You never deal with the AI. We do. Every dollar figure in TrueK is worked out by ordinary software from your own sales and your own supplier bills; the AI chooses what is worth telling you and writes it in plain English. If a figure is wrong, that is ours to fix, and we say so in the next report. Here is what the AI actually does, and what stops it doing anything else.
The AI may only speak in numbers the code worked out and handed it. It chooses which facts matter to your week and writes them in plain English. It never works a figure out itself.
That is not a preference. A number in a report has to be reproducible and testable, and a number worked out in a model's head has no test that can hold it. So the arithmetic is code, with hand-checked examples behind it, and the AI is the writer.
The order of work
TrueK runs like a small finance firm. The checkers are pure arithmetic: they read every sale and every supplier bill, and they cannot guess. The analyst turns what they find into plain English. The reviewer can stop the send. You decide what to do about it.
Six checkers · pure arithmetic, no AI
The senior desk
Before anything is sent, each number in what the AI wrote is matched against the numbers it was handed. One digit that did not come from the arithmetic and the whole passage is thrown away. The plainly written version the code produced is sent instead, so the check can only ever cost you some polish, never a figure.
A number that cannot be traced back to an invoice is not printed. That is the whole rule, and it fails towards saying less.
Every finding has the same shape: what happened, what it is worth in dollars, what would have to change to cancel it. When the numbers in front of it back an obvious option, it may say so, as an option offered, never an order given. Wording that strays into instructing you is blocked before the email goes out, by a check that reads the finished text.
You run the cafe. You know which regular would notice a 50 cent rise. That is why a suggestion always arrives with the working that produced it, and why it never promises what a change will earn: it shows the arithmetic under a stated assumption, and that is all an honest number can do.
What goes to the model is numbers and findings. Your cafe's name, your name, your address and your contact details do not. It processes in Sydney, in the same region your data is stored, and nothing you connect is used to train anybody's model.
We do not hand you a model and leave you to check it. TrueK runs the AI, tests what it writes against the arithmetic, and answers for what it sends. A wrong figure is corrected in the next report and named as a correction, not buried. That is part of what you pay for.
Nothing reaches a founding cafe's inbox until the founder has read it and verified the figures himself. If the AI is unavailable or its wording fails a check, the report still arrives: written by the software, in the same plain style, with the same numbers.
A whole quarter read at once, and what TrueK caught in it. It is planned, not built. Founding cafes do not get it today. It is made of the weekly Keeps it reads, so it arrives once a cafe has a quarter of them.
We flag an item as high-prep and we will not put a dollar figure on the minutes. Splitting a wage bill across plates would be a guess, so we do not do it.
TrueK reads. Nothing in it, the AI included, is connected to anything that writes back.
Ask, and you get the arithmetic, not a summary of it.
Email the founderOr read what happens to your data, in full.