AI proposes. Who signs?

IThree kinds of deciders

Every company now has three kinds of deciders. There are people, who have limits written down. There are algorithms, which have run inside the ERP for years. And there are AI assistants and agents, which arrived in the last three years and mostly walked past the rulebook.

One in five European companies used AI in 2025, up from 13.5% the year before. Among medium-sized companies it was 30.4%. Those that considered AI and held back gave three main reasons: they lacked the expertise, the legal consequences were unclear, and they worried about their data.

Those are sensible reasons. They describe a company that does not yet know how to stay in control, and that is a different problem from not wanting the technology.

IIThe question the budget cannot answer

Ask a finance director what the company's AI spending returned last year. Most cannot say. The tools were bought by seat and the pilots ran without a starting figure, so there is nothing to compare the result against.

A result without a starting figure proves nothing. Success and failure look the same.

IIIThree things every decision should leave behind

Whatever proposed it, a decision that matters should leave three things.

A name. One person answers for this class of decision. A committee is not a name.

A limit. Inside it, the machine may act. Outside it, a person signs. The limit is a number, written by the owner, with a review date.

A record. What was true, what was proposed, why, who approved, and what happened afterwards in money.

None of this is new. Banks have worked with signature limits and second signatures for centuries. We are only extending the oldest instrument in the building to the newest employee, as I argued in the previous article.

IVSeven rules

We wrote the practice down as seven rules and published them for anyone to use, as the Simon G. Method.

1. Every class of decision has a named owner.
2. Every freedom given to a machine has a signed limit.
3. Every proposal carries its evidence.
4. A person signs.
5. The record cannot be rewritten.
6. The effect is measured against a baseline fixed beforehand.
7. Data and models stay inside a perimeter the company controls.

A company can reach the first level in an afternoon, with one page and no software. Software is needed only at the last level, where the system refuses what the charter does not allow.

VWhat it looks like on a Tuesday

Take the most ordinary decision in a distribution company: what to order today.

A planner exports the opening stock and a year of transactions. The calculation rebuilds the stock of every day, so it knows when a product ran out and corrects for it. It tests several forecasting methods on past dates and uses a complex one only when it beats a simple average. Then it builds the order step by step and shows each step.

The company's limit says that a low-value order for a minor product may go ahead on one signature, and that anything larger goes to the head of supply chain. Each proposal is marked accordingly and leaves a record.

Nothing in this needs a language model, and nothing leaves the company's computer. We built it as a free tool so that anyone can check the claim on their own data.

VIWhat this does not solve

It does not make a model accurate. It does not make a company compliant with any regulation by itself. It does not decide where a limit belongs; two honest companies will put it in different places.

It also takes work. Someone has to write the names down, and that conversation is often the first time a company finds out that nobody owned a decision.

VIIWhere to start

Start with one page this week. List the decisions that matter, put a name beside each, and write one limit. The Decision Charter gives you fifteen classes to begin from. It will be wrong. Yesterday the same wrongness was invisible and belonged to nobody.

If you would rather start from a figure, the Working Capital Check takes two balance sheets and shows what one day of stock is worth. The rest is on the Simon G. page.