Answers with sources, and why "I don't know" is a good answer

AI models write well. That is their strength, and it is also what makes them misleading. A wrong sentence can be just as fluent and confident as a right one. Reading it, you cannot tell them apart.

For a general knowledge question, a mistake rarely matters. For a question about your company, it can be costly: a wrong quote amount, an invented delivery date, a commitment attributed to a client who never made it. That is why the central brain we set up for you follows four strict rules. A central brain is a single place linked to all your tools, which knows what they know and answers you with its sources.

Every answer cites its sources

Fictional example: a fifteen-person joinery company is preparing a site meeting. The site manager asks: "What did we agree with the client about the installation date for the windows?" A useful answer does not just give a date. It shows where each piece comes from:

• the manufacturing lead time announced in the supplier's e-mail, with the date of the message;
• the installation date set during the meeting with the client, from the calendar event and the notes written afterwards;
• the version of the shared schedule in which that date appears.

Each source shows its date and the tool it comes from. The date matters as much as the content: a lead time announced in March may have changed since. You do not have to take the answer on trust. You can check it yourself.

"I don't know", a useful answer

An AI that always answers something looks confident. But a made-up answer costs you more time than no answer at all: first you have to find out it is wrong, then you still have to find the right one.

When the information is not in your tools, the answer says "I don't know" instead of making something up. If nobody ever wrote down a client's budget in an e-mail, a document or a note, the question "What is their budget?" gets a clear answer: the information is not in your data. You then know you need to ask the client.

The rule also applies to part of an answer: what rests on no source is not stated.

No source, no statement.

Numbers come from your data, not from the AI

"How many invoices are overdue?" "What is the total of the quotes sent in September?" These are simple questions. They are also the ones where a poorly designed AI goes wrong most easily, and most quietly. An AI model writes text, one word after another: it is not built to add up a long series of amounts. And to answer, a model usually only receives a selection of passages. If it did the counting itself, it would count what it was shown, not everything that exists. The risks are real:

• a total can leave out lines the model never received;
• a count can include the same invoice twice, because it appears in two different e-mails;
• a wrong number looks just as neat as a right one: nothing gives the error away.

So we keep the two jobs apart. The model understands the question. The calculation itself is done by the database, from your data, through a fixed list of calculations, such as counting records or adding up amounts. It covers all the data concerned, not just what a model would have seen. The model puts the number into its sentence; it never produces it.

Why a fixed list? Because a calculation set in advance can be checked in advance, whereas a calculation improvised for each question cannot. No planned calculation, no number: the model never estimates one.

Everyone only sees what they are allowed to see

A central brain reads a lot. That raises a fair question: who sees what? Our rule: each person only sees in an answer what they are allowed to see in the tools.

Fictional example: in a twelve-person agency, Claire, an account manager, asks: "What did the Dupont client say about the budget?" The answer can draw on her own e-mails with that client and on the documents shared with her. If the managing director asks the same question, his answer draws on what he can see: Claire's e-mails are not part of it. The same goes for numbers: a total only counts what the person asking is allowed to see.

An answer you can check is worth more than a brilliant one. That is our choice: less bluster, more evidence. To see which questions you could ask about your own company, the online diagnostic takes about four minutes and gives you a personalized report.