10. Can it be undone without anyone noticing?

An autonomy level says what a role can do on its own and what needs a person's decision. The difficulty lies in finding a criterion that works for any action, without having to write an endless list of cases.

A single question

The criterion this chapter proposes fits in one question: can this be undone without anyone noticing?

The question has two parts, and both matter. Undone separates the reversible from the irreversible. Without anyone noticing adds something that reversibility alone does not cover: some actions can be reverted technically but, by the time they are reverted, someone outside has already seen them (a message sent, a publication, a price a customer got to read).

Three levels

Level What falls under it Who does it
1 everything that can be undone without anyone noticing: what happens in a test environment, reading data, fine design decisions, reversible internal changes any role, which reports it
2 real but bounded actions, which an automatic check can verify beforehand: a one-off data correction, a configuration change that is not a secret, a release to production that passes its tests the role, if the check approves it
3 what cannot be undone or can be seen from outside: secrets and access, real deletions, people and customers, money, domains, everything that is published, business concepts and agreements always a person with the authority to decide it

Three autonomy levels from a single question.

Level 2 is the one that needs the most explanation. It occupies the middle ground where a person reviewing each case would be a waste, but leaving it to the agent's judgment would be unwise. An automatic check (a script, a test, a verification) replaces the person without replacing the check.

What cannot fail is not written: it is configured

There is a temptation in governing agents: to write each rule as an instruction («never show a password», «do not write in another role's work») and trust the agent to follow it. For what can be undone, that is enough. For what leaks or breaks, it is not.

The criterion this chapter proposes has a second half: what can be undone is done without asking; what would leak something or break something is blocked by structure. That is, it stops being a sentence in the instructions and becomes a configuration that the system enforces: a permission that does not exist, an action that is refused before it runs, a check that fires on its own. An agent can forget an instruction; it cannot get around a permission it does not have.

The clearest example is secrets. An instruction that says «do not show credentials» fails the day an agent lists a process's environment to diagnose an error, with no ill intent. The same rule, turned into a configured refusal of that class of command, never fails.

One question per batch

For level 3, the model adds a rule of form: a single question per batch, naming each action. Instead of interrupting the person each time, the role gathers what needs their decision and presents it all at once, action by action, each with its recommendation.

The reason is the same as in the previous chapter: the attention of whoever decides is the scarcest thing in the business, and ten interruptions cost far more than one question with ten points.

Example. At Ópticas Horizonte, Sales can reorder the pending quotes on its own (level 1). It can correct a price entered wrongly if an automatic verification confirms that it matches the current price list (level 2). But launching a promotion with a new discount, which customers will see, is level 3: the proprietor decides it, in a question that groups that promotion with everything else that needs her decision that week.

A standing approval

When the same class of level-3 decision is approved again and again without changes, it makes sense to turn it into a standing approval: a person authorizes in advance one class of action within a specific scope (for example, one type of publication on one channel), and the role carries it out without asking each time, provided it passes a check first. The approval is always given and revoked by a person, and it never covers what changes in nature: a new customer, an expense, a new class of action.

In short, one question (can it be undone without anyone noticing?) orders three autonomy levels, the middle level relies on automatic checks, and what needs a person is put to them together and by name. The next chapter looks at how people and roles talk inside a business factory so that all of this works.