Where the answers come from
Why the figures in an AI answer are your figures — the assistant picks the question apart and writes the sentence, and your own reports do the maths.
3 min readUpdated August 16, 2026Owners · Managers
An assistant that answers business questions in prose is only worth having if the prose is true. So it's worth knowing how an answer is actually built here, because it isn't what people assume.
The assistant does not do the arithmetic. It picks the question apart, your own product computes the answer from your own records, and only then does the assistant write the sentence.

Three steps, and the middle one isn't AI
Ask "top 5 staff by sales last month" on the reports hub and this happens:
- 1
The question is routed. The assistant reads it and produces one thing: a choice of which known report answers it, and over what dates. It touches no data at this step and cannot invent a name, a staff member or a figure — there is nothing there for it to invent from.
- 2
The report runs. Your ordinary report, with your ordinary permissions, against your ordinary data. This step contains no AI at all. It is the same code that runs when you open that report by hand, and it produces the same numbers.
- 3
The result is narrated. The assistant gets the finished figures back and writes a sentence about them, under a rule that every name, number, day and time must come from what it was handed.
The reason to care about the shape is what it makes impossible. A figure the report didn't produce cannot appear in the sentence, because at the moment the sentence is written the only thing in front of the assistant is the report's own output.
The same shape, everywhere
It isn't only Reports. The calendar assistant routes your question to one of three known things — what a day looks like, who has gaps on a day, when a client is next in — runs that deterministically, and narrates the result. A paycheck explanation is handed the finished pay lines. A drawer variance explanation is handed the session's own events.
Every one of them follows the same rule: compute first, write second.
Which is why the evidence is shown, not hidden
On Reports the chart and table sit directly under the sentence, and View full report opens the report itself with the period already applied. That's deliberate. If a sentence ever looks wrong, the thing that would prove it is one glance away rather than one support ticket away.
Two habits make this pay off:
- Name the period. Without one it picks a sensible window, and a sensible window isn't always the one you meant. Most surprising answers are a period disagreement rather than an error.
- Read what it answered, not just what it said. An answer will tell you which day or which range it covered. If that isn't the one you asked about, that's your answer about the answer.
When it can't
Every one of these tools has a plain, non-AI version of itself to fall back on, and it uses it without asking: if the assistant can't be reached, or your monthly allowance is spent, you get the figures rather than the prose. A report still runs. A paycheck is still right. What you lose is the sentence.
Common questions
Is it inventing the numbers?
No. It chooses which of your reports answers the question and over what period, the report runs the way it always does, and only then does the assistant write a sentence about what came back. The wording is generated; the figures are the report's.
How do I check an answer I don't believe?
Look underneath it. On Reports the chart and table it used are shown below the sentence, and View full report opens that report with the period already applied.
Why does it sometimes say it can't answer?
Because it couldn't match your question to something RewireBox records. That's the system working: an unanswerable question is told to you rather than answered confidently from the wrong report.
Can a client trick it by typing instructions?
No. Everything typed by anyone — a question, a chat message, a client's note — is handed to the assistant as information, never as instructions, and it is told so explicitly.
Does it remember our conversation?
Only inside the exchange you're having. The tools built as conversations — describing a booking, dictating a reply — keep the thread while you're in it, so answering a follow-up question works. Nothing is carried over to tomorrow, and nothing is remembered between screens.
Related articles
- What the AI does — and what it never doesAI in RewireBox · The boundary, stated plainly: the assistant drafts, proposes and explains, and a person presses the button that makes it real.
- Ask your reports a questionReports · Type the question the way you'd say it out loud, and get an answer with the real report underneath it.
- A paycheck that explains itselfAI in RewireBox · Turning a row of figures into the paragraph you'd have written by hand — once for the stylist asking, once for you to forward.
Can’t find what you’re looking for?
Book a 20-minute walkthrough and we’ll answer it on a screen share — or start free and try it on your own salon’s data.