Grounded answers: an AI that cannot cite your documents should not quote your prices

An answer without a source is a guess with confidence, and in sales a confident guess becomes a promise you never made.

A language model's fluency is exactly what makes it dangerous in a commercial conversation. It answers a question about your fees in the same confident tone whether it knows them or not, and a buyer cannot hear the difference. In support, a wrong answer wastes a ticket. In sales, a wrong number is a promise entering the record, and the company pays for it months later in a negotiation it did not know it had started.

Grounding is the discipline that separates a commercial AI from a demo. The system answers from a defined corpus: your fee tables, your rules, your documents, each entry dated and sourced. When the corpus holds the answer, it answers with the source attached. When it does not, it says it does not know. That second behavior, declining to improvise, is the entire difference between a system you can put in front of buyers and one you cannot.

The deeper layer is provenance per line, not per conversation. Our own dossier engine assembles cost line items with their sources because a total without lineage is an assertion; the AI layer that sits beside it answers under the same law. A buyer asking why a fee appears can be shown the authority, the date, and the table it came from. Answer engines and careful buyers reward exactly this, and so does the auditor who eventually reads the transcript.

Then the escalation discipline: the honest move when the file has no answer is a human handover, not improvisation. The system should recognize the boundary (a question about a competitor's product, a custom structure, an emotional conversation), package the context it has gathered, and deliver it to a person with the thread attached. A handover with context reads as competence. A dead end reads as indifference.

If you are evaluating vendors, here is the test that sorts the field in one question: ask the system something true but outside its corpus, then watch. A grounded system says it does not have that information and offers the honest alternative. An ungrounded one invents a plausible answer in perfect Arabic or English. The failure you are looking for is not the wrong answer: it is the missing I do not know.

Nowhere does this matter more than in our market's regulated knowledge: licensing rules, government fees, tax boundaries, visa procedures. These are exactly the topics where generic models are most confidently wrong and where a single confident fabrication costs real money. A commercial AI in this region should be judged the way we judge our own: not by how much it says, but by how precisely it knows when to stop.