Five rungs of letting go, and what licenses each one

Autonomy is not one decision

People talk about automation as though it were a switch: either a person does the job or the computer does. In practice there are five distinct positions, and most of the bad outcomes in small-business AI come from skipping two of them.

Rung one: suggest. You are writing; the model offers phrasing, an outline, three subject lines. You type the thing that goes out. Errors cost seconds.

Rung two: draft. The model produces the whole item. You read all of it, every time, and you press send. This is where the great majority of useful small-business AI lives, and a lot of businesses would be better off never leaving it.

Rung three: queue. The model produces the item and prepares the action: the reply sits in a drafts folder, the invoice is created but unsent, the stock order is filled in but not placed. You approve. The difference from rung two is that the work of doing is already finished, so your job shrinks to judging.

Rung four: act, then tell you. It happens. You get a notification and a window in which you can undo it. Nothing is lost if you are quick.

Rung five: act silently. It happens. You find out from the monthly log, or from a customer.

What actually changes as you climb

Not the error rate. The model is the same model on rung two and rung five; it makes mistakes at the same rate either way.

What changes is time to detection. On rung two an error is caught in the ten seconds before you send. On rung three it is caught at the next approval sweep, so within a day. On rung four it is caught within your undo window. On rung five an error can run for weeks, which means it is no longer one error: it is one error multiplied by the number of times the job ran.

That multiplication is the whole risk. A 2% error rate is unremarkable at rung two and is roughly six wrong messages a week at rung five in a business handling three hundred enquiries.

The reversibility test, done honestly

Before any rung above three, ask what undoing this actually involves.

Cheaply reversible: creating a record, tagging something, filing a document, scheduling a task, drafting, adding a row, raising an unsent invoice.

Not reversible, whatever the software claims: sending an email or a message, taking a payment, placing an order with a supplier, publishing anything, deleting anything, making a promise to a customer, submitting a form to a government body.

“Unsend” features are a fourteen-second courtesy, not a control. Once a message has arrived in somebody’s inbox on their phone, the fact that you later retracted it changes nothing about what they read.

What licenses a climb

Not confidence. Evidence, and the evidence is boringly specific: a run of consecutive items where you corrected nothing.

Fifty consecutive clean items is a reasonable bar for moving from rung two to rung three on ordinary work. It is worth being clear-eyed about what that buys you: module five’s arithmetic applies unchanged here. Fifty clean items is consistent with a true error rate of about 6%. It is not proof of safety; it is the point at which watching every single one has stopped being the best use of your attention.

Moving to rung four needs two further things: an undo path you have actually tested, and a notification that a human being reads. A notification going to an address nobody opens turns rung four into rung five with extra steps.

The failure the ladder does not protect you from

Sampling estimates the rate of independent errors: the ones that happen one at a time for uncorrelated reasons.

It tells you nothing about correlated failure. If your facts sheet still says the old price, every single output is wrong, in the same way, at once. Fifty clean items last month is no evidence at all about the morning after the price changed, because the thing that changed was not the model.

This is why the climb is not permanent. Anything that changes the inputs (a price rise, a new product line, a model upgrade, a change in your terms) drops you back to rung two for a fortnight. Not as a punishment. As the only way to find out whether the thing still works.

Where to start

Pick the job you already trust most at rung two, and move it to rung three. Not to rung four. Rung three costs you almost nothing, because approving a prepared action takes about four seconds, and it gives back nearly all the time that reviewing was costing.

Most small businesses that get value from AI over a year are sitting at rung three on two or three tasks, and nowhere higher. That is not timidity. It is where the arithmetic lands.