There is a category of AI product that derives its perceived usefulness from confidence. Ask it a question and it answers, fluently, in the register of your own documentation. The problem surfaces two weeks later when someone shares the answer with a customer and a colleague quietly notes that the specific detail is not quite right. A threshold was paraphrased. A date was off by a quarter. The answer sounded authoritative and was not.
Confident is not the same as trustworthy
Language models are good at synthesis and at generating prose that reads like your own documentation. They are also capable of drifting from what a source actually says to what a source would plausibly say. The drift is smooth and hard to detect because the output reads the same whether the model stayed on ground or wandered a little. The only way to catch it consistently is to show the source.
Not a link to a document somewhere in the footnotes. The exact passage the answer was drawn from, shown alongside the claim it supports.
Why verbatim matters
A paraphrase that is 95 percent accurate can still be practically wrong. Policy language, support commitments, contractual thresholds: the precision of the original wording matters. A paraphrase of a policy is, in effect, a new policy that neither you nor the person who wrote the original intended. The only way to preserve what the document actually says is to quote it directly.
Showing the verbatim passage is not an aesthetic decision. It is the only mechanism that lets the person reading the answer verify it against the original without trusting that the AI interpreted the source correctly. That trust should be earned by the quote, not assumed.
What grounding means in practice
When Frond answers a question, it retrieves the passages most relevant to that question from your connected sources. A model reads those passages and composes a plain-language answer. Then every claim in the answer is matched back to the passage it came from, and that passage is returned verbatim alongside the claim.
- Each citation is verified character-for-character against the source document before it is returned.
- A quote that does not match the source document is dropped rather than shown.
- The source document is linked so you can open it and read the surrounding context.
- If multiple sources say different things about the same topic, both are shown and the difference is noted.
The second point is the one that matters most. A citation that appears but does not match the document is worse than no citation, because it creates false confidence. Frond verifies each quote inline, before the answer is returned, on every quote, every time. If you can click through and open a cited source, the passage is there.
When the answer is not in your sources
Grounding also means knowing when to stop. If the question touches something not covered in the connected sources, Frond does not fill the gap from general knowledge. It says it could not find the answer, and it tells you where in your sources it looked.
An honest gap is useful. It tells you there is something missing that should be documented. A confident wrong answer tells you the opposite: that the gap is filled when it is not. A knowledge assistant that fabricates answers to avoid appearing uncertain is less useful than no tool at all, because it creates the appearance of coverage without the reality of it.
Every answer Frond returns is a reading of your material, not a synthesis of the internet dressed in your company's voice.
Permission-aware by design
There is a related question that comes up in every enterprise conversation: what happens when someone asks a question that touches a document they should not have access to?
Frond retrieves only from documents the person asking is permitted to access, based on the access already configured in each connected source. The permission check happens at retrieval, not after the answer is composed. A person does not get a partial answer drawn from a restricted document. They do not learn the document exists. The answer space for each person is a strict subset of what they are allowed to see, and that is always true regardless of how the question is phrased.