Grounded AI is only as good as the documents you feed it

A grounded system answers from what you gave it, so a stale document does not weaken the answer: it poisons it, confidently.

The market's version is comforting: connect an assistant, and it knows your business. The honest version is narrower and more useful. Grounding means the system answers from your documents instead of its general memory. It does not know your business. It knows exactly what you gave it, and nothing else.

That mechanism has an edge people prefer not to look at. A stale price list does not produce a hesitant answer; it produces a confident stale answer, delivered in the same calm tone as everything else. A missing page produces either a polite guess or an admission that the system does not know. Which one you get depends on the guardrails, but neither fixes the underlying fact: the input decided the output before the question was asked.

So the quality bar moves upstream, to the documents. Good inputs carry their dates: a version from last quarter behaves differently in your head than a version from last week, and the system cannot guess which one you meant. Good inputs have one owner each, a person whose name sits on the document and whose job includes retiring it when it expires. A file that belongs to everyone belongs to no one, and grounding quietly depends on files that belong to someone.

Structure matters as much as freshness. A document that separates facts from prose, one line per fact, is something the system can quote accurately. A document where facts are woven into paragraphs of persuasion is something it paraphrases, and paraphrase is where drift begins. Coverage completes the picture: the source set should answer the questions customers actually ask, not the questions you prefer to answer.

The practice we landed on is restraint. Keep the source set small, current, and owned. Feed the system what you would swear to in front of a client, and leave out the rest. Feeding it everything feels thorough and behaves like the opposite: more stale files, more contradictions, more confident nonsense dressed as knowledge.

We say this from inside the practice, not from theory: our own consultant answers from the dossier and the verified sources we maintain, and when a fact is not in them, it says so, and it says plainly that it is AI. That is the standard we suggest to any business grounding an assistant on its own material. Grounding is not magic. It is a library discipline applied to machines, and the library is yours to keep honest.