Not ready to let AI into your systems, not willing to fall behind.

There is a conversation I have had a dozen times this year, always with the same person. He owns or runs a company that makes or moves things: film, parts, food, grain. He is not against technology. His plant has more sensors than his house. But when the word "AI" comes up, he folds his arms.

"I am not putting my numbers in there."

He is then told, by people selling him something, that this is fear, that he will be left behind, that his competitors are already doing it. He is told he is a paranoiac.

He is not. He is doing his job.

What the owner actually knows

He knows three things that the people calling him paranoid tend to skip.

First, he knows what his numbers are worth. His price list, his margins per customer, his supplier terms, his stock positions, the payment behaviour of every client: that is the company. Not the building, not the machines. Anyone who holds those numbers holds a map of how to take his customers.

Second, he knows he cannot evaluate the systems he is being offered. Where does the data go when it leaves the building? Which model reads it? Who trains on it? Who can be subpoenaed for it, in which country? The honest answer from most vendors is a privacy page, written by a lawyer, that reserves the right to change.

Third, he knows that a decision taken on a wrong answer is still his decision. If a system he does not understand proposes a price, an order or a credit limit, and he signs it, the signature is his. "The tool said so" is not a defence he has ever accepted from his own people, and he will not use it himself.

None of this is paranoia. It is the ordinary discipline of someone who is personally liable for a company. In any other domain we would call it risk management and put it in the annual report.

What he is right to want anyway

Here is the part the folded arms hide. The same owner knows that his company decides too slowly, and that it decides badly when the right person is on holiday.

His purchasing manager spends the first two hours of every day building a list from the ERP, two spreadsheets and the planner's e-mail. His planner reworks the week three times because nobody saw the material shortage until the line stopped. His finance director gets the cash position on Thursday, decides on Friday, and the overdraft was needed on Tuesday. He pays for all of this. Nobody books it to a cost centre, so nobody sees it.

He would like what the vendors are promising: the decisions prepared before people arrive, with the reasons beside them, instantly, every day. He is right to want it. He is also right to refuse the way it is usually offered.

The mistake is to think these are one choice.

The flow that matters, and where it breaks

Every operating decision in a company follows the same path, whether anyone draws it or not.

The systems record what happened. Someone reconciles it, usually in a spreadsheet. Someone prepares the decision: what to order, what to produce, whom to give credit to, where the cash will be short. A person decides. Something is, or is not, written down. Later, something is, or is not, measured.

Look at where the money leaks. It is not in the recording; the ERP does that well enough. It is in the preparation and the record. Preparation is slow because it is done by hand, in the gaps between other work. The record is missing because nobody has time to write down why. And without a record there is no measurement, so the same mistake is made next quarter with the same confidence.

Now look at what the usual AI offer changes. It takes the preparation step out of the building, into a system you cannot evaluate, and often blurs the decision step too: a chat window that answers in confident prose, with no limit on what it may suggest and no record of what it was shown. It fixes the slow part by breaking the part that was working: the fact that a named person, in the building, decided, and could be asked why.

That trade is not worth making. And it does not have to be made.

Doing it in your own house

The preparation of a decision is arithmetic on data the company already has. Stock today, demand, lead times, open orders, payment histories, machine speeds, the plan. There is nothing in that arithmetic that needs to leave the building. There is nothing in it that needs a conversation with a model on someone else's server.

So do it inside. A system that runs on a server the company owns, reads exports from the systems it already has, and never writes back to them. One that needs no internet connection to run, so the question "where does the data go" has a one-word answer: nowhere. One whose calculations are deterministic, so the same inputs give the same proposal and your own people can trace every step when they disagree with it. One that produces, every night, a list of prepared decisions, each with its evidence and the inputs it did not have, for the person who signs.

Then keep the human step exactly where it was, and make it better. A name beside every kind of decision. A written limit: what may go through alone, what needs a signature. A record that cannot be rewritten, by anyone, including the administrator. A baseline written down before a decision is applied, so that afterwards the effect can be measured honestly, in either direction.

That is the whole method. Seven rules; we published them under an open licence so that an auditor can read them before anyone buys anything. It is not a technology position. It is a governance position that happens to be implementable.

What "instantly" means when it is done this way

The owner's objection was never to speed. It was to blindness. Once the preparation runs in the building, speed is the easy part.

The purchasing manager opens one list at 08:10: twelve orders proposed overnight, the stock, the need, the lead time, the supplier's record beside each. Nine go through inside her limit; three need her signature. Done by 08:40. The planner reads a plan per machine and shift built from the orders that actually exist; the shortage was flagged yesterday and the order already went out. The finance director sees the thirteen-week cash position with the two largest customers' real payment behaviour, and the decision on terms for one of them already prepared.

Nobody learned a new system. Nobody's numbers left the room. And when the board asks, in March, why the credit limit for that customer was raised in October, there is a record with a name, a reason and the measured effect.

The question to ask any vendor, including us

If you are the owner with the folded arms, keep them folded until you have four answers in writing.

Where does my data go, and can it run with the internet cable unplugged? Who decides, and is there a limit written down for what the system may do alone? Can my own people trace how a proposal was computed, or is it a model's opinion? What is measured afterwards, against what baseline, and who sees it?

A vendor who cannot answer all four plainly is asking you to lower your risk management to his convenience. You should not. Not because you are afraid of technology, but because you are responsible for a company, and the people who call that paranoia are not.

You can stay in your own house and still have the decisions ready at eight in the morning. That is not a compromise. It is the point.