The first week with KYBERIOS: what to connect first

Picture a company that makes kitchen furniture to order. It is Monday, the new system has been installed on the server in the back office over the weekend, and the owner has a list. The list has eleven items on it: the invoicing program, the customer spreadsheet, the stock sheet, the shop, the bank, the email, the accountant’s portal, the two scanners in the warehouse, the driver’s phone and the CNC machine.

By Wednesday six of them are connected, in the order they happened to be easiest. On Thursday the system says there are more open orders than the invoicing program knows about, the stock figure disagrees with the sheet by a few dozen hinges, and a reminder has been drafted for a customer who paid last week by a transfer the bank feed has not shown yet. Nobody trusts any of it, and the owner spends Friday deciding whether the system is wrong or the company was.

The real cause: connecting by convenience, not by trust

Nothing on the list was a bad idea. The order was. When several sources are joined in one week, every disagreement between them lands on the new system, and nobody can tell which side is right. A system that is supposed to earn write access one process at a time cannot earn anything while it is being asked to referee six processes at once.

Where the first week goes wrong

Starting with the messiest source. The temptation is to fix the worst spreadsheet first, because that is where the pain is. But the worst spreadsheet is also the one nobody agrees about, so the first thing the system learns is a dispute.

Giving everyone an assistant on day one. Twelve people, twelve KyberBots, twelve different ways of asking, and no agreed idea yet of what the system knows. The first wrong answer travels through the whole company before anyone has learned how to check one.

Agents before rules. A swarm that drafts confirmations, reserves stock and proposes reminders is only as good as the limits it runs under. With nothing written down, “it did what it was told” is true and useless at the same time.

Devices before the data they report into. A scanner in the warehouse confirms a pick into a stock figure. If that figure has not been agreed with the stock sheet yet, every scan adds to the confusion with a very confident timestamp.

Write access as a reward for installing. The system can be read-only for weeks and still be useful. Letting it write to the invoicing program in week one because “it is connected anyway” skips the only test that matters: whether it has been right about that process for long enough.

Connecting the bank because it is the most important. The bank is where a mistake costs the most and where the least can be undone. It belongs at the end of the list, not the top.

What to connect first, in order

One source the business already trusts. Pick the record the company argues about least, usually orders or issued invoices. Point the system at it, reading only. For a week the whole job is to see whether what the system shows matches what the person who owns that record would say. One data model beats eleven integrations explains why one clean source is worth more than six joined ones.

One assistant for one person. Give the first KyberBot to the person who owns that source, not to the owner and not to everyone. They learn what it knows, where it is thin, and how to ask a question so the answer can be checked. Their company memory and their private memory are kept apart from the first day, which is also the day to explain that.

Policies before agents. Before the first swarm runs, write the rules it will run under: what it may suggest, what it may do, up to which amount, and what waits for a person. In week one the honest answer for most lines is “suggest only”. Policy, not prompts is the longer argument for why this comes first.

Then one process, as a swarm, under those rules. Order intake or invoice drafting, whichever the trusted source feeds. The system proposes, a person approves, and the log of what was proposed and what was changed is the evidence for widening the limits later. Processes as swarms describes what that looks like.

Devices last. Bands, Cards and scanners earn their place once the stock figure they report into is one the warehouse already agrees with. Then a confirmed pick changes a record people trust, instead of a record people are still checking.

What stays disconnected in week one

The bank, payroll, the email inbox and anything that sends a message to a customer. Not because the system could not read them, but because there is nothing yet to compare their answers against, and because these are the places where the cost of being wrong is highest. The accountant’s portal and the tax side of the business can wait a little longer still; the sibling piece on where a firm leaves its ground covers how those borders are meant to work.

A test you can run today

Take your own list of systems, however long it is, and sort it by one question: if this source and the new system disagreed tomorrow, who in the company could say which one is right, and how long would it take them? Sources with a named person and an answer in minutes go to the top. Sources where the answer is “we would have to check” go to the bottom.

The top item is what you connect first. The bottom items are what stays unplugged, and the gap between them is how much of your company still runs on people remembering.

KYBERIOS is built to start this way: beside what you already run, reading first, earning each process. It is pre-launch; the homepage keeps the honest state of the work.