AI-drafted quotes need a verification step before the client sees them

A fluent draft is a hypothesis about the truth, not the truth; no number reaches a client until someone traces it to a system of record.

Somewhere between the draft and the send button sits the moment this essay is about: an AI assistant has produced a message for a client, the message reads well, and inside it stands a figure, a total, a fee, a timeline, a discount. The question is not whether the language is good. The question is whether anyone has proven the number true before it left the building.

AI drafts are fluent, and fluency is precisely what makes an invented figure dangerous. A clumsy forgery trips the reader's suspicion; a well-formed sentence lowers it. The model does not consult your price list and does not know what your operator actually promised: it produces language shaped like an answer. The draft is a hypothesis, not a source, and treating it as a source is where the damage starts.

The failure versions repeat in every field that has adopted this tooling. There is the confident wrong total, delivered in perfect grammar. There is the fee that was true for a different jurisdiction, or true last year, or true for a package the client never asked about. And there is the promise the operator never made, now standing in writing under the company's name, waiting for the client to hold it against you.

The practice that closes this gap is unglamorous and absolute: every number in an AI draft gets traced to a system of record before the client sees it. The total comes from the price list. The license condition comes from the license itself. The timeline comes from the person who will actually do the work. If a figure cannot be traced, it is not sent, however convincing the sentence around it reads.

We run an AI assistant on our own front line, and we built it with this rule at the core: the assistant never invents figures. Totals come from the operator, from verified line items, not from the model's imagination. And we say plainly that the assistant is AI, because transparency about the machine is what makes strictness about the numbers credible; a company that hides the machine has little standing to demand trust in its figures.

The verification step costs minutes. The wrong total costs the client's trust, and a client who was quoted confidently and wrongly rarely returns to ask why. Whatever tool drafts your client messages, keep the send button behind a person who can trace every figure in them. Guidance, not guarantees: the tools will keep getting more fluent, which is exactly why the habit of tracing numbers cannot be delegated to them.