Suppose a new colleague walks into your office and says, without hesitation: “That contract has a six-month notice period.”
You ask: “Where does it say that?”
That is not distrust. That is simply your job. With people, we ask that question automatically.
With AI, we do not. And that is because of the tone: an AI answer arrives in tidy sentences, without hesitation, without an “I think”. It sounds like knowledge. But sounding certain and being right are two different things.
AI can make things up — with a straight face
AI systems have a well-known weakness: when they do not know the answer, they sometimes invent something that looks right. There is a technical term for it, but what you need to remember is simpler: the wrong answer sounds exactly as convincing as the right one. No warning light comes on.
At home, that is annoying. At work, it is something else. An invented notice period in advice to a client is not a slip — it is your name under something that is not true.
What “showing a source” really means
Some systems put a row of links under the answer. That looks like a source, but often it is decoration: a list of places where you get to figure it out yourself.
A real source is more precise. Not “it is somewhere in your documents”, but: this document, this passage, this sentence. Precise enough that you can check it yourself in ten seconds.
That difference changes everything. With a real source you do not have to believe the AI — you simply check, as easily as with that colleague. The answer changes from “trust me” to “see for yourself”.
Three questions to ask of every AI answer
These work on any system, including ours.
- Which document does this come from — and which passage exactly?
- Can I open that spot with one click and read it myself?
- Does the system say so honestly when it cannot find something in my documents — or does a nice story appear anyway?
If the answer to question one is “the system cannot show that”, you know enough: then every answer is an opinion you have to track down yourself.
How we approach it
Provenize is built around exactly this idea. An answer names the document and the passage it rests on, so you can read it back yourself. The search can show how it arrived at an answer and how sure it is — including when the answer is: “this is not in your documents”. And every question you ask is recorded in the log, so you can show afterwards what was looked up.
Everything happens on your own computer. Your documents go nowhere — here is how that works, in eight steps.
Curious how your current AI use scores on this point? Take the AI scan — five minutes, in your browser, nothing is sent.
To be honest about it: the software that does this (the SDK) is built and public, but still pre-1.0. Provenize Desktop — the app around it for people who do not program — is not out yet; the release is planned for late 2026. Nothing runs in production at a customer yet. It is not a stamp of approval, not a certificate and not a guarantee; we deliberately do not use those words.
Want to think along before it is finished? We are looking for a limited number of firms and teams that use and shape the product before release. It costs nothing. What we ask in return is honest feedback: info@provenize.ai.
Further reading: AI on client files without anything leaving the office, or the honest comparison cloud AI or local AI.
v0.1 · September 2026 · Back to the blog overview