.

### The actual problem: the AI does not know what it is bound by

A general-purpose AI answers from everything it has ever read. It does not know your requirements, nor your contracts, and certainly not which law applies to your project. If something is missing, it fills the gap, and it does so with whatever sounds plausible.

That is exactly where what is called hallucination arises: the answer is not bound to what is actually there.

This leads to the idea on which AIGIS is built. The question is not whether the AI is clever enough. The question is whether it is bound to a framework that it is not allowed to leave.

### What AIGIS does differently

AIGIS is itself an AI. We do not claim that it cannot make mistakes. We have built it so that it is not allowed to simply say whatever it likes.

**1. First the framework, then the review.**
Before AIGIS reviews a document, the requirements are fixed: internal policies, contractual terms, standards, legal requirements, project decisions. The review is carried out against this framework, not against a model's general knowledge of the world.

**2. No statement without a source reference.**
Every requirement that AIGIS draws from a specification carries the verbatim quote and the document it comes from. Every finding names the point at which two things do not match. Anyone who wants to check does not have to search.

**3. AIGIS states how certain it is.**
Every requirement is given a provenance:

  • *verbatim*: appears in the document as written;
  • *derived*: sharpened from a documented passage;
  • *inferred*: an interpretation.

This label is checked, not taken on faith. If the AI claims a verbatim quote that does not appear word for word in the document, the statement is downgraded. And anything that is only inferred never becomes a finding that blocks a project.

**4. Gaps are named, not filled.**
If information is missing, the result states that it is missing. AIGIS does not invent a deadline, a responsibility or a rule to round out the picture. A result with an open point is more honest than one that only looks complete.

**5. AIGIS does not decide who is right.**
If two requirements contradict each other, for example an internal policy and a legal requirement, AIGIS makes the contradiction visible: the question, the options with their consequences, and who has to decide. The judgment remains with the people who are responsible for it. As long as it is pending, the result carries a visible caveat.

### What comes out of it

A result that contains only what is in your documents. Nothing more, and nothing added.

That sounds modest, but it is the difference between an AI you have to believe and one you can check. Your team no longer cross-reads every sentence to find an invented one. It jumps to the source reference, checks it and decides. The work shifts from distrust to decision.

### Trust is not a feeling but a matter of proof

We do not ask anyone to trust an AI. Not even ours. We require our AI to show, for every statement, where it comes from, and to stay silent where there is nothing to show.

Nobody trusts their AI. Rightly so. Trust the source reference!